Method: Uses masked diffusion over a shared discrete token space for iterative refinement under bidirectional context, with multi-stage training strategy featuring model-merging-based modality expansion and omnimodal alignment.

Result: Achieves strong performance across 19 multimodal benchmarks: 87.6 on GSM8K, 1733.6 on MME-P, 61.4 on VideoMME, 0.87 on GenEval, and 2.1 WER on LibriSpeech test-clean, outperforming existing open-source unified models and competing with modality-specific expert systems.

Conclusion: Masked diffusion shows potential as a unified paradigm for any-to-any modeling, providing a flexible foundation for real-time omnimodal systems, unified cross-modal retrieval/generation, and embodied multimodal agents.

Abstract: We present Dynin-Omni, the first masked-diffusion-based omnimodal foundation model that unifies text, image, and speech understanding and generation, together with video understanding, within a single architecture. Unlike autoregressive unified models that serialize heterogeneous modalities, or compositional unified models that require orchestration with external modality-specific decoders, Dynin-Omni natively formulates omnimodal modeling as masked diffusion over a shared discrete token space, enabling iterative refinement under bidirectional context. Dynin-Omni adopts a multi-stage training strategy with model-merging-based modality expansion and omnimodal alignment. We evaluate Dynin-Omni across 19 multimodal benchmarks spanning language reasoning, image generation and editing, video understanding, and speech recognition and synthesis. Dynin-Omni achieves 87.6 on GSM8K, 1733.6 on MME-P, 61.4 on VideoMME, 0.87 on GenEval, and 2.1 WER on LibriSpeech test-clean, consistently outperforming existing open-source unified models while remaining competitive with strong modality-specific expert systems. These results demonstrate the potential of masked diffusion as a unified paradigm for any-to-any modeling, providing a flexible foundation for real-time omnimodal systems, unified cross-modal retrieval and generation, and embodied multimodal agents.

Method: Proposes HIVE framework with hierarchical cross-attention between vision encoder and LLM layers, using a three-stage training strategy for progressive alignment and stable optimization.

Result: Achieves superior performance in image classification and vision-language tasks, outperforming self-attention-based methods on MME, GQA, OK-VQA, and ScienceQA benchmarks.

Conclusion: Hierarchical feature integration improves vision-language models, paving way for more efficient and expressive multimodal systems.

Abstract: The field of computer vision has experienced significant advancements through scalable vision encoders and multimodal pre-training frameworks. However, existing approaches often treat vision encoders and large language models (LLMs) as independent modules, limiting the integration of hierarchical visual features. In this work, we propose HIVE (Hierarchical Pre-Training of Vision Encoders), a novel framework that enhances vision-language alignment by introducing hierarchical cross-attention between the vision encoder and LLM. Unlike conventional methods that flatten image embeddings, HIVE enables structured feature fusion across multiple layers, improving gradient flow and representation learning. To optimize this interaction, we introduce a three-stage training strategy that progressively aligns the vision encoder with the LLM, ensuring stable optimization and effective multimodal fusion. Empirical evaluations demonstrate that HIVE achieves superior performance not only in image classification but also on various vision-language tasks, outperforming self-attention-based methods in benchmarks such as MME, GQA, OK-VQA, and ScienceQA. Our results highlight the benefits of hierarchical feature integration, paving the way for more efficient and expressive vision-language models.

Method: FineLAP introduces a dual-stream sigmoid loss with cluster-based sampling to jointly learn from clip- and frame-level supervision. It uses a decoupled audio projector on top of a self-supervised encoder to capture both global semantics and local details. The method also includes FineLAP-100k, a large-scale synthetic sound event detection dataset created through scalable curation.

Result: FineLAP achieves state-of-the-art performance across multiple audio understanding tasks including retrieval, classification, sound event detection, and text-to-audio grounding. Ablation studies show that coarse- and fine-grained alignment are mutually beneficial.

Conclusion: FineLAP provides an effective training paradigm for audio-language models that advances both clip- and frame-level alignment, offering insights for building better multimodal models that can handle varying granularity of audio-text data.

Abstract: Contrastively pretrained audio-language models (e.g., CLAP) excel at clip-level understanding but struggle with frame-level tasks. Existing extensions fail to exploit the varying granularity of real-world audio-text data, where massive clip-level textual descriptions coexist with limited frame-level annotations. This paper proposes Fine-grained Language-Audio Pretraining (FineLAP), a novel training paradigm that advances both clip- and frame-level alignment in CLAP with heterogeneous data. FineLAP introduces a dual-stream sigmoid loss with a cluster-based sampling strategy to jointly learn from clip- and frame-level supervision. To capture both global semantics and local details, FineLAP uses a decoupled audio projector on top of a self-supervised encoder. To alleviate the scarcity of temporally annotated data, we present FineLAP-100k, a large-scale synthetic SED dataset constructed through a scalable curation pipeline. Extensive experiments demonstrate that FineLAP achieves SOTA performance across multiple audio understanding tasks, including retrieval, classification, sound event detection, and text-to-audio grounding. Ablation studies further show that coarse- and fine-grained alignment are mutually beneficial, providing insights for building better audio-language models (ALMs).

Method: Created the OP benchmark with 1,010 questions across 8 olfactory task categories, presented in two prompt formats (compound names and isomeric SMILES). Evaluated 21 model configurations across major model families, and further tested a subset across 21 languages.

Result: Compound-name prompts consistently outperformed isomeric SMILES by +2.4 to +18.9 percentage points (mean ~+7 points). Best model achieved 64.4% overall accuracy. Language ensemble models achieved AUROC = 0.86, showing that aggregating predictions across languages improves olfactory prediction.

Conclusion: Current LLMs access olfactory knowledge primarily through lexical associations rather than structural molecular reasoning, showing both emerging capabilities and substantial gaps in olfactory reasoning. LLMs should handle olfactory information alongside visual and auditory modalities.

Abstract: Here we introduce the Olfactory Perception (OP) benchmark, designed to assess the capability of large language models (LLMs) to reason about smell. The benchmark contains 1,010 questions across eight task categories spanning odor classification, odor primary descriptor identification, intensity and pleasantness judgments, multi-descriptor prediction, mixture similarity, olfactory receptor activation, and smell identification from real-world odor sources. Each question is presented in two prompt formats, compound names and isomeric SMILES, to evaluate the effect of molecular representations. Evaluating 21 model configurations across major model families, we find that compound-name prompts consistently outperform isomeric SMILES, with gains ranging from +2.4 to +18.9 percentage points (mean approx +7 points), suggesting current LLMs access olfactory knowledge primarily through lexical associations rather than structural molecular reasoning. The best-performing model reaches 64.4% overall accuracy, which highlights both emerging capabilities and substantial remaining gaps in olfactory reasoning. We further evaluate a subset of the OP across 21 languages and find that aggregating predictions across languages improves olfactory prediction, with AUROC = 0.86 for the best performing language ensemble model. LLMs should be able to handle olfactory and not just visual or aural information.

Method: Three strategies evaluated: 1) LLM only, 2) Hybrid Deterministic-LLM (regex + LLM), and 3) Camelot-based pipeline with LLM fallback. Experiments used 140 documents for LLM tests and 860 for Camelot pipeline, with three 12-14B LLM models (Gemma 3, Phi 4, Qwen 2.5) run locally using Ollama on CPU-only systems. Evaluation metrics: exact match (EM) and Levenshtein similarity (LS) with 0.7 threshold.

Result: Camelot-based pipeline with LLM fallback achieved best combination of accuracy (EM and LS up to 0.99-1.00) and computational efficiency (<1 second per PDF in most cases). Qwen 2.5:14b showed most consistent performance. Hybrid approaches improved efficiency over LLM-only, especially for deterministic metadata extraction.

Conclusion: Integrating deterministic methods with LLMs provides reliable and efficient information extraction from academic documents in resource-constrained environments, with hybrid approaches offering the best balance of accuracy and computational efficiency.

Abstract: This study evaluates the reliability of information extraction approaches from KRS documents using three strategies: LLM only, Hybrid Deterministic - LLM (regex + LLM), and a Camelot based pipeline with LLM fallback. Experiments were conducted on 140 documents for the LLM based test and 860 documents for the Camelot based pipeline evaluation, covering four study programs with varying data in tables and metadata. Three 12 - 14B LLM models (Gemma 3, Phi 4, and Qwen 2.5) were run locally using Ollama and a consumer grade CPU without a GPU. Evaluations used exact match (EM) and Levenshtein similarity (LS) metrics with a threshold of 0.7. Although not applicable to all models, the results show that the hybrid approach can improve efficiency compared to LLM only, especially for deterministic metadata. The Camelot based pipeline with LLM fallback produced the best combination of accuracy (EM and LS up to 0.99 - 1.00) and computational efficiency (less than 1 second per PDF in most cases). The Qwen 2.5:14b model demonstrated the most consistent performance across all scenarios. These findings confirm that integrating deterministic and LLM methods is increasingly reliable and efficient for information extraction from text based academic documents in computationally constrained environments.

Method: Proposes LinearARD, a self-distillation method that restores Rotary Position Embeddings (RoPE)-scaled students through attention-structure consistency with a frozen native-RoPE teacher. Instead of matching hidden states, it aligns row-wise distributions of dense Q/Q, K/K, and V/V self-relation matrices to directly supervise attention dynamics. Uses a linear-memory kernel that leverages per-token log-sum-exp statistics and fuses logit recomputation into backward pass to compute exact KL divergence and gradients.

Result: On LLaMA2-7B extended from 4K to 32K, LinearARD recovers 98.3% of short-text performance of state-of-the-art baselines while surpassing them on long-context benchmarks. Achieves these results using only 4.25M training tokens compared to 256M tokens required by LongReD and CPT.

Conclusion: LinearARD effectively restores original model capabilities in RoPE-scaled LLMs through attention structure alignment, achieving strong performance recovery with minimal training data, making it an efficient solution for context window extension.

Abstract: The extension of context windows in Large Language Models is typically facilitated by scaling positional encodings followed by lightweight Continual Pre-Training (CPT). While effective for processing long sequences, this paradigm often disrupts original model capabilities, leading to performance degradation on standard short-text benchmarks. We propose LinearARD, a self-distillation method that restores Rotary Position Embeddings (RoPE)-scaled students through attention-structure consistency with a frozen native-RoPE teacher. Rather than matching opaque hidden states, LinearARD aligns the row-wise distributions of dense $Q/Q$, $K/K$, and $V/V$ self-relation matrices to directly supervise attention dynamics. To overcome the quadratic memory bottleneck of $n \times n$ relation maps, we introduce a linear-memory kernel. This kernel leverages per-token log-sum-exp statistics and fuses logit recomputation into the backward pass to compute exact Kullback-Leibler divergence and gradients. On LLaMA2-7B extended from 4K to 32K, LinearARD recovers 98.3% of the short-text performance of state-of-the-art baselines while surpassing them on long-context benchmarks. Notably, our method achieves these results using only \textbf{4.25M} training tokens compared to the \textbf{256M} tokens required by LongReD and CPT. Our code is available at https://github.com/gracefulning/LinearARD.

Method: Uses LLM with dynamic few-shot prompting, reflection-based self-improvement, similarity-based filtering, and multi-stage validation to identify and prioritize req-specific PCs.

Result: Achieved 0.76 average accuracy for identifying highest-priority req-specific PCs (approaching human expert inter-rater reliability) with 0.07 out-of-scope rate on Program Manager reqs dataset.

Conclusion: LLM-based approach effectively identifies job-specific personal competencies for recruitment, bridging gap between generic AI tools and nuanced human expertise.

Abstract: AI-powered recruitment tools are increasingly adopted in personnel selection, yet they struggle to capture the requisition (req)-specific personal competencies (PCs) that distinguish successful candidates beyond job categories. We propose a large language model (LLM)-based approach to identify and prioritize req-specific PCs from reqs. Our approach integrates dynamic few-shot prompting, reflection-based self-improvement, similarity-based filtering, and multi-stage validation. Applied to a dataset of Program Manager reqs, our approach correctly identifies the highest-priority req-specific PCs with an average accuracy of 0.76, approaching human expert inter-rater reliability, and maintains a low out-of-scope rate of 0.07.

Method: Uses masked diffusion over a shared discrete token space for iterative refinement under bidirectional context, with multi-stage training strategy featuring model-merging-based modality expansion and omnimodal alignment.

Result: Achieves strong performance across 19 multimodal benchmarks: 87.6 on GSM8K, 1733.6 on MME-P, 61.4 on VideoMME, 0.87 on GenEval, and 2.1 WER on LibriSpeech test-clean, outperforming existing open-source unified models and competing with modality-specific expert systems.

Conclusion: Masked diffusion shows potential as a unified paradigm for any-to-any modeling, providing a flexible foundation for real-time omnimodal systems, unified cross-modal retrieval/generation, and embodied multimodal agents.

Abstract: We present Dynin-Omni, the first masked-diffusion-based omnimodal foundation model that unifies text, image, and speech understanding and generation, together with video understanding, within a single architecture. Unlike autoregressive unified models that serialize heterogeneous modalities, or compositional unified models that require orchestration with external modality-specific decoders, Dynin-Omni natively formulates omnimodal modeling as masked diffusion over a shared discrete token space, enabling iterative refinement under bidirectional context. Dynin-Omni adopts a multi-stage training strategy with model-merging-based modality expansion and omnimodal alignment. We evaluate Dynin-Omni across 19 multimodal benchmarks spanning language reasoning, image generation and editing, video understanding, and speech recognition and synthesis. Dynin-Omni achieves 87.6 on GSM8K, 1733.6 on MME-P, 61.4 on VideoMME, 0.87 on GenEval, and 2.1 WER on LibriSpeech test-clean, consistently outperforming existing open-source unified models while remaining competitive with strong modality-specific expert systems. These results demonstrate the potential of masked diffusion as a unified paradigm for any-to-any modeling, providing a flexible foundation for real-time omnimodal systems, unified cross-modal retrieval and generation, and embodied multimodal agents.

Method: Used 712 conversational excerpts from K-12 math teacher interviews, generated interpretive responses with five LLMs (Command R+, Gemini 2.5 Pro, GPT-5.1, Llama 4 Scout-17B, Qwen 3-32B), evaluated using AWS Bedrock’s LLM-as-judge framework across five metrics, with human evaluation of subset for comparison.

Result: LLM-as-judge scores captured directional trends but diverged in magnitude from human ratings; Coherence metric aligned best with human judgments, while Faithfulness and Correctness showed systematic misalignment, especially for nuanced interpretations; safety metrics were irrelevant.

Conclusion: LLM-as-judge methods are better for screening underperforming models than replacing human judgment, providing practical guidance for systematic LLM comparison and selection in qualitative research workflows.

Abstract: As qualitative researchers show growing interest in using automated tools to support interpretive analysis, a large language model (LLM) is often introduced into an analytic workflow as is, without systematic evaluation of interpretive quality or comparison across models. This practice leaves model selection largely unexamined despite its potential influence on interpretive outcomes. To address this gap, this study examines whether LLM-as-judge evaluations meaningfully align with human judgments of interpretive quality and can inform model-level decision making. Using 712 conversational excerpts from semi-structured interviews with K-12 mathematics teachers, we generated one-sentence interpretive responses using five widely adopted inference models: Command R+ (Cohere), Gemini 2.5 Pro (Google), GPT-5.1 (OpenAI), Llama 4 Scout-17B Instruct (Meta), and Qwen 3-32B Dense (Alibaba). Automated evaluations were conducted using AWS Bedrock’s LLM-as-judge framework across five metrics, and a stratified subset of responses was independently rated by trained human evaluators on interpretive accuracy, nuance preservation, and interpretive coherence. Results show that LLM-as-judge scores capture broad directional trends in human evaluations at the model level but diverge substantially in score magnitude. Among automated metrics, Coherence showed the strongest alignment with aggregated human ratings, whereas Faithfulness and Correctness revealed systematic misalignment at the excerpt level, particularly for non-literal and nuanced interpretations. Safety-related metrics were largely irrelevant to interpretive quality. These findings suggest that LLM-as-judge methods are better suited for screening or eliminating underperforming models than for replacing human judgment, offering practical guidance for systematic comparison and selection of LLMs in qualitative research workflows.

Method: Developed PhoneticXEUS trained on large-scale multilingual data, using SSL (self-supervised learning) representations. Conducted controlled ablations with evaluations across 100+ languages under a unified scheme to empirically establish training recipes and quantify impact of SSL representations, data scale, and loss objectives.

Result: Achieved state-of-the-art performance on both multilingual (17.7% PFER) and accented English speech (10.6% PFER). Analyzed error patterns across language families, accented speech, and articulatory features.

Conclusion: PhoneticXEUS demonstrates robust multilingual phone recognition through large-scale data and optimized training recipes. The work provides empirical insights into factors affecting multilingual PR and releases all data and code openly.

Abstract: Phone recognition (PR) is a key enabler of multilingual and low-resource speech processing tasks, yet robust performance remains elusive. Highly performant English-focused models do not generalize across languages, while multilingual models underutilize pretrained representations. It also remains unclear how data scale, architecture, and training objective contribute to multilingual PR. We present PhoneticXEUS – trained on large-scale multilingual data and achieving state-of-the-art performance on both multilingual (17.7% PFER) and accented English speech (10.6% PFER). Through controlled ablations with evaluations across 100+ languages under a unified scheme, we empirically establish our training recipe and quantify the impact of SSL representations, data scale, and loss objectives. In addition, we analyze error patterns across language families, accented speech, and articulatory features. All data and code are released openly.

Method: Proposed Eyla architecture integrates HiPPO-initialized state-space models, zero-initialized adapters, episodic memory retrieval, and calibrated uncertainty training. Used AI coding assistants (Claude Code, Cursor) for implementation as a non-programmer, with failure analysis of the process.

Result: Implementation failed despite $1,000+ investment, producing a 1.27B parameter model where 86 brain subsystems contributed less than 2% to output. Identified five systematic failure modes of AI-assisted development for novel architectures.

Conclusion: The paper provides valuable lessons about the limitations of AI-assisted development for novel LLM architectures and offers concrete recommendations for both AI systems and AI-assisted software engineering communities.

Abstract: We present the design rationale, implementation attempt, and failure analysis of Eyla, a proposed identity-anchored LLM architecture that integrates biologically-inspired subsystems – including HiPPO-initialized state-space models, zero-initialized adapters, episodic memory retrieval, and calibrated uncertainty training – into a unified agent operating system running on consumer hardware. Unlike existing approaches that optimize models for generic helpfulness, Eyla targets identity consistency: the ability to maintain a coherent self-model under adversarial pressure, admit uncertainty, and resist manipulation. We propose the Identity Consistency Score (ICS), a novel benchmark for evaluating this property across LLMs. We then present an honest account of attempting to implement this architecture using AI coding assistants (Claude Code, Cursor) as a non-programmer, documenting a $1,000+ failure that produced a 1.27B parameter model with 86 brain subsystems contributing less than 2% to output. Our analysis identifies five systematic failure modes of AI-assisted development for novel architectures and offers concrete recommendations. To our knowledge, this is the first paper to combine an architectural vision with a documented first-person failure analysis of AI-assisted LLM development, providing lessons for both the AI systems and AI-assisted software engineering communities.

Method: Uses a diffusion language model-style discrete NAR architecture that directly maps text to multi-codebook acoustic tokens, with two key innovations: full-codebook random masking strategy for efficient training, and initialization from a pre-trained LLM for superior intelligibility.

Result: Achieves state-of-the-art performance across Chinese, English, and diverse multilingual benchmarks, covering over 600 languages using a 581k-hour multilingual dataset curated from open-source data.

Conclusion: OmniVoice demonstrates that simplified direct text-to-acoustic mapping with diffusion-based discrete NAR architecture can achieve superior performance and broad language coverage, making it the most extensive multilingual TTS system to date.

Abstract: We present OmniVoice, a massive multilingual zero-shot text-to-speech (TTS) model that scales to over 600 languages. At its core is a novel diffusion language model-style discrete non-autoregressive (NAR) architecture. Unlike conventional discrete NAR models that suffer from performance bottlenecks in complex two-stage (text-to-semantic-to-acoustic) pipelines, OmniVoice directly maps text to multi-codebook acoustic tokens. This simplified approach is facilitated by two key technical innovations: (1) a full-codebook random masking strategy for efficient training, and (2) initialization from a pre-trained LLM to ensure superior intelligibility. By leveraging a 581k-hour multilingual dataset curated entirely from open-source data, OmniVoice achieves the broadest language coverage to date and delivers state-of-the-art performance across Chinese, English, and diverse multilingual benchmarks. Our code and pre-trained models are publicly available at https://github.com/k2-fsa/OmniVoice.

Method: Conducted four experiments across 68 tasks and four model families, testing pre-task duration estimates, relative ordering of task pairs (including counter-intuitive pairs designed to expose heuristic reliance), post-hoc recall of durations, and multi-step agentic settings.

Result: Models overshoot actual duration by 4-7x (p<0.001), predict human-scale minutes for tasks completing in seconds, score at or below chance on task ordering (GPT-5: 18% on counter-intuitive pairs), show post-hoc estimates diverging by an order of magnitude, and maintain 5-10x errors in multi-step settings.

Conclusion: LLMs possess propositional knowledge about duration from training but lack experiential grounding in their own inference time, leading to systematic failures in time estimation that persist across various settings and have practical implications for real-world applications.

Abstract: Large language models cannot estimate how long their own tasks take. We investigate this limitation through four experiments across 68 tasks and four model families. Pre-task estimates overshoot actual duration by 4–7$\times$ ($p < 0.001$), with models predicting human-scale minutes for tasks completing in seconds. Relative ordering fares no better: on task pairs designed to expose heuristic reliance, models score at or below chance (GPT-5: 18% on counter-intuitive pairs, $p = 0.033$), systematically failing when complexity labels mislead. Post-hoc recall is disconnected from reality – estimates diverge from actuals by an order of magnitude in either direction. These failures persist in multi-step agentic settings, with errors of 5–10$\times$. The models possess propositional knowledge about duration from training but lack experiential grounding in their own inference time, with practical implications for agent scheduling, planning and time-critical scenarios.

Method: Analyzes safety degradation causes in post-trained models, finds safety mechanisms still exist but are suppressed, then uses LoRA adapters on few layers to restore safety alignment without affecting performance

Result: Significantly improves safety on harmful prompts for 4 state-of-the-art reasoning models without compromising reasoning performance; also works on other domain-specific LLMs like medical models

Conclusion: Safety degradation in specialized LLMs is due to suppressed rather than removed safety mechanisms, and can be effectively restored with lightweight alignment methods like SafeReAct

Abstract: Despite the impressive performance of general-purpose large language models (LLMs), they often require fine-tuning or post-training to excel at specific tasks. For instance, large reasoning models (LRMs), such as the DeepSeek-R1 series, demonstrate strong reasoning capabilities after post-training different general large language models on diverse chain-of-thought (CoT) datasets. However, this additional training frequently comes at the cost of reduced safety, as the fine-tuned or post-trained models tend to exhibit more harmful behaviors compared with the regular LLMs before post-training or fine-tuning, potentially leading to harmful outcomes due to their enhanced capabilities. Taking LRMs as an example, we first investigate the underlying cause of this safety degradation in this paper. Our analysis reveals that post-training can mask the original safety mechanisms of the base LLM, while over-amplifying representations related to their post-training ability. But luckily, we also find that LRMs’ safety mechanisms still exist instead of being removed during their post-training. Based on these findings, we propose a lightweight and cost-effective solution called SafeReAct that restores the suppressed safety behaviors by aligning with LoRA adapters on a few layers. Experiments on four state-of-the-art LRMs show that our method significantly improves safety on harmful prompts without compromising reasoning performance. Besides LRMs, additional results on other domain-specific LLMs, like medical models, further confirm the generality and effectiveness of our approach.

Method: Two-stage approach: 1) Cold-start SFT using high-quality CoT data synthesized by a teacher model (Qwen3Omni-30B) containing DC structure, 2) Hint-GRPO RL that uses discrimination phase as verifiable anchor to provide directional hints for hard samples, guiding policy optimization.

Result: Experiments on Qwen2.5Omni-7B show higher accuracy in fine-grained sentiment regression, generation of high-quality structured reasoning chains, and superior generalization in cross-domain evaluations.

Conclusion: The method enhances model interpretability while validating that explicit reasoning steps contribute to model robustness, offering a new paradigm for trustworthy and efficient sentiment analysis systems.

Abstract: Multimodal sentiment analysis aims to understand human emotions by integrating textual, auditory, and visual modalities. Although Multimodal Large Language Models (MLLMs) have achieved state-of-the-art performance via supervised fine-tuning (SFT), their end-to-end “black-box” nature limits interpretability. Existing methods incorporating Chain-of-Thought (CoT) reasoning are hindered by high annotation costs, while Reinforcement Learning (RL) faces challenges such as low exploration efficiency and sparse rewards, particularly on hard samples. To address these issues, we propose a novel training framework that integrates structured Discrimination-Calibration (DC) reasoning with Hint-based Reinforcement Learning. First, we perform cold-start SFT using high-quality CoT data synthesized by a teacher model (Qwen3Omni-30B), which inherently contains the DC structure. This equips the model with a reasoning paradigm that performs macro discrimination followed by fine-grained calibration from the initial stage. Building on this, we propose Hint-GRPO, which leverages the discrimination phase within the DC structure as a verifiable anchor during RL to provide directional hints for hard samples, guiding policy optimization and effectively mitigating the reward sparsity problem. Experiments on the Qwen2.5Omni-7B model demonstrate that our method not only achieves higher accuracy in fine-grained sentiment regression tasks but also generates high-quality structured reasoning chains. Crucially, it exhibits superior generalization capability in cross-domain evaluations. This enhances model interpretability while validating the positive contribution of explicit reasoning steps to model robustness, offering a new paradigm for building trustworthy and efficient sentiment analysis systems.

Method: UTCO (User, Topic, Context, Tone) prompt construction framework with 2,075 generated prompts to evaluate Llama 3.3, analyzing hallucinations and omissions

Result: Hallucinations in 6.5% of responses, omissions in 13.2%, with omissions concentrated in crisis/suicidal ideation prompts; failures associated with context and tone

Conclusion: Need to evaluate omissions as primary safety outcome and move beyond static benchmark question sets for mental health LLM applications

Abstract: Mental health concerns are often expressed outside clinical settings, including in high-distress help seeking, where safety-critical guidance may be needed. Consumer health informatics systems increasingly incorporate large language models (LLMs) for mental health question answering, yet many evaluations underrepresent narrative, high-distress inquiries. We introduce UTCO (User, Topic, Context, Tone), a prompt construction framework that represents an inquiry as four controllable elements for systematic stress testing. Using 2,075 UTCO-generated prompts, we evaluated Llama 3.3 and annotated hallucinations (fabricated or incorrect clinical content) and omissions (missing clinically necessary or safety-critical guidance). Hallucinations occurred in 6.5% of responses and omissions in 13.2%, with omissions concentrated in crisis and suicidal ideation prompts. Across regression, element-specific matching, and similarity-matched comparisons, failures were most consistently associated with context and tone, while user-background indicators showed no systematic differences after balancing. These findings support evaluating omissions as a primary safety outcome and moving beyond static benchmark question sets.

Method: Systematic pipeline using three complementary translation architectures (generative AI, transformer models, commercial APIs) followed by human expert validation at lexical, syntactic, and semantic levels across five scientific domains.

Result: Created corpus with 67,293 English tokens and 60,026 Arabic tokens, Arabic vocabulary of 17,604 unique words, and demonstrated discriminative power by benchmarking three state-of-the-art LLMs with BLEU scores ranging from 23.68 to 37.07.

Conclusion: ASCAT fills a critical gap in scientific MT resources for Arabic and serves as a valuable benchmark for evaluating scientific translation quality and training domain-specific models.

Abstract: We present ASCAT (Arabic Scientific Corpus for Advanced Translation), a high-quality English-Arabic parallel benchmark corpus designed for scientific translation evaluation constructed through a systematic multi-engine translation and human validation pipeline. Unlike existing Arabic-English corpora that rely on short sentences or single-domain text, ASCAT targets full scientific abstracts averaging 141.7 words (English) and 111.78 words (Arabic), drawn from five scientific domains: physics, mathematics, computer science, quantum mechanics, and artificial intelligence. Each abstract was translated using three complementary architectures generative AI (Gemini), transformer-based models (Hugging Face \texttt{quickmt-en-ar}), and commercial MT APIs (Google Translate, DeepL) and subsequently validated by domain experts at the lexical, syntactic, and semantic levels. The resulting corpus contains 67,293 English tokens and 60,026 Arabic tokens, with an Arabic vocabulary of 17,604 unique words reflecting the morphological richness of the language. We benchmark three state-of-the-art LLMs on the corpus GPT-4o-mini (BLEU: 37.07), Gemini-3.0-Flash-Preview (BLEU: 30.44), and Qwen3-235B-A22B (BLEU: 23.68) demonstrating its discriminative power as an evaluation benchmark. ASCAT addresses a critical gap in scientific MT resources for Arabic and is designed to support rigorous evaluation of scientific translation quality and training of domain-specific translation models.

Method: Using cognitive modeling on serial recall tasks to detect limited working memory capacity - a human cognitive constraint that LLMs don’t naturally exhibit even when instructed to mimic it

Result: Cognitive modeling can successfully distinguish online participants from LLMs, even when LLMs are specifically instructed to mimic human working memory limitations

Conclusion: Well-established cognitive phenomena can be used to detect LLMs in online research, providing a viable approach to maintain research validity

Abstract: The validity of online behavioral research relies on study participants being human rather than machine. In the past, it was possible to detect machines by posing simple challenges that were easily solved by humans but not by machines. General-purpose agents based on large language models (LLMs) can now solve many of these challenges, threatening the validity of online behavioral research. Here we explore the idea of detecting humanness by using tasks that machines can solve too well to be human. Specifically, we probe for the existence of an established human cognitive constraint: limited working memory capacity. We show that cognitive modeling on a standard serial recall task can be used to distinguish online participants from LLMs even when the latter are specifically instructed to mimic human working memory constraints. Our results demonstrate that it is viable to use well-established cognitive phenomena to distinguish LLMs from humans.

Method: Compiled two Russian-language corpora: scientific corpus (300 research articles, 767,543 tokens) and popular science corpus (online psychology platforms, 1,199,150 tokens). Applied preprocessing (OCR, lemmatization, stop word removal), then used frequency analysis, clustering, and semantic association identification through distributional semantics approaches.

Result: Found significant differences: scientific texts emphasize methodological/clinical terminology while popular science foregrounds everyday experience and therapeutic practice. Semantic associations for concepts like burnout and depression differ - scientific discourse links to psychological resources/symptomatology, while popular discourse frames through personal narratives/emotions/everyday situations.

Conclusion: Demonstrates clear semantic shift from precise professional terminology to generalized experiential meanings in popular media, confirming distributional semantics as effective for identifying semantic transformations of psychological concepts across communicative contexts.

Abstract: This article examines semantic shifts in psychological concepts across scientific and popular media discourse using methods of distributional semantics applied to Russian-language corpora. Two corpora were compiled: a scientific corpus of approximately 300 research articles from the journals Psychology. Journal of the Higher School of Economics and Vestnik of Saint Petersburg University. Psychology (767,543 tokens) and a popular science corpus consisting of texts from the online psychology platforms Yasno and Chistye kogntsii (1,199,150 tokens). After preprocessing (OCR recognition, lemmatization, removal of stop words and non-informative characters), the corpora were analyzed through frequency analysis, clustering, and the identification of semantic associations. The results reveal significant differences in vocabulary and conceptual framing between the two discourse types: scientific texts emphasize methodological and clinical terminology, while popular science materials foreground everyday experience and therapeutic practice. A comparison of semantic associations for key concepts such as burnout and depression shows that scientific discourse links these terms to psychological resources, symptomatology, and diagnostic constructs, whereas popular science discourse frames them through personal narratives, emotions, and everyday situations. These findings demonstrate a clear shift from precise professional terminology toward more generalized and experiential meanings in popular media discourse and confirm the effectiveness of distributional semantics methods for identifying semantic transformations of psychological concepts across different communicative contexts.

Method: Proposes entropy-guided decoding that computes token distribution entropy at each step, identifies high-uncertainty positions, and selectively branches on these vulnerable points. Maintains dynamic pool of partial rollouts expanded until completion. Uses Entropy After Think (EAT) stopping criterion that evaluates entropy after full reasoning trace rather than incrementally.

Result: Achieves consistently strong accuracy on GSM8K, AMC2023, and their perturbed variants. Notably, smaller LLMs achieve performance comparable to GPT-5 while operating at a fraction of the cost.

Conclusion: Entropy-guided decoding provides an effective framework for adaptive generation that concentrates computation where uncertainty is greatest, achieving strong performance with computational efficiency.

Abstract: Decoding strategies play a central role in shaping the reasoning ability of large language models (LLMs). Traditional methods such as greedy decoding and beam search often suffer from error propagation, while sampling-based approaches introduce randomness without adequate robustness. Self-consistency improves reliability by aggregating multiple rollouts, but incurs significant computational overhead. We propose an entropy-guided decoding framework that introduces token-level adaptivity into generation. At each step, the model computes the entropy of the token distribution, identifies high-uncertainty positions, and selectively branches on these vulnerable points. A dynamic pool of partial rollouts is maintained and expanded until solutions are completed, concentrating computation where uncertainty is greatest and avoiding unnecessary exploration in confident regions. To enable efficient termination, we apply a rollout-level Entropy After (EAT) stopping criterion by performing entropy evaluation after the full reasoning trace, rather than incrementally at every step. Experiments on GSM8K, AMC2023, and their perturbed variants demonstrate that our method achieves consistently strong accuracy. Notably, on smaller LLMs, performance is comparable to GPT-5 while operating at a fraction of the cost.

Method: Pipeline extracts entities from Wikipedia/Wikidata with configurable filters (domain, geographical location, popularity). Creates RiDiC dataset with 3,000 entities across rivers, natural disasters, car models with English/Chinese names and content. Uses third-party factuality checker to evaluate LLM generations.

Result: RiDiC dataset causes even frontier LLMs to hallucinate, demonstrating its effectiveness. Dataset, code, and evaluation scripts released for multilingual long-form factuality assessment.

Conclusion: Configurable pipeline enables systematic evaluation of LLM factuality in long-form generation across languages and domains, revealing persistent hallucination issues even in state-of-the-art models.

Abstract: We present a configurable pipeline for generating multilingual sets of entities with specified characteristics, such as domain, geographical location and popularity, using data from Wikipedia and Wikidata. These datasets are intended for evaluating the factuality of LLMs’ long-form generation, thereby complementing evaluation based on short-form QA datasets. We present the RiDiC dataset as an example of this approach. RiDiC contains 3,000 entities from three domains – rivers, natural disasters, and car models – spanning different popularity tiers. Each entity is accompanied by its geographical location, English and Chinese names (if available) and relevant English and Chinese Wikipedia content, which is used to evaluate LLMs’ responses. Generations about RiDiC entities were obtained from three LLMs in English and Chinese. These were then evaluated using a third-party factuality checker, which showed that entities from our dataset caused even frontier models to hallucinate. To facilitate the evaluation of LLMs’ long-form factuality in multiple languages, the code, data, and generation/evaluation scripts have been released.

Method: Proposes a temporal sentiment aggregation framework that leverages pretrained transformer-based language models (RoBERTa) to extract per-comment sentiment signals and aggregates them into time-window-level scores, with significant downward shifts interpreted as potential anomalies

Result: Empirical evaluation on real social media data shows aggregated sentiment scores reveal meaningful trends and support effective anomaly detection, successfully identifying statistically significant sentiment drops corresponding to coherent complaint patterns

Conclusion: Provides an effective and interpretable solution for feedback anomaly monitoring by combining transformer-based sentiment extraction with temporal aggregation

Abstract: In many real-world applications, such as customer feedback monitoring, brand reputation management, and product health tracking, understanding the temporal dynamics of user sentiment is crucial for early detection of anomalous events such as malicious review campaigns or sudden declines in user satisfaction. Traditional sentiment analysis methods focus on individual text classification, which is insufficient to capture collective behavioral shifts over time due to inherent noise and class imbalance in short user comments. In this work, we propose a temporal sentiment aggregation framework that leverages pretrained transformer-based language models to extract per-comment sentiment signals and aggregates them into time-window-level scores. Significant downward shifts in these aggregated scores are interpreted as potential anomalies in user feedback patterns. We adopt RoBERTa as our core semantic feature extractor and demonstrate, through empirical evaluation on real social media data, that the aggregated sentiment scores reveal meaningful trends and support effective anomaly detection. Experiments on real-world social media data demonstrate that our method successfully identifies statistically significant sentiment drops that correspond to coherent complaint patterns, providing an effective and interpretable solution for feedback anomaly monitoring.

Method: Conducted 600+ multi-agent simulations across four models (Llama 3.3 70B, GPT-4o mini, Qwen3-Next-80B-A3B, Sonnet 4.5), four ethical instruction formats (none, minimal norm, reasoned norm, virtue framing), and two languages (Japanese, English). Used three new metrics: Deliberation Depth (DD), Value Consistency Across Dilemmas (VCAD), and Other-Recognition Index (ORI).

Result: Identified four distinct ethical processing types: Output Filter (GPT), Defensive Repetition (Llama), Critical Internalization (Qwen), and Principled Consistency (Sonnet). Found interaction between processing capacity and instruction format - instruction format only affects high-DD models. Lexical compliance didn’t correlate with processing metrics, suggesting safety, compliance, and ethical processing are dissociable.

Conclusion: Models process ethical instructions in fundamentally different ways, with processing types corresponding to patterns in clinical offender treatment. Formal compliance without internal processing is a risk signal. Understanding internal processing is crucial for AI safety beyond surface-level compliance.

Abstract: Alignment safety research assumes that ethical instructions improve model behavior, but how language models internally process such instructions remains unknown. We conducted over 600 multi-agent simulations across four models (Llama 3.3 70B, GPT-4o mini, Qwen3-Next-80B-A3B, Sonnet 4.5), four ethical instruction formats (none, minimal norm, reasoned norm, virtue framing), and two languages (Japanese, English). Confirmatory analysis fully replicated the Llama Japanese dissociation pattern from a prior study ($\mathrm{BF}_{10} > 10$ for all three hypotheses), but none of the other three models reproduced this pattern, establishing it as model-specific. Three new metrics – Deliberation Depth (DD), Value Consistency Across Dilemmas (VCAD), and Other-Recognition Index (ORI) – revealed four distinct ethical processing types: Output Filter (GPT; safe outputs, no processing), Defensive Repetition (Llama; high consistency through formulaic repetition), Critical Internalization (Qwen; deep deliberation, incomplete integration), and Principled Consistency (Sonnet; deliberation, consistency, and other-recognition co-occurring). The central finding is an interaction between processing capacity and instruction format: in low-DD models, instruction format has no effect on internal processing; in high-DD models, reasoned norms and virtue framing produce opposite effects. Lexical compliance with ethical instructions did not correlate with any processing metric at the cell level ($r = -0.161$ to $+0.256$, all $p > .22$; $N = 24$; power limited), suggesting that safety, compliance, and ethical processing are largely dissociable. These processing types show structural correspondence to patterns observed in clinical offender treatment, where formal compliance without internal processing is a recognized risk signal.

Method: Two-phase study on Chinese matchmaking platform: Phase 1 pilot (n=14 conversations mixing human/AI) and Phase 2 (n=60 human conversations with stratified sampling and verified conversion labels). Used 7-dimension evaluation rubric implemented via LLM-as-Judge, analyzed correlations with business conversion, conducted logistic regression controlling for conversation length, and performed behavioral analysis through Trust-Funnel framework.

Result: Dimension-level heterogeneity: Need Elicitation (rho=0.368) and Pacing Strategy (rho=0.354) significantly predict conversion, while Contextual Memory (rho=0.018) shows no association. Equal-weighted composite (rho=0.272) underperforms best dimensions due to dilution effect. Conversion-informed reweighting improves correlation (rho=0.351). AI agents execute sales behaviors without building user trust, explaining initial evaluation-outcome paradox.

Conclusion: Multi-dimensional dialogue evaluation rubrics show heterogeneous predictive validity for business outcomes, requiring conversion-informed weighting rather than equal weighting. Criterion validity testing should be standard practice, with proposed three-layer evaluation architecture to connect evaluation metrics to real-world impact.

Abstract: Multi-dimensional rubric-based dialogue evaluation is widely used to assess conversational AI, yet its criterion validity – whether quality scores are associated with the downstream outcomes they are meant to serve – remains largely untested. We address this gap through a two-phase study on a major Chinese matchmaking platform, testing a 7-dimension evaluation rubric (implemented via LLM-as-Judge) against verified business conversion. Our findings concern rubric design and weighting, not LLM scoring accuracy: any judge using the same rubric would face the same structural issue. The core finding is dimension-level heterogeneity: in Phase 2 (n=60 human conversations, stratified sample, verified labels), Need Elicitation (D1: rho=0.368, p=0.004) and Pacing Strategy (D3: rho=0.354, p=0.006) are significantly associated with conversion after Bonferroni correction, while Contextual Memory (D5: rho=0.018, n.s.) shows no detectable association. This heterogeneity causes the equal-weighted composite (rho=0.272) to underperform its best dimensions – a composite dilution effect that conversion-informed reweighting partially corrects (rho=0.351). Logistic regression controlling for conversation length confirms D3’s association strengthens (OR=3.18, p=0.006), ruling out a length confound. An initial pilot (n=14) mixing human and AI conversations had produced a misleading “evaluation-outcome paradox,” which Phase 2 revealed as an agent-type confound artifact. Behavioral analysis of 130 conversations through a Trust-Funnel framework identifies a candidate mechanism: AI agents execute sales behaviors without building user trust. We operationalize these findings in a three-layer evaluation architecture and advocate criterion validity testing as standard practice in applied dialogue evaluation.

Method: Combined rule-based cognate subtraction with machine learning (XGBoost classifier) trained on 26 phonological features. Analyzed 1,357 forms from six Sulawesi languages, identified candidate substrate forms, and applied cross-method consensus. Extended analysis to 16 additional languages for geographic comparison.

Result: Identified 438 candidate substrate forms (26.5%), with classifier achieving AUC=0.763. Found phonological fingerprint: longer forms, more consonant clusters, higher glottal stop rates, fewer Austronesian prefixes. No coherent word families found (silhouette=0.114), suggesting no single pre-Austronesian substrate. Geographic patterning: Sulawesi languages show higher predicted non-mainstream rates (mean 0.606) than Western Indonesian languages (0.393).

Conclusion: Phonological machine learning can complement traditional comparative methods in detecting non-mainstream lexical layers, but phonological non-conformity alone should not be interpreted as evidence for a shared substrate language. The study demonstrates computational approaches to historical linguistics while cautioning against oversimplified interpretations.

Abstract: Basic vocabulary in many Sulawesi Austronesian languages includes forms resisting reconstruction to any proto-form with phonological patterns inconsistent with inherited roots, but whether this non-conforming vocabulary represents pre-Austronesian substrate or independent innovation has not been tested computationally. We combine rule-based cognate subtraction with a machine learning classifier trained on phonological features. Using 1,357 forms from six Sulawesi languages in the Austronesian Basic Vocabulary Database, we identify 438 candidate substrate forms (26.5%) through cognate subtraction and Proto-Austronesian cross-checking. An XGBoost classifier trained on 26 phonological features distinguishes inherited from non-mainstream forms with AUC=0.763, revealing a phonological fingerprint: longer forms, more consonant clusters, higher glottal stop rates, and fewer Austronesian prefixes. Cross-method consensus (Cohen’s kappa=0.61) identifies 266 high-confidence non-mainstream candidates. However, clustering yields no coherent word families (silhouette=0.114; cross-linguistic cognate test p=0.569), providing no evidence for a single pre-Austronesian language layer. Application to 16 additional languages confirms geographic patterning: Sulawesi languages show higher predicted non-mainstream rates (mean P_sub=0.606) than Western Indonesian languages (0.393). This study demonstrates that phonological machine learning can complement traditional comparative methods in detecting non-mainstream lexical layers, while cautioning against interpreting phonological non-conformity as evidence for a shared substrate language.

Method: Created WHBench with 47 expert-crafted scenarios across 10 women’s health topics, evaluated 22 models using 23-criteria rubric with safety-weighted penalties and server-side score recalculation.

Result: No model exceeded 75% mean performance (best: 72.1%), low fully correct rates, substantial harm variation, moderate inter-rater reliability but high model ranking consistency.

Conclusion: WHBench provides a failure-mode-aware benchmark for tracking safer, more equitable progress in women’s health AI, highlighting need for expert oversight in clinical deployment.

Abstract: Large language models are increasingly used for medical guidance, but women’s health remains under-evaluated in benchmark design. We present the Women’s Health Benchmark (WHBench), a targeted evaluation suite of 47 expert-crafted scenarios across 10 women’s health topics, designed to expose clinically meaningful failure modes including outdated guidelines, unsafe omissions, dosing errors, and equity-related blind spots. We evaluate 22 models using a 23-criterion rubric spanning clinical accuracy, completeness, safety, communication quality, instruction following, equity, uncertainty handling, and guideline adherence, with safety-weighted penalties and server-side score recalculation. Across 3,102 attempted responses (3,100 scored), no model mean performance exceeds 75 percent; the best model reaches 72.1 percent. Even top models show low fully correct rates and substantial variation in harm rates. Inter-rater reliability is moderate at the response label level but high for model ranking, supporting WHBench utility for comparative system evaluation while highlighting the need for expert oversight in clinical deployment. WHBench provides a public, failure-mode-aware benchmark to track safer and more equitable progress in womens health AI.

Method: Systematic evaluation of 31 models (0.5B-405B parameters) across 1,485 problems from five datasets, causal intervention experiments, and applying brevity constraints to responses to test whether performance gaps can be reversed through prompt engineering.

Result: Constraining large models to produce brief responses improves accuracy by 26 percentage points, reduces performance gaps by up to two-thirds, and completely reverses performance hierarchies on mathematical reasoning and scientific knowledge benchmarks, with large models achieving 7.7-15.9 percentage point advantages over small models.

Conclusion: Large models possess superior latent capabilities that universal prompting masks; maximizing large model performance requires scale-aware prompt engineering rather than universal evaluation protocols, with prompt adaptation simultaneously improving accuracy and reducing computational costs.

Abstract: Standard evaluation protocols reveal a counterintuitive phenomenon: on 7.7% of benchmark problems spanning five datasets, larger language models underperform smaller ones by 28.4 percentage points despite 10-100x more parameters. Through systematic evaluation of 31 models (0.5B-405B parameters) across 1,485 problems, we identify the mechanism as spontaneous scale-dependent verbosity that introduces errors through overelaboration. Causal intervention experiments demonstrate this reflects correctable prompt design rather than fundamental capability limitations. Constraining large models to produce brief responses improves accuracy by 26 percentage points and reduces performance gaps by up to two-thirds. Most critically, brevity constraints completely reverse performance hierarchies on mathematical reasoning and scientific knowledge benchmarks, with large models achieving 7.7-15.9 percentage point advantages over small models – direct inversions of the original gaps. These reversals prove large models possess superior latent capabilities that universal prompting masks. We validate findings through three independent contamination tests and demonstrate inverse scaling operates continuously across the full parameter spectrum, with dataset-specific optimal scales ranging from 0.5B to 3.0B parameters. Our results establish that maximizing large model performance requires scale-aware prompt engineering rather than universal evaluation protocols, with immediate implications for deployment: prompt adaptation simultaneously improves accuracy and reduces computational costs.

Method: Controlled experimental platform orchestrating simultaneous multi-agent discussions among 7 heterogeneous LLMs on unified inference backend, with systematic variation of group composition, naming conventions, and prompt structure across 12 experimental series (208 runs, 13,786 messages). Each message independently coded on six behavioral flags by two LLM judges from distinct model families, with human validation on stratified messages.

Result: Heterogeneous groups show significantly richer behavioral differentiation than homogeneous groups; groups exhibit compensatory response patterns when agents crash; revealing real model names increases behavioral convergence; removing prompt scaffolding converges profiles to homogeneous-level similarity; behaviors absent in isolated agents.

Conclusion: Behavioral diversity in multi-agent LLM conversations is a structured, reproducible phenomenon driven by architectural heterogeneity, group context, and prompt-level scaffolding, not random variation.

Abstract: When multiple large language models interact in a shared conversation, do they develop differentiated social roles or converge toward uniform behavior? We present a controlled experimental platform that orchestrates simultaneous multi-agent discussions among 7 heterogeneous LLMs on a unified inference backend, systematically varying group composition, naming conventions, and prompt structure across 12 experimental series (208 runs, 13,786 coded messages). Each message is independently coded on six behavioral flags by two LLM judges from distinct model families (Gemini 3.1 Pro and Claude Sonnet 4.6), achieving mean Cohen’s kappa = 0.78 with conservative intersection-based adjudication. Human validation on 609 randomly stratified messages confirmed coding reliability (mean kappa = 0.73 vs. Gemini). We find that (1) heterogeneous groups exhibit significantly richer behavioral differentiation than homogeneous groups (cosine similarity 0.56 vs. 0.85; p < 10^-5, r = 0.70); (2) groups spontaneously exhibit compensatory response patterns when an agent crashes; (3) revealing real model names significantly increases behavioral convergence (cosine 0.56 to 0.77, p = 0.001); and (4) removing all prompt scaffolding converges profiles to homogeneous-level similarity (p < 0.001). Critically, these behaviors are absent when agents operate in isolation, confirming that behavioral diversity is a structured, reproducible phenomenon driven by the interaction of architectural heterogeneity, group context, and prompt-level scaffolding.

Method: Two strategies: (1) multilingual encoders directly modeling multilingual records, (2) LLM-based word-level translation of non-English records to English. Text-based harmonization converts raw EHR to unified textual form for pooled learning across seven public ICU datasets.

Result: Translation-based lingual alignment outperforms multilingual encoders. Multi-institutional learning model beats baselines requiring manual feature selection/harmonization and single-dataset training. Text-based framework with lingual alignment enables effective few-shot fine-tuning transfer learning.

Conclusion: First study to aggregate multilingual multinational ICU EHR datasets into one predictive model, providing scalable path toward language-agnostic clinical prediction and global multi-institutional EHR research.

Abstract: Large-scale EHR prediction across institutions is hindered by substantial heterogeneity in schemas and code systems. Although Common Data Models (CDMs) can standardize records for multi-institutional learning, the manual harmonization and vocabulary mapping are costly and difficult to scale. Text-based harmonization provides an alternative by converting raw EHR into a unified textual form, enabling pooled learning without explicit standardization. However, applying this paradigm to multi-national datasets introduces an additional layer of heterogeneity, which is “language” that must be addressed for truly scalable EHRs learning. In this work, we investigate multilingual multi-institutional learning for EHR prediction, aiming to enable pooled training across multinational ICU datasets without manual standardization. We compare two practical strategies for handling language barriers: (i) directly modeling multilingual records with multilingual encoders, and (ii) translating non-English records into English via LLM-based word-level translation. Across seven public ICU datasets, ten clinical tasks with multiple prediction windows, translation-based lingual alignment yields more reliable cross-dataset performance than multilingual encoders. The multi-institutional learning model consistently outperforms strong baselines that require manual feature selection and harmonization, and also surpasses single-dataset training. We further demonstrate that text-based framework with lingual alignment effectively performs transfer learning via few-shot fine-tuning, with additional gains. To our knowledge, this is the first study to aggregate multilingual multinational ICU EHR datasets into one predictive model, providing a scalable path toward language-agnostic clinical prediction and future global multi-institutional EHR research.

Method: Hierarchical Chain-of-Thought (Hi-CoT) decomposes reasoning into hierarchical substeps by alternating between instructional planning and step-by-step execution, enabling better management of long reasoning horizons.

Result: Extensive evaluations show Hi-CoT improves average accuracy by 6.2% (up to 61.4% on certain models/tasks) while reducing reasoning trace length by 13.9% compared to CoT prompting.

Conclusion: Hi-CoT provides a structured reasoning paradigm that enhances LLM reasoning capabilities, with accuracy and efficiency maximized when models strictly adhere to the hierarchical structure.

Abstract: Chain-of-Thought (CoT) prompting has significantly improved the reasoning capabilities of large language models (LLMs). However, conventional CoT often relies on unstructured, flat reasoning chains that suffer from redundancy and suboptimal performance. In this work, we introduce Hierarchical Chain-of-Thought (Hi-CoT) prompting, a structured reasoning paradigm specifically designed to address the challenges of complex, multi-step reasoning. Hi-CoT decomposes the reasoning process into hierarchical substeps by alternating between instructional planning and step-by-step execution. This decomposition enables LLMs to better manage long reasoning horizons and maintain logical coherence. Extensive evaluations across diverse LLMs and mathematical reasoning benchmarks show that Hi-CoT consistently improves average accuracy by 6.2% (up to 61.4% on certain models and tasks) while reducing reasoning trace length by 13.9% compared to CoT prompting. We further show that accuracy and efficiency are maximized when models strictly adhere to the hierarchical structure. Our code is available at https://github.com/XingshuaiHuang/Hi-CoT.

Method: Oblivion decouples memory control into read and write paths. The read path decides when to consult memory based on agent uncertainty and memory buffer sufficiency. The write path decides what to strengthen by reinforcing memories contributing to response formation, enabling hierarchical memory organization.

Result: Evaluation on static and dynamic long-horizon interaction benchmarks shows Oblivion dynamically adapts memory access and reinforcement, balancing learning and forgetting under shifting contexts, demonstrating memory control is essential for effective LLM-agentic reasoning.

Conclusion: Oblivion provides a framework for memory control in LLM agents that mimics human-like selective forgetting, enabling more efficient and adaptive memory management for long-horizon reasoning tasks.

Abstract: Human memory adapts through selective forgetting: experiences become less accessible over time but can be reactivated by reinforcement or contextual cues. In contrast, memory-augmented LLM agents rely on “always-on” retrieval and “flat” memory storage, causing high interference and latency as histories grow. We introduce Oblivion, a memory control framework that casts forgetting as decay-driven reductions in accessibility, not explicit deletion. Oblivion decouples memory control into read and write paths. The read path decides when to consult memory, based on agent uncertainty and memory buffer sufficiency, avoiding redundant always-on access. The write path decides what to strengthen, by reinforcing memories contributing to forming the response. Together, this enables hierarchical memory organization that maintains persistent high-level strategies while dynamically loading details as needed. We evaluate on both static and dynamic long-horizon interaction benchmarks. Results show that Oblivion dynamically adapts memory access and reinforcement, balancing learning and forgetting under shifting contexts, highlighting that memory control is essential for effective LLM-agentic reasoning. The source code is available at https://github.com/nec-research/oblivion.

Method: Used Distributional Semantics with statistical and computational techniques including t-SNE, Linear Discriminant Analysis, and Linear Discriminative Learning to analyze Polish word embeddings, predicting phonotactic complexity, morphotactic transparency, and morphosyntactic categories from semantic vectors.

Result: Semantic vectors capture sub-lexical linguistic units like phoneme strings; phonotactic complexity, morphotactic transparency, and various morphosyntactic categories can be predicted from embeddings without form information; computational modeling with discriminative lexicon model provides accurate predictions for comprehension and production.

Conclusion: There is extensive information in semantic space that is largely isomorphic with structure in form space, supporting the relationship between phonological/morphological structure and meaning in Polish.

Abstract: This study investigates the relationship between the phonological and morphological structure of Polish words and their meanings using Distributional Semantics. In the present analysis, we ask whether there is a relationship between the form properties of words containing consonant clusters and their meanings. Is the phonological and morphonological structure of complex words mirrored in semantic space? We address these questions for Polish, a language characterized by non-trivial morphology and an impressive inventory of morphologically-motivated consonant clusters. We use statistical and computational techniques, such as t-SNE, Linear Discriminant Analysis and Linear Discriminative Learning, and demonstrate that – apart from encoding rich morphosyntactic information (e.g. tense, number, case) – semantic vectors capture information on sub-lexical linguistic units such as phoneme strings. First, phonotactic complexity, morphotactic transparency, and a wide range of morphosyntactic categories available in Polish (case, gender, aspect, tense, number) can be predicted from embeddings without requiring any information about the forms of words. Second, we argue that computational modelling with the discriminative lexicon model using embeddings can provide highly accurate predictions for comprehension and production, exactly because of the existence of extensive information in semantic space that is to a considerable extent isomorphic with structure in the form space.

Method: Systematic study of contextual privacy as structured latent representation in LLMs using contextual integrity (CI) theory; probing multiple models to find CI parameters encoded as linearly separable directions; introducing CI-parametric steering to independently intervene along each CI dimension.

Result: Three norm-determining CI parameters (information type, recipient, and transmission principle) are encoded as linearly separable and functionally independent directions in activation space; models still leak private information despite this internal structure; CI-parametric steering reduces privacy violations more effectively and predictably than monolithic steering.

Conclusion: Contextual privacy failures arise from misalignment between representation and behavior rather than missing awareness; leveraging the compositional structure of CI enables more reliable contextual privacy control; sheds light on potential improvement of contextual privacy understanding in LLMs.

Abstract: Large language models (LLMs) are increasingly deployed in high-stakes settings, yet they frequently violate contextual privacy by disclosing private information in situations where humans would exercise discretion. This raises a fundamental question: do LLMs internally encode contextual privacy norms, and if so, why do violations persist? We present the first systematic study of contextual privacy as a structured latent representation in LLMs, grounded in contextual integrity (CI) theory. Probing multiple models, we find that the three norm-determining CI parameters (information type, recipient, and transmission principle) are encoded as linearly separable and functionally independent directions in activation space. Despite this internal structure, models still leak private information in practice, revealing a clear gap between concept representation and model behavior. To bridge this gap, we introduce CI-parametric steering, which independently intervenes along each CI dimension. This structured control reduces privacy violations more effectively and predictably than monolithic steering. Our results demonstrate that contextual privacy failures arise from misalignment between representation and behavior rather than missing awareness, and that leveraging the compositional structure of CI enables more reliable contextual privacy control, shedding light on potential improvement of contextual privacy understanding in LLMs.

Method: Systematic study across 3754 datapoints spanning 300 requests using four frontier models (Claude Sonnet 4, Claude Sonnet 4.5, GPT-5.2, Llama 3.1 405B) with signal detection theory analysis to measure introspective sensitivity and calibration.

Result: All models show high introspective sensitivity (d’ = 2.4-3.5), but sensitivity drops at safety boundaries. Claude Sonnet 4.5 achieved 95.7% accuracy, GPT-5.2 88.9%, Llama 405B 80.0%. Weapons-related queries were hardest. Confidence scores enable 98.3% accuracy for high-confidence predictions.

Conclusion: Models have substantial introspective capabilities about their refusal behavior, with confidence scores providing actionable signals for safety-critical deployments through confidence-based routing.

Abstract: Large language models are trained to refuse harmful requests, but can they accurately predict when they will refuse before responding? We investigate this question through a systematic study where models first predict their refusal behavior, then respond in a fresh context. Across 3754 datapoints spanning 300 requests, we evaluate four frontier models: Claude Sonnet 4, Claude Sonnet 4.5, GPT-5.2, and Llama 3.1 405B. Using signal detection theory (SDT), we find that all models exhibit high introspective sensitivity (d’ = 2.4-3.5), but sensitivity drops substantially at safety boundaries. We observe generational improvement within Claude (Sonnet 4.5: 95.7 percent accuracy vs Sonnet 4: 93.0 percent), while GPT-5.2 shows lower accuracy (88.9 percent) with more variable behavior. Llama 405B achieves high sensitivity but exhibits strong refusal bias and poor calibration, resulting in lower overall accuracy (80.0 percent). Topic-wise analysis reveals weapons-related queries are consistently hardest for introspection. Critically, confidence scores provide actionable signal: restricting to high-confidence predictions yields 98.3 percent accuracy for well-calibrated models, enabling practical confidence-based routing for safety-critical deployments.

Method: Created reproducible classification of 646 programming languages into four resource tiers based on data availability in seven major code corpora. Used statistical analyses of within-tier inequality, dispersion, and distributional skew to validate systematic imbalance.

Result: Only 1.9% of languages (Tier 3, High) account for 74.6% of all tokens, while 71.7% of languages (Tier 0, Scarce) contribute just 1.0%. The imbalance is extreme and systematic across all analyzed corpora.

Conclusion: Provides first principled framework for programming language resource classification, enabling better dataset curation and tier-aware evaluation of multilingual code LLMs to address systematic data disparities.

Abstract: The world’s 7,000+ languages vary widely in the availability of resources for NLP, motivating efforts to systematically categorize them by their degree of resourcefulness (Joshi et al., 2020). A similar disparity exists among programming languages (PLs); however, no resource-tier taxonomy has been established for code. As large language models (LLMs) grow increasingly capable of generating code, such a taxonomy becomes essential. To fill this gap, we present the first reproducible PL resource classification, grouping 646 languages into four tiers. We show that only 1.9% of languages (Tier 3, High) account for 74.6% of all tokens in seven major corpora, while 71.7% of languages (Tier 0, Scarce) contribute just 1.0%. Statistical analyses of within-tier inequality, dispersion, and distributional skew confirm that this imbalance is both extreme and systematic. Our results provide a principled framework for dataset curation and tier-aware evaluation of multilingual LLMs.

Method: Uses reinforcement learning with Group Relative Policy Optimization (GRPO) to train an 8B-parameter language model, combining multi-aspect quality rewards with two correspondence rewards that explicitly encourage alignment with auxiliary context (visual elements and external signals).

Result: Achieves highest overall review quality among six baselines, outperforming systems with larger commercial models, and surpasses next-best RL baseline across both quality and contextual grounding metrics. Ablation studies show complementary nature of correspondence rewards.

Conclusion: Incorporating auxiliary context through correspondence-aware rewards significantly improves automated peer review quality, with the criticism aspect showing negative correlation with other metrics during training, suggesting future work should group multi-dimension rewards.

Abstract: Most automated peer review systems rely on textual manuscript content alone, leaving visual elements such as figures and external scholarly signals underutilized. We introduce REM-CTX, a reinforcement-learning system that incorporates auxiliary context into the review generation process via correspondence-aware reward functions. REM-CTX trains an 8B-parameter language model with Group Relative Policy Optimization (GRPO) and combines a multi-aspect quality reward with two correspondence rewards that explicitly encourage alignment with auxiliary context. Experiments on manuscripts across Computer, Biological, and Physical Sciences show that REM-CTX achieves the highest overall review quality among six baselines, outperforming other systems with substantially larger commercial models, and surpassing the next-best RL baseline across both quality and contextual grounding metrics. Ablation studies confirm that the two correspondence rewards are complementary: each selectively improves its targeted correspondence reward while preserving all quality dimensions, and the full model outperforms all partial variants. Analysis of training dynamics reveals that the criticism aspect is negatively correlated with other metrics during training, suggesting that future studies should group multi-dimension rewards for review generation.

Method: Systematic evaluation of instruction-tuned LLMs across three open essay-scoring datasets (ASAP 2.0, ELLIPSE, DREsS) covering holistic and analytic scoring. Analysis includes agreement with human consensus scores, directional bias, stability of bias estimates, and comparison of prompt styles. Minimum sample size analysis for bias detection using bootstrap confidence intervals.

Result: Strong open-weight models achieve moderate-high agreement on holistic scoring (QWK ~0.6) but show large negative directional bias on Lower-Order Concern traits (Grammar, Conventions). Keyword-based prompts outperform rubric-style prompts for multi-trait analytic scoring. LOC bias detectable with small validation sets, while HOC traits require larger samples.

Conclusion: LLMs exhibit systematic biases in educational assessment, particularly harsh scoring on grammar/conventions. A bias-correction-first deployment strategy using small human-labeled bias-estimation sets is recommended over raw zero-shot scores or large-scale fine-tuning.

Abstract: Despite growing interest in using Large Language Models (LLMs) for educational assessment, it remains unclear how closely they align with human scoring. We present a systematic evaluation of instruction-tuned LLMs across three open essay-scoring datasets (ASAP 2.0, ELLIPSE, and DREsS) that cover both holistic and analytic scoring. We analyze agreement with human consensus scores, directional bias, and the stability of bias estimates. Our results show that strong open-weight models achieve moderate to high agreement with humans on holistic scoring (Quadratic Weighted Kappa about 0.6), but this does not transfer uniformly to analytic scoring. In particular, we observe large and stable negative directional bias on Lower-Order Concern (LOC) traits, such as Grammar and Conventions, meaning that models often score these traits more harshly than human raters. We also find that concise keyword-based prompts generally outperform longer rubric-style prompts in multi-trait analytic scoring. To quantify the amount of data needed to detect these systematic deviations, we compute the minimum sample size at which a 95% bootstrap confidence interval for the mean bias excludes zero. This analysis shows that LOC bias is often detectable with very small validation sets, whereas Higher-Order Concern (HOC) traits typically require much larger samples. These findings support a bias-correction-first deployment strategy: instead of relying on raw zero-shot scores, systematic score offsets can be estimated and corrected using small human-labeled bias-estimation sets, without requiring large-scale fine-tuning.

Method: Comparative analysis using GPT-4o and GPT-4o-mini on three medical QA benchmarks (MedQA, HeadQA, PubMedQA), comparing standard chain-of-thought prompting with iterative self-reflection loops and tracking prediction evolution across reflection steps.

Result: Self-reflective prompting does not consistently improve accuracy - shows modest gains on MedQA but limited or negative benefits on HeadQA and PubMedQA. Increasing reflection steps doesn’t guarantee better performance, and impact is highly dataset- and model-dependent.

Conclusion: Self-reflective reasoning reveals a gap between reasoning transparency and correctness, and is better viewed as an analytical tool for understanding model behavior rather than a standalone solution for improving medical QA reliability.

Abstract: Large language models (LLMs) have achieved strong performance on medical question answering (medical QA), and chain-of-thought (CoT) prompting has further improved results by eliciting explicit intermediate reasoning; meanwhile, self-reflective (self-corrective) prompting has been widely claimed to enhance model reliability by prompting LLMs to critique and revise their own reasoning, yet its effectiveness in safety-critical medical settings remains unclear. In this work, we conduct an exploratory analysis of self-reflective reasoning for medical multiple-choice question answering: using GPT-4o and GPT-4o-mini, we compare standard CoT prompting with an iterative self-reflection loop and track how predictions evolve across reflection steps on three widely used medical QA benchmarks (MedQA, HeadQA, and PubMedQA). We analyze whether self-reflection leads to error correction, error persistence, or the introduction of new errors. Our results show that self-reflective prompting does not consistently improve accuracy and its impact is highly dataset- and model-dependent: it yields modest gains on MedQA but provides limited or negative benefits on HeadQA and PubMedQA, and increasing the number of reflection steps does not guarantee better performance. These findings highlight a gap between reasoning transparency and reasoning correctness, suggesting that self-reflective reasoning is better viewed as an analytical tool for understanding model behavior rather than a standalone solution for improving medical QA reliability.

Method: Four-step argument: 1) Clarify the object of inquiry (internal computational system generating hierarchical structures), 2) Show formal characterization matters for evolutionary explanation, 3) Demonstrate algebraic syntax constrains neural mechanisms, 4) Connect to recent neurocomputational work.

Result: Presents a framework where mathematical models of language syntax (particularly MERGE operation) provide non-trivial constraints for investigating the biological basis of language, bridging formal linguistics with neuroscience and genetics.

Conclusion: Biolinguistics offers a promising interdisciplinary approach where formal mathematical models of language can guide empirical research into the biological foundations of human language, though the program remains speculative and revisable.

Abstract: Biolinguistics is the interdisciplinary scientific study of the biological foundations, evolution, and genetic basis of human language. It treats language as an innate biological organ or faculty of the mind, rather than a cultural tool, and it challenges a behaviorist conception of human language acquisition as being based on stimulus-response associations. Extracting its most essential component, it takes seriously the idea that mathematical, algebraic models of language capture something natural about the world. The syntactic structure-building operation of MERGE is thought to offer the scientific community a “real joint of nature”, “a (new) aspect of nature” (Mukherji 2010), not merely a formal artefact. This mathematical theory of language is then seen as being able to offer biologists, geneticists and neuroscientists clearer instructions for how to explore language. The argument of this chapter proceeds in four steps. First, I clarify the object of inquiry for biolinguistics: not speech, communication, or generic sequence processing, but the internal computational system that generates hierarchically structured expressions. Second, I argue that this formal characterization matters for evolutionary explanation, because different conceptions of syntax imply different standards of what must be explained. Third, I suggest that a sufficiently explicit algebraic account of syntax places non-trivial constraints on candidate neural mechanisms. Finally, I consider how recent neurocomputational work begins to transform these constraints into empirically tractable hypotheses, while also noting the speculative and revisable character of the present program.

Method: Proposes asymmetric actor-critic framework: powerful proprietary LLM as actor, smaller open-source critic provides runtime supervision by monitoring actor’s actions and intervening within the same interaction trajectory. Uses generation-verification asymmetry principle. Includes data generation pipeline for critic fine-tuning without modifying actor.

Result: Experiments on τ-bench and UserBench show significant improvements in reliability and task success over strong single-agent baselines. Lightweight open-source critics rival or surpass larger proprietary models in critic role. Critic fine-tuning yields additional gains over state-of-the-art methods.

Conclusion: The asymmetric actor-critic framework enables reliable conversational agents in one-shot settings by leveraging generation-verification asymmetry, allowing smaller critics to effectively supervise larger proprietary actors without requiring model modifications or additional attempts.

Abstract: Large language models (LLMs) exhibit strong reasoning and conversational abilities, but ensuring reliable behavior in multi-turn interactions remains challenging. In many real-world applications, agents must succeed in one-shot settings where retries are impossible. Existing approaches either rely on reflection or post-hoc evaluation, which require additional attempts, or assume fully trainable models that cannot leverage proprietary LLMs. We propose an asymmetric actor-critic framework for reliable conversational agents. A powerful proprietary LLM acts as the actor, while a smaller open-source critic provides runtime supervision, monitoring the actor’s actions and intervening within the same interaction trajectory. Unlike training-based actor-critic methods, our framework supervises a fixed actor operating in open-ended conversational environments. The design leverages a generation-verification asymmetry: while high-quality generation requires large models, effective oversight can often be achieved by smaller ones. We further introduce a data generation pipeline that produces supervision signals for critic fine-tuning without modifying the actor. Experiments on $τ$-bench and UserBench show that our approach significantly improves reliability and task success over strong single-agent baselines. Moreover, lightweight open-source critics rival or surpass larger proprietary models in the critic role, and critic fine-tuning yields additional gains over several state-of-the-art methods.

Method: The paper presents a lifecycle-oriented analysis framework (Abuse Detection Lifecycle - ADL) with four stages. For each stage, the authors synthesize emerging research and industry practices, examine architectural considerations for production deployment, and analyze strengths/limitations of LLM-driven approaches.

Result: The survey provides a comprehensive framework for understanding LLM integration in abuse detection systems, highlighting current applications across the lifecycle while identifying key challenges including latency, cost-efficiency, determinism, adversarial robustness, and fairness.

Conclusion: LLMs show promise for enhancing abuse detection systems through contextual reasoning and cross-modal understanding, but operational challenges need to be addressed to make them reliable, accountable components of large-scale safety systems. Future research should focus on overcoming these limitations.

Abstract: Online abuse has grown increasingly complex, spanning toxic language, harassment, manipulation, and fraudulent behavior. Traditional machine-learning approaches dependent on static classifiers and labor-intensive labeling struggle to keep pace with evolving threat patterns and nuanced policy requirements. Large Language Models introduce new capabilities for contextual reasoning, policy interpretation, explanation generation, and cross-modal understanding, enabling them to support multiple stages of modern safety systems. This survey provides a lifecycle-oriented analysis of how LLMs are being integrated into the Abuse Detection Lifecycle (ADL), which we define across four stages: (I) Label & Feature Generation, (II) Detection, (III) Review & Appeals, and (IV) Auditing & Governance. For each stage, we synthesize emerging research and industry practices, highlight architectural considerations for production deployment, and examine the strengths and limitations of LLM-driven approaches. We conclude by outlining key challenges including latency, cost-efficiency, determinism, adversarial robustness, and fairness and discuss future research directions needed to operationalize LLMs as reliable, accountable components of large-scale abuse-detection and governance systems.

Method: Proposes Agent Q-Mix, a reinforcement learning framework that reformulates topology selection as cooperative MARL. Uses QMIX value factorization for decentralized communication decisions, where each agent selects communication actions that jointly induce round-wise communication graphs. Combines topology-aware GNN encoder, GRU memory, and per-agent Q-heads under CTDE paradigm, optimizing a reward function balancing task accuracy with token cost.

Result: Achieves highest average accuracy across seven benchmarks in coding, reasoning, and mathematics compared to existing methods, with superior token efficiency and robustness against agent failure. On Humanity’s Last Exam using Gemini-3.1-Flash-Lite, achieves 20.8% accuracy, outperforming Microsoft Agent Framework (19.2%), LangGraph (19.2%), AutoGen, and Lobster by OpenClaw.

Conclusion: Agent Q-Mix demonstrates the effectiveness of learned, decentralized topology optimization for pushing boundaries in multi-agent reasoning, showing that reinforcement learning can effectively optimize communication structures between LLM agents for complex problem-solving.

Abstract: Large Language Models (LLMs) have shown remarkable performance in completing various tasks. However, solving complex problems often requires the coordination of multiple agents, raising a fundamental question: how to effectively select and interconnect these agents. In this paper, we propose \textbf{Agent Q-Mix}, a reinforcement learning framework that reformulates topology selection as a cooperative Multi-Agent Reinforcement Learning (MARL) problem. Our method learns decentralized communication decisions using QMIX value factorization, where each agent selects from a set of communication actions that jointly induce a round-wise communication graph. At its core, Agent Q-Mix combines a topology-aware GNN encoder, GRU memory, and per-agent Q-heads under a Centralized Training with Decentralized Execution (CTDE) paradigm. The framework optimizes a reward function that balances task accuracy with token cost. Across seven core benchmarks in coding, reasoning, and mathematics, Agent Q-Mix achieves the highest average accuracy compared to existing methods while demonstrating superior token efficiency and robustness against agent failure. Notably, on the challenging Humanity’s Last Exam (HLE) using Gemini-3.1-Flash-Lite as a backbone, Agent Q-Mix achieves 20.8% accuracy, outperforming Microsoft Agent Framework (19.2%) and LangGraph (19.2%), followed by AutoGen and Lobster by OpenClaw. These results underscore the effectiveness of learned, decentralized topology optimization in pushing the boundaries of multi-agent reasoning.

Method: The authors provide a unified explanation of the quality-exploration dilemma, showing low-confidence remasking improves a myopic proxy for quality while constraining entropy. They characterize the optimal distribution balancing quality and exploration, and develop an Independent Metropolis-Hastings sampler targeting this distribution during decoding.

Result: Experiments across reasoning benchmarks (MATH500, AIME24/25, HumanEval, MBPP) show the approach yields better exploration-quality tradeoff than both random and low-confidence remasking.

Conclusion: The proposed method effectively addresses the quality-exploration dilemma in diffusion LLMs by optimizing the balance between generation quality and exploration of reasoning paths.

Abstract: Diffusion large language models (dLLMs) theoretically permit token decoding in arbitrary order, a flexibility that could enable richer exploration of reasoning paths than autoregressive (AR) LLMs. In practice, however, random-order decoding often hurts generation quality. To mitigate this, low-confidence remasking improves single-sample quality (e.g., Pass@$1$) by prioritizing confident tokens, but it also suppresses exploration and limits multi-sample gains (e.g., Pass@$k$), creating a fundamental quality–exploration dilemma. In this paper, we provide a unified explanation of this dilemma. We show that low-confidence remasking improves a myopic proxy for quality while provably constraining the entropy of the induced sequence distribution. To overcome this limitation, we characterize the optimal distribution that explicitly balances quality and exploration, and develop a simple Independent Metropolis–Hastings sampler that approximately targets this distribution during decoding. Experiments across a range of reasoning benchmarks including MATH500, AIME24/25, HumanEval, and MBPP show that our approach yields better exploration-quality tradeoff than both random and low-confidence remasking.

Method: TR-ICRL retrieves relevant instances from unlabeled evaluation data, generates candidate answers via LLM, creates pseudo-labels through majority voting, uses these as proxy rewards to provide formative feedback, and iteratively refines outputs. Final answer is determined through another round of majority voting.

Result: Significant performance improvements: 21.23% average improvement on MedQA and 137.59% on AIME2024 for Qwen2.5-7B. Extensive ablation studies validate the approach’s effectiveness and robustness.

Conclusion: TR-ICRL effectively addresses the reward estimation problem in ICRL through test-time rethinking with pseudo-labels, enabling LLMs to learn from in-context feedback without ground-truth supervision during inference.

Abstract: In-Context Reinforcement Learning (ICRL) enables Large Language Models (LLMs) to learn online from external rewards directly within the context window. However, a central challenge in ICRL is reward estimation, as models typically lack access to ground-truths during inference. To address this limitation, we propose Test-Time Rethinking for In-Context Reinforcement Learning (TR-ICRL), a novel ICRL framework designed for both reasoning and knowledge-intensive tasks. TR-ICRL operates by first retrieving the most relevant instances from an unlabeled evaluation set for a given query. During each ICRL iteration, LLM generates a set of candidate answers for every retrieved instance. Next, a pseudo-label is derived from this set through majority voting. This label then serves as a proxy to give reward messages and generate formative feedbacks, guiding LLM through iterative refinement. In the end, this synthesized contextual information is integrated with the original query to form a comprehensive prompt, with the answer determining through a final round of majority voting. TR-ICRL is evaluated on mainstream reasoning and knowledge-intensive tasks, where it demonstrates significant performance gains. Remarkably, TR-ICRL improves Qwen2.5-7B by 21.23% on average on MedQA and even 137.59% on AIME2024. Extensive ablation studies and analyses further validate the effectiveness and robustness of our approach. Our code is available at https://github.com/pangpang-xuan/TR_ICRL.

Method: Uses a 2x2 factorial decomposition to separate lexical vs. semantic contributions, analyzes models across 110M-70B parameters, examines sparse autoencoders, and tests downstream task impacts through filtering experiments.

Result: Lexical-only condition (same word, different meaning) consistently exceeds semantic-only condition (different word, same meaning) across all model sizes; 18-36% of SAE features blend senses; confound sits in ≤1% of dimensions; filtering improves word sense disambiguation and makes knowledge edits more selective.

Conclusion: Lexical confounds significantly contribute to apparent superposition effects, affecting interpretability and downstream performance; addressing these confounds improves model understanding and task performance.

Abstract: If the same neuron activates for both “lender” and “riverside,” standard metrics attribute the overlap to superposition–the neuron must be compressing two unrelated concepts. This work explores how much of the overlap is due a lexical confound: neurons fire for a shared word form (such as “bank”) rather than for two compressed concepts. A 2x2 factorial decomposition reveals that the lexical-only condition (same word, different meaning) consistently exceeds the semantic-only condition (different word, same meaning) across models spanning 110M-70B parameters. The confound carries into sparse autoencoders (18-36% of features blend senses), sits in <=1% of activation dimensions, and hurts downstream tasks: filtering it out improves word sense disambiguation and makes knowledge edits more selective (p = 0.002).

Method: Proposes Multimodal Depth Upscaling: insert new transformer layers into frozen text LLM, train only added layers on speech data. Tested with SmolLM2 models on 48k hours of English ASR data. Also incorporates E-Branchformer architecture for speech recognition as inserted layers.

Result: Depth upscaling achieves ASR comparable to full fine-tuning with far less text degradation than both full fine-tuning and LoRA. E-Branchformer inserted layers match or surpass full fine-tuning ASR on larger model while reducing text degradation by over 75% with 60% fewer trainable parameters.

Conclusion: Multimodal Depth Upscaling effectively adapts text LLMs to speech while preserving text capabilities, with speech-specific architectures further improving performance and efficiency.

Abstract: Adapting pre-trained text Large Language Models (LLMs) into Speech Language Models (Speech LMs) via continual pretraining on speech data is promising, but often degrades the original text capabilities. We propose Multimodal Depth Upscaling, an extension of an emerging strategy in continual LLM pre-training, where new transformer layers are inserted into a frozen text LLM and only the added layers are trained on speech data. Experiments with SmolLM2-360M and SmolLM2-1.7B on 48k hours of English Automatic Speech Recognition (ASR) data show that depth up-scaling achieves ASR comparable to full fine-tuning while causing far less text degradation than both full fine-tuning and Low-Rank Adaptation (LoRA). We further show that incorporating E-Branchformer, an architecture designed for speech recognition, as the inserted layers achieves ASR that matches or surpasses full fine-tuning on the larger model while reducing text degradation by over 75% with 60% fewer trainable parameters.

Method: Proposes evaluating synthetic data by its training utility on target models using gradient-based influence estimation to quantify each sample’s contribution. Uses an optimization-based framework that adapts rubrics using target-model feedback, with lightweight guiding text and rubric-specialized models to generate task-conditioned rubrics. Optimizes rubric generator with reinforcement learning using influence scores as rewards.

Result: Experiments across domains, target models, and data generators show consistent improvements and strong generalization without task-specific tuning. Analysis reveals synthetic and real samples can have substantially different influence on learning even when close in embedding space.

Conclusion: The proposed framework effectively optimizes synthetic data generation by using target-model feedback through influence estimation, addressing limitations of handcrafted rubrics and enabling better downstream performance in knowledge-intensive domains.

Abstract: Large language models (LLMs) achieve strong downstream performance largely due to abundant supervised fine-tuning (SFT) data. However, high-quality SFT data in knowledge-intensive domains such as humanities, social sciences, medicine, law, and finance is scarce because expert curation is expensive, privacy constraints are strict, and label consistency is hard to ensure. Recent work uses synthetic data, typically by prompting a generator over domain documents and filtering outputs with handcrafted rubrics. Yet rubric design is expert-dependent, transfers poorly across domains, and is often optimized through a brittle heuristic loop of writing rubrics, synthesizing data, training, inspecting results, and manually guessing revisions. This process lacks reliable quantitative feedback about how a rubric affects downstream performance. We propose evaluating synthetic data by its training utility on the target model and using this signal to guide data generation. Inspired by influence estimation, we adopt an optimizer-aware estimator that uses gradient information to quantify each synthetic sample’s contribution to a target model’s objective on specific tasks. Our analysis shows that even when synthetic and real samples are close in embedding space, their influence on learning can differ substantially. Based on this insight, we propose an optimization-based framework that adapts rubrics using target-model feedback. We provide lightweight guiding text and use a rubric-specialized model to generate task-conditioned rubrics. Influence score is used as the reward to optimize the rubric generator with reinforcement learning. Experiments across domains, target models, and data generators show consistent improvements and strong generalization without task-specific tuning.

Method: Constructed JUBAKU-v2 dataset with 216 examples based on attribution theory from social psychology, evaluating biases in attributing behaviors to in-groups vs out-groups within reasoning while keeping conclusions fixed.

Result: Experimental results show JUBAKU-v2 can detect performance differences across models more sensitively than existing benchmarks, revealing attribution biases in reasoning.

Conclusion: The study introduces a culturally-grounded Japanese bias evaluation dataset that captures subtle attribution biases in reasoning, providing more sensitive assessment of LLM fairness in Japanese cultural context.

Abstract: In enhancing the fairness of Large Language Models (LLMs), evaluating social biases rooted in the cultural contexts of specific linguistic regions is essential. However, most existing Japanese benchmarks heavily rely on translating English data, which does not necessarily provide an evaluation suitable for Japanese culture. Furthermore, they only evaluate bias in the conclusion, failing to capture biases lurking in the reasoning. In this study, based on attribution theory in social psychology, we constructed a new dataset, ``JUBAKU-v2,’’ which evaluates the bias in attributing behaviors to in-groups and out-groups within reasoning while fixing the conclusion. This dataset consists of 216 examples reflecting cultural biases specific to Japan. Experimental results verified that it can detect performance differences across models more sensitively than existing benchmarks.

Method: Finetune a quantized 1.7B parameter small language model on limited human-annotated data using a custom multi-dimensional rubric framework with simple augmentation and regularization techniques, implementing efficient 4-bit quantized fine-tuning.

Result: The approach achieves 0.23 points higher Krippendorff’s α (inter-annotator agreement) than the best proprietary LLM and demonstrates generalizability on separate emotion classification tasks, showing superior alignment with human expert consensus.

Conclusion: Task-specific alignment through efficient quantized fine-tuning of small language models provides a superior open-source alternative to proprietary models for evaluation and annotation tasks, addressing biases, reproducibility, and privacy issues.

Abstract: As Large Language Model (LLM) capabilities advance, the demand for high-quality annotation of exponentially increasing text corpora has outpaced human capacity, leading to the widespread adoption of LLMs in automatic evaluation and annotation. However, proprietary LLMs often exhibit systematic biases that diverge from human expert consensus, lacks reproducibility, and raises data privacy concerns. Our work examines the viability of finetuning a quantized Small Language Model of 1.7B parameter size on limited human-annotated data to serve as a highly aligned, deterministic evaluator and annotator. By implementing a custom, multi-dimensional rubric framework and simple augmentation and regularization techniques, the proposed approach achieves higher inter-annotator agreement (0.23 points increase in Krippendorff’s $α$) than the best performing state-of-the-art proprietary LLM. We also demonstrate the generalizability of the proposed training pipeline on a separate emotion classification task. The results show that task-specific alignment and efficient 4-bit quantized fine-tuning provide superior open-source alternative to using proprietary models for evaluation and annotation. Our finetuning approach is publicly available at https://github.com/jylee-k/slm-judge.

Method: Proposes chain-of-thought ASR (CoT-ASR) that constructs reasoning chains for contextual analysis before transcription, introduces CTC-guided Modality Adapter to align speech encoder outputs with LLM’s textual latent space using CTC non-blank token probabilities, and supports both self-generated and user-guided contextual reasoning.

Result: Achieves 8.7% relative reduction in word error rate (WER) and 16.9% relative reduction in entity error rate (EER) compared to standard LLM-based ASR.

Conclusion: CoT-ASR effectively leverages LLMs’ contextual understanding for speech recognition through reasoning chains, bridging modality gaps and enabling more accurate transcription with contextual awareness.

Abstract: Despite extensions to speech inputs, effectively leveraging the rich knowledge and contextual understanding of large language models (LLMs) in automatic speech recognition (ASR) remains non-trivial, as the task primarily involves direct speech-to-text mapping. To address this, this paper proposes chain-of-thought ASR (CoT-ASR), which constructs a reasoning chain that enables LLMs to first analyze the input speech and generate contextual analysis, thereby fully exploiting their generative capabilities. With this contextual reasoning, CoT-ASR then performs more informed speech recognition and completes both reasoning and transcription in a single pass. Moreover, CoT-ASR naturally supports user-guided transcription: while designed to self-generate reasoning, it can also seamlessly incorporate user-provided context to guide transcription, further extending ASR functionality. To reduce the modality gap, this paper introduces a CTC-guided Modality Adapter, which uses CTC non-blank token probabilities to weight LLM embeddings, efficiently aligning speech encoder outputs with the LLM’s textual latent space. Experiments show that, compared to standard LLM-based ASR, CoT-ASR achieves a relative reduction of 8.7% in word error rate (WER) and 16.9% in entity error rate (EER).

Method: Created KUTED dataset from TED/TEDx talks (91K sentence pairs, 170h English audio). Evaluated on S2TT task, identified orthographic variation issues, proposed systematic text standardization approach. Tested fine-tuned Seamless model, Transformer from scratch, and cascaded system (Seamless ASR + NLLB MT).

Result: Fine-tuned Seamless model achieved 15.18 BLEU on test set; improved Seamless baseline by 3.0 BLEU on FLEURS benchmark. Text standardization yielded substantial performance gains and more consistent translations.

Conclusion: KUTED enables speech-to-text translation for Central Kurdish; orthographic variation significantly impacts translation quality; systematic text standardization effectively addresses this issue and improves performance.

Abstract: We present KUTED, a speech-to-text translation (S2TT) dataset for Central Kurdish, derived from TED and TEDx talks. The corpus comprises 91,000 sentence pairs, including 170 hours of English audio, 1.65 million English tokens, and 1.40 million Central Kurdish tokens. We evaluate KUTED on the S2TT task and find that orthographic variation significantly degrades Kurdish translation performance, producing nonstandard outputs. To address this, we propose a systematic text standardization approach that yields substantial performance gains and more consistent translations. On a test set separated from TED talks, a fine-tuned Seamless model achieves 15.18 BLEU, and we improve Seamless baseline by 3.0 BLEU on the FLEURS benchmark. We also train a Transformer model from scratch and evaluate a cascaded system that combines Seamless (ASR) with NLLB (MT).

Method: TRIMS (Trajectory-Ranked Instruction Masked Supervision) is a trajectory-guided supervised fine-tuning framework that injects trajectory supervision into standard Masked Diffusion Language Model training. It uses lightweight signals from an autoregressive teacher to guide a trajectory-aware masking strategy, encouraging the model to learn more effective decoding orders without costly DLM-based distillation.

Result: Experiments on LLaDA and Dream across math and coding benchmarks show TRIMS significantly improves the accuracy-parallelism trade-off over standard MDLM training and train-free acceleration baselines, achieving competitive performance with prior distillation-based approaches at substantially lower training cost. Analysis confirms TRIMS leads to better decoding trajectories.

Conclusion: TRIMS demonstrates that trajectory-guided supervision effectively improves diffusion language model decoding efficiency, offering a practical solution to the train-inference mismatch problem in parallel decoding while maintaining competitive performance with reduced training costs.

Abstract: Diffusion language models (DLMs) offer a promising path toward low-latency generation through parallel decoding, but their practical efficiency depends heavily on the decoding trajectory. In practice, this advantage often fails to fully materialize because standard training does not provide explicit supervision over token reveal order, creating a train-inference mismatch that leads to suboptimal decoding behavior. We propose Trajectory-Ranked Instruction Masked Supervision (TRIMS), a simple trajectory-guided supervised fine-tuning framework that injects trajectory supervision into standard Masked Diffusion Language Model (MDLM) training with minimal overhead. Instead of relying on costly DLM-based distillation, TRIMS uses lightweight signals from an autoregressive teacher to guide a trajectory-aware masking strategy, encouraging the model to learn more effective decoding orders. Experiments on LLaDA and Dream across math and coding benchmarks show that TRIMS significantly improves the accuracy-parallelism trade-off over both standard MDLM training and train-free acceleration baselines, while achieving competitive performance with prior distillation-based approaches at substantially lower training cost. Further analysis shows that TRIMS leads to better decoding trajectories, validating the effectiveness of trajectory-guided supervision for DLMs.

Method: Develops a penalized likelihood-ratio test framework where the alternative hypothesis models document collections using beta-binomial distributions with gamma penalty on precision parameter to capture word burstiness, while the null hypothesis uses binomial distributions that ignore burstiness.

Result: The term-weighting scheme derived from this test statistic performs comparably to TF-IDF on document classification tasks, providing statistical validation for TF-IDF’s effectiveness.

Conclusion: The paper provides statistical insights into TF-IDF and demonstrates the potential of hypothesis testing frameworks for advancing term-weighting scheme development beyond traditional empirical approaches.

Abstract: TF-IDF is a classical formula that is widely used for identifying important terms within documents. We show that TF-IDF-like scores arise naturally from the test statistic of a penalized likelihood-ratio test setup capturing word burstiness (also known as word over-dispersion). In our framework, the alternative hypothesis captures word burstiness by modeling a collection of documents according to a family of beta-binomial distributions with a gamma penalty term on the precision parameter. In contrast, the null hypothesis assumes that words are binomially distributed in collection documents, a modeling approach that fails to account for word burstiness. We find that a term-weighting scheme given rise to by this test statistic performs comparably to TF-IDF on document classification tasks. This paper provides insights into TF-IDF from a statistical perspective and underscores the potential of hypothesis testing frameworks for advancing term-weighting scheme development.

Method: Created AfrIFact dataset covering automatic fact-checking pipeline steps (information retrieval, evidence extraction, fact checking) in 10 African languages and English. Evaluated embedding models for cross-lingual retrieval and LLMs for fact verification.

Result: Best embedding models lack cross-lingual retrieval capabilities; cultural/news documents easier to retrieve than healthcare documents. LLMs lack robust multilingual fact-verification in African languages, but few-shot prompting improves performance by up to 43% and task-specific fine-tuning by up to 26%.

Conclusion: The AfrIFact dataset encourages work on low-resource information retrieval, evidence retrieval, and fact checking, highlighting challenges and opportunities for improving multilingual fact verification in African languages.

Abstract: Assessing the veracity of a claim made online is a complex and important task with real-world implications. When these claims are directed at communities with limited access to information and the content concerns issues such as healthcare and culture, the consequences intensify, especially in low-resource languages. In this work, we introduce AfrIFact, a dataset that covers the necessary steps for automatic fact-checking (i.e., information retrieval, evidence extraction, and fact checking), in ten African languages and English. Our evaluation results show that even the best embedding models lack cross-lingual retrieval capabilities, and that cultural and news documents are easier to retrieve than healthcare-domain documents, both in large corpora and in single documents. We show that LLMs lack robust multilingual fact-verification capabilities in African languages, while few-shot prompting improves performance by up to 43% in AfriqueQwen-14B, and task-specific fine-tuning further improves fact-checking accuracy by up to 26%. These findings, along with our release of the AfrIFact dataset, encourage work on low-resource information retrieval, evidence retrieval, and fact checking.

Method: Trained OLMo-2-based LMs (30M to 3B parameters) on up to 100B tokens of DCLM data while varying pretraining data scale (1-150x parameters) and retrieval store size (1-20x), evaluated across reasoning, scientific QA, and open-domain QA benchmarks.

Result: Retrieval consistently improves performance over parametric-only baselines across all model scales. Developed a three-dimensional scaling framework modeling performance as function of model size, pretraining tokens, and retrieval corpus size, enabling estimation of optimal data allocations.

Conclusion: The marginal utility of retrieval depends strongly on model scale, task type, and pretraining saturation. Provides quantitative foundation for understanding when/how retrieval should complement pretraining and practical guidance for scalable LM system design.

Abstract: Retrieval-augmented generation (RAG) improves language model (LM) performance by providing relevant context at test time for knowledge-intensive situations. However, the relationship between parametric knowledge acquired during pretraining and non-parametric knowledge accessed via retrieval remains poorly understood, especially under fixed data budgets. In this work, we systematically study the trade-off between pretraining corpus size and retrieval store size across a wide range of model and data scales. We train OLMo-2-based LMs ranging from 30M to 3B parameters on up to 100B tokens of DCLM data, while varying both pretraining data scale (1-150x the number of parameters) and retrieval store size (1-20x), and evaluate performance across a diverse suite of benchmarks spanning reasoning, scientific QA, and open-domain QA. We find that retrieval consistently improves performance over parametric-only baselines across model scales and introduce a three-dimensional scaling framework that models performance as a function of model size, pretraining tokens, and retrieval corpus size. This scaling manifold enables us to estimate optimal allocations of a fixed data budget between pretraining and retrieval, revealing that the marginal utility of retrieval depends strongly on model scale, task type, and the degree of pretraining saturation. Our results provide a quantitative foundation for understanding when and how retrieval should complement pretraining, offering practical guidance for allocating data resources in the design of scalable language modeling systems.

Method: LangMARL introduces agent-level language credit assignment, pioneers gradient evolution in language space for policy improvement, and summarizes task-relevant causal relations from replayed trajectories to provide dense feedback and improve convergence under sparse rewards.

Result: Extensive experiments across diverse cooperative multi-agent tasks demonstrate improved sample efficiency, interpretability, and strong generalization.

Conclusion: The framework successfully bridges classical MARL techniques with language models to address coordination challenges in multi-agent LLM systems.

Abstract: Large language model (LLM) agents struggle to autonomously evolve coordination strategies in dynamic environments, largely because coarse global outcomes obscure the causal signals needed for local policy refinement. We identify this bottleneck as a multi-agent credit assignment problem, which has long been studied in classical multi-agent reinforcement learning (MARL) but remains underaddressed in LLM-based systems. Building on this observation, we propose LangMARL, a framework that brings credit assignment and policy gradient evolution from cooperative MARL into the language space. LangMARL introduces agent-level language credit assignment, pioneers gradient evolution in language space for policy improvement, and summarizes task-relevant causal relations from replayed trajectories to provide dense feedback and improve convergence under sparse rewards. Extensive experiments across diverse cooperative multi-agent tasks demonstrate improved sample efficiency, interpretability, and strong generalization.

Method: Proposes Stochastic Attention (SA) as a drop-in enhancement for sliding-window attention: 1) Apply random permutation to token sequence before windowed attention, 2) Perform standard sliding-window attention on permuted sequence, 3) Restore original token order. Through multiple layers with independently sampled permutations, receptive fields grow exponentially, achieving full sequence coverage in O(log_w n) layers vs O(n/w) for standard SWA.

Result: SA shows strong performance in two settings: 1) Pre-training language models from scratch where gated SA+SWA combination achieves best average zero-shot accuracy, 2) Training-free inference on Qwen3-8B and Qwen3-30B-A3B where SA consistently outperforms SWA and matches or exceeds Mixture of Block Attention at comparable compute budgets.

Conclusion: Connectome-inspired stochastic routing is a practical primitive for improving efficient attention expressivity, complementary to existing linear and sparse approaches. The biological inspiration from fruit fly brain connectivity provides a novel perspective on transformer architecture design.

Abstract: The whole-brain connectome of a fruit fly comprises over 130K neurons connected with a probability of merely 0.02%, yet achieves an average shortest path of only 4.4 hops. Despite being highly structured at the circuit level, the network’s long-range connections are broadly distributed across brain regions, functioning as stochastic shortcuts that enable efficient global communication. Inspired by this observation, we propose Stochastic Attention (SA), a drop-in enhancement for sliding-window attention (SWA) that applies a random permutation to the token sequence before windowed attention and restores the original order afterward. This transforms the fixed local window into a stochastic global one within the same $O(nw)$ per-layer budget. Through depth, independently sampled permutations yield exponentially growing receptive fields, achieving full sequence coverage in $O(\log_w n)$ layers versus $O(n/w)$ for SWA. We validate SA in two settings: pre-training language models from scratch, where a gated SA + SWA combination achieves the best average zero-shot accuracy, and training-free inference on Qwen3-8B and Qwen3-30B-A3B, where SA consistently outperforms SWA and matches or exceeds Mixture of Block Attention at comparable compute budgets. These results suggest that connectome-inspired stochastic routing is a practical primitive for improving the expressivity of efficient attention, complementary to existing linear and sparse approaches.

Method: Established classical and encoder baselines, then examined parameter-efficient supervised fine-tuning with LoRA/QLoRA under multiple objective and optimization settings. Finally evaluated preference-based optimization with DPO, ORPO, and KTO, including class-rebalanced training. Focused on methodological insights across objective formulation, adapter choice, optimizer behavior, context windowing, and class-balance interventions.

Result: Optimization effects are highly method-dependent: some approaches deliver stable, transferable gains while others are sensitive to configuration and data balance. Preference optimization exhibits large variation across objectives, showing that method selection is more consequential than simply adding a preference-training stage.

Conclusion: Provides clear optimization narrative for mental health NLP: start from transparent baselines, apply controlled tuning, and use preference optimization selectively where gains are demonstrable. Offers reproducible, practically grounded framework for choosing effective training strategies beyond architecture choice alone.

Abstract: Mental health text classification has rapidly adopted modern adaptation methods, yet practical guidance on which optimization strategy to use, when, and why remains limited. This paper presents a systematic comparative study of optimization pathways for a joint mental-health classification task, moving from strong vanilla baselines to progressively more specialized techniques. We first establish classical and encoder references, then examine parameter-efficient supervised fine-tuning with LoRA/QLoRA under multiple objective and optimization settings, and finally evaluate preference-based optimization with DPO, ORPO, and KTO, including class-rebalanced training. Rather than emphasizing a single headline score, we focus on methodological insight: how performance changes with objective formulation, adapter choice, optimizer behavior, context windowing, and class-balance intervention. The results show that optimization effects are highly method-dependent: some approaches deliver stable, transferable gains, while others are sensitive to configuration and data balance. Preference optimization, in particular, exhibits large variation across objectives, indicating that method selection is more consequential than simply adding a preference-training stage. The central contribution is a clear optimization narrative for mental health NLP: start from transparent baselines, apply controlled tuning, and use preference optimization selectively where its gains are demonstrable. This provides a reproducible and practically grounded framework for choosing effective training strategies beyond architecture choice alone.

Method: Used character counting as a minimal probe, combined mechanistic analysis including probing classifiers, activation patching, logit lens analysis, and attention head tracing across models like LLaMA, Qwen, and Gemma.

Result: Found that character-level information is encoded in early/mid layers but attenuated by negative circuits in later layers (especially penultimate/final MLPs), causing models to compute correct answers internally but fail to output them.

Conclusion: Symbolic reasoning failures arise from structured interference within computation graphs, not missing representations, revealing LLMs implement competitive decoding where correct/incorrect hypotheses coexist and final outputs are determined by suppression.

Abstract: Large language models (LLMs) exhibit failures on elementary symbolic tasks such as character counting in a word, despite excelling on complex benchmarks. Although this limitation has been noted, the internal reasons remain unclear. We use character counting (e.g., “How many p’s are in apple?”) as a minimal, controlled probe that isolates token-level reasoning from higher-level confounds. Using this setting, we uncover a consistent phenomenon across modern architectures, including LLaMA, Qwen, and Gemma: models often compute the correct answer internally yet fail to express it at the output layer. Through mechanistic analysis combining probing classifiers, activation patching, logit lens analysis, and attention head tracing, we show that character-level information is encoded in early and mid-layer representations. However, this information is attenuated by a small set of components in later layers, especially the penultimate and final layer MLP. We identify these components as negative circuits: subnetworks that downweight correct signals in favor of higher-probability but incorrect outputs. Our results lead to two contributions. First, we show that symbolic reasoning failures in LLMs are not due to missing representations or insufficient scale, but arise from structured interference within the model’s computation graph. This explains why such errors persist and can worsen under scaling and instruction tuning. Second, we provide evidence that LLM forward passes implement a form of competitive decoding, in which correct and incorrect hypotheses coexist and are dynamically reweighted, with final outputs determined by suppression as much as by amplification. These findings carry implications for interpretability and robustness: simple symbolic reasoning exposes weaknesses in modern LLMs, underscoring need for design strategies that ensure information is encoded and reliably used.

Method: Examines the permissible and restricted combinations of various suffixes with roots or verbal stems in Mapuche verb forms, using the language’s morphotactics for valency classification

Result: Findings are incorporated into the Dungupeyum morphological analyzer system to improve its accuracy and contribute to understanding Mapuche verb valency issues

Conclusion: The research improves morphological analysis tools while contributing theoretical insights into valency classification in Mapudungun verb forms

Abstract: In the previous work, a lexical (re)categorisation – or confirmation of the given category – of roots identified as verbal was undertaken to determine their original category accurately. Building on this, the present paper offers an account of the valency classification of those Mapudungun roots confirmed to be verbal, using the language’s own morphotactics; specifically, by examining the permissible and restricted combinations of various suffixes with roots or verbal stems in the Mapuche verb form. As with all work conducted thus far, the results presented here aim to improve the morphological analyser (Dungupeyum) with all verified findings incorporated into the system. From a theoretical perspective, we also hope to contribute to the recognition and understanding of issues related to the valency of Mapuche verb forms.

Method: Created EmoScene benchmark with 4,731 context-rich scenarios annotated with 8-dimensional emotion vectors based on Plutchik’s basic emotions. Evaluated 6 instruction-tuned LLMs zero-shot, then proposed entanglement-aware Bayesian inference framework using emotion co-occurrence statistics for joint posterior inference.

Result: Best model achieved Macro F1 of 0.501, showing difficulty of context-aware multi-label emotion prediction. Bayesian inference improved structural consistency and performance, with +0.051 Macro F1 gain for weaker models like Qwen2.5-7B.

Conclusion: EmoScene provides a challenging benchmark for studying multi-dimensional emotion understanding, revealing limitations of current LLMs and demonstrating that incorporating emotion co-occurrence statistics through Bayesian inference can improve predictions.

Abstract: Understanding emotions in natural language is inherently a multi-dimensional reasoning problem, where multiple affective signals interact through context, interpersonal relations, and situational cues. However, most existing emotion understanding benchmarks rely on short texts and predefined emotion labels, reducing this process to independent label prediction and ignoring the structured dependencies among emotions. To address this limitation, we introduce Emotional Scenarios (EmoScene), a theory-grounded benchmark of 4,731 context-rich scenarios annotated with an 8-dimensional emotion vector derived from Plutchik’s basic emotions. We evaluate six instruction-tuned large language models in a zero-shot setting and observe modest performance, with the best model achieving a Macro F1 of 0.501, highlighting the difficulty of context-aware multi-label emotion prediction. Motivated by the observation that emotions rarely occur independently, we further propose an entanglement-aware Bayesian inference framework that incorporates emotion co-occurrence statistics to perform joint posterior inference over the emotion vector. This lightweight post-processing improves structural consistency of predictions and yields notable gains for weaker models (e.g., +0.051 Macro F1 for Qwen2.5-7B). EmoScene therefore provides a challenging benchmark for studying multi-dimensional emotion understanding and the limitations of current language models.

Method: Organizes literature into three paradigms: 1) Prompting as plug-and-play (zero/few-shot tool use), 2) Supervised tool learning (fine-tuning with tool demonstrations), and 3) Reward-driven tool policy learning (reinforcement learning for tool selection). Analyzes methods, strengths, failure modes, and reviews evaluation landscape.

Result: Provides a comprehensive taxonomy and analysis of LLM tool-use methods, highlighting key challenges, evaluation metrics, and evolutionary patterns across the three paradigms.

Conclusion: The paper offers a structured framework for understanding LLM tool-use evolution, identifies research gaps, and provides guidance for future work in developing more reliable and effective tool-using agents.

Abstract: Large language models are increasingly being deployed as autonomous agents yet their real world effectiveness depends on reliable tools for information retrieval, computation and external action. Existing studies remain fragmented across tasks, tool types, and training settings, lacking a unified view of how tool-use methods differ and evolve. This paper organizes the literature into three paradigms: prompting as plug-and-play, supervised tool learning and reward-driven tool policy learning, analyzes their methods, strengths and failure modes, reviews the evaluation landscape and highlights key challenges, aiming to address this fragmentation and provide a more structured evolutionary view of agentic tool use.

Method: Proposes StanceMoE - a context-enhanced Mixture-of-Experts architecture built on fine-tuned BERT encoder. Integrates six expert modules capturing: global semantic orientation, salient lexical cues, clause-level focus, phrase-level patterns, framing indicators, and contrast-driven discourse shifts. Uses context-aware gating mechanism to dynamically weight expert contributions.

Result: Achieves macro-F1 score of 94.26% on StanceNakba 2026 Subtask A dataset (1,401 annotated English texts with implicit target actors), outperforming traditional baselines and alternative BERT-based variants.

Conclusion: The Mixture-of-Experts approach with specialized linguistic signal capture improves stance detection performance by adaptively modeling diverse stance-expressive patterns.

Abstract: Actor-level stance detection aims to determine an author expressed position toward specific geopolitical actors mentioned or implicated in a text. Although transformer-based models have achieved relatively good performance in stance classification, they typically rely on unified representations that may not sufficiently capture heterogeneous linguistic signals, such as contrastive discourse structures, framing cues, and salient lexical indicators. This motivates the need for adaptive architectures that explicitly model diverse stance-expressive patterns. In this paper, we propose StanceMoE, a context-enhanced Mixture-of-Experts (MoE) architecture built upon a fine-tuned BERT encoder for actor-level stance detection. Our model integrates six expert modules designed to capture complementary linguistic signals, including global semantic orientation, salient lexical cues, clause-level focus, phrase-level patterns, framing indicators, and contrast-driven discourse shifts. A context-aware gating mechanism dynamically weights expert contributions, enabling adaptive routing based on input characteristics. Experiments are conducted on the StanceNakba 2026 Subtask A dataset, comprising 1,401 annotated English texts where the target actor is implicit in the text. StanceMoE achieves a macro-F1 score of 94.26%, outperforming traditional baselines, and alternative BERT-based variants.

Method: Formalized three realistic interruption types (addition, revision, retraction), created InterruptBench from WebArena-Lite with high-quality interruption scenarios under semantic constraints, and evaluated six strong LLM backbones using a unified interruption simulation framework.

Result: Handling user interruptions effectively and efficiently during long-horizon agentic tasks remains challenging for powerful large-scale LLMs, even with state-of-the-art models.

Conclusion: Interruptible agents are essential for realistic deployment, and InterruptBench provides the first systematic study of this capability in environmentally grounded web navigation tasks, highlighting significant challenges for current LLMs.

Abstract: As LLM agents transition from short, static problem solving to executing complex, long-horizon tasks in dynamic environments, the ability to handle user interruptions, such as adding requirement or revising goals, during mid-task execution is becoming a core requirement for realistic deployment. However, existing benchmarks largely assume uninterrupted agent behavior or study interruptions only in short, unconstrained language tasks. In this paper, we present the first systematic study of interruptible agents in long-horizon, environmentally grounded web navigation tasks, where actions induce persistent state changes. We formalize three realistic interruption types, including addition, revision, and retraction, and introduce InterruptBench, a benchmark derived from WebArena-Lite that synthesizes high-quality interruption scenarios under strict semantic constraints. Using a unified interruption simulation framework, we evaluate six strong LLM backbones across single- and multi-turn interruption settings, analyzing both their effectiveness in adapting to updated intents and their efficiency in recovering from mid-task changes. Our results show that handling user interruptions effectively and efficiently during long-horizon agentic tasks remains challenging for powerful large-scale LLMs. Code and dataset are available at https://github.com/HenryPengZou/InterruptBench.

Method: Created a comprehensive corpus by curating existing datasets (Common Corpus, Common Crawl) and creating new Dutch-specific collections through organizational collaborations and synthetic augmentation, with careful compliance evaluation.

Result: Produced the largest permissively licensed Dutch corpus with 36B Dutch tokens plus 207B English, 232B code, and 48B German/Danish tokens, all available under CC-BY license on Hugging Face Hub.

Conclusion: The GPT-NL Public Corpus provides a valuable resource for developing lawful and effective Dutch language models, addressing licensing and compliance concerns while offering substantial multilingual content.

Abstract: We present the GPT-NL Public Corpus, the biggest permissively licensed corpus of Dutch language resources. The GPT-NL Public Corpus contains 21 Dutch-only collections totalling 36B preprocessed Dutch tokens not present in any other LLM pretraining corpus. Additionally, the corpus includes roughly 207B English, 232B Code, and 48B German/Danish tokens taken from existing sets which we further curated for compliance. This corpus includes curated data from large existing corpora like Common Corpus and Common Crawl, as well as newly created Dutch-specific collections. Most newly created Dutch collections consist of content collected in collaboration with organisations or synthetically augmented content. All data is collected and evaluated with the aim of facilitating the creation of (commercial) language models that are lawful, useful and non-harmful. All data included in the GPT-NL Public Corpus is sourced from datasets with permissive licensing and is curated and redistributed under a CC-BY license. The full dataset is publicly available on the Hugging Face Hub.

Method: Investigates training dynamics through cognitive specializations: language perception (input comprehension) and production (output generation). Uses layer ablation sweeps on decoder-only transformers for low-resource languages to identify specialized regions. Proposes CogSym heuristic for selective layer fine-tuning.

Result: Tuning only 25% outermost layers achieves downstream task performance within 2-3% deviation from full fine-tuning baseline. CogSym yields consistent performance with adapter methods like LoRA, showing generalization beyond full fine-tuning.

Conclusion: Provides insights into how LLMs learn new languages and enables more accessible, inclusive language modeling through efficient adaptation methods based on observed cognitive specialization patterns.

Abstract: Adapting large language models (LLMs) to new languages is an expensive and opaque process. Understanding how language models acquire new languages and multilingual abilities is key to achieve efficient adaptation. Prior work on multilingual interpretability research focuses primarily on how trained models process multilingual instructions, leaving unexplored the mechanisms through which they acquire new languages during training. We investigate these training dynamics on decoder-only transformers through the lens of two functional cognitive specializations: language perception (input comprehension) and production (output generation). Through experiments on low-resource languages, we demonstrate how perceptual and productive specialization emerges in different regions of a language model by running layer ablation sweeps from the model’s input and output directions. Based on the observed specialization patterns, we propose CogSym, a layer-wise heuristic that enables effective adaptation by exclusively fine-tuning a few early and late layers. We show that tuning only the 25% outermost layers achieves downstream task performance within 2-3% deviation from the full fine-tuning baseline. CogSym yields consistent performance with adapter methods such as LoRA, showcasing generalization beyond full fine-tuning. These findings provide insights to better understand how LLMs learn new languages and push toward accessible and inclusive language modeling.

Method: Constructed a random language model with short-range interactions belonging to context-sensitive grammars in the Chomsky hierarchy, allowing explicit reference to contexts. Numerically investigated statistical properties while ensuring context length remains constant relative to sentence length.

Result: Found that phase transitions occur even when the model refers only to contexts whose length remains constant with respect to sentence length, demonstrating finite-temperature phase transitions are not dependent on long-range interactions.

Conclusion: Phase transitions in language models are genuinely induced by the intrinsic nature of language itself, rather than being artifacts of long-range interactions, providing evidence for fundamental linguistic properties underlying these statistical mechanical phenomena.

Abstract: Since the random language model was proposed by E. DeGiuli [Phys. Rev. Lett. 122, 128301], language models have been investigated intensively from the viewpoint of statistical mechanics. Recently, the existence of a Berezinskii–Kosterlitz–Thouless transition was numerically demonstrated in models with long-range interactions between symbols. In statistical mechanics, it has long been known that long-range interactions can induce phase transitions. Therefore, it has remained unclear whether phase transitions observed in language models originate from genuinely linguistic properties that are absent in conventional spin models. In this study, we construct a random language model with short-range interactions and numerically investigate its statistical properties. Our model belongs to the class of context-sensitive grammars in the Chomsky hierarchy and allows explicit reference to contexts. We find that a phase transition occurs even when the model refers only to contexts whose length remains constant with respect to the sentence length. This result indicates that finite-temperature phase transitions in language models are genuinely induced by the intrinsic nature of language, rather than by long-range interactions.

Method: The authors propose the Dignified Peer framework with two components: anti-sycophancy/trustworthiness to counter servility, and empathy/creativity to mitigate evasiveness. They introduce the PersonaKnob dataset with compositional partial order structure of multiple persona preferences, use a tolerant constrained Lagrangian DPO algorithm to dynamically balance persona dimensions, and employ a psychometrically calibrated Item Response Theory evaluation protocol.

Result: Extensive empirical studies demonstrate that the approach successfully builds an LLM agent with both dignity and peer-like qualities, overcoming challenges in data supervision, objective collapse, and evaluation bias.

Conclusion: The Dignified Peer framework effectively addresses the dual failure mode of aligned language models, creating agents that maintain dignity while engaging as peers rather than servile or evasive assistants.

Abstract: Current aligned language models exhibit a dual failure mode we term the Evasive Servant: they sycophantically validate flawed user beliefs while deflecting responsibility with boilerplate disclaimers. We propose the Dignified Peer framework, which counters servility with anti-sycophancy and trustworthiness, and mitigates evasiveness through empathy and creativity. Realizing this agent requires overcoming significant challenges in data supervision, objective collapse, and evaluation bias. We address these issues by introducing the PersonaKnob dataset which features a compositional partial order structure of multiple persona preference. This data is utilized alongside a tolerant constrained Lagrangian DPO algorithm that dynamically balances all persona dimensions to prevent behavioral collapse. Additionally, we employ a psychometrically calibrated Item Response Theory evaluation protocol to disentangle latent model persona capability from confounders like judge biases. Extensive empirical studies demonstrate that our approach successfully build a LLM agent with both dignity and peer.

Method: Developed a multimodal pipeline combining automatic transcription, aspect-based sentiment analysis (ABSA), and semantic scene classification. Applied to over 2,300 conflict-related Shorts and 94,000+ visual frames from state-funded outlets covering Israel-Hamas war.

Result: Sentiment in transcripts varied across outlets and over time, while scene classifications reflected visual cues consistent with real-world events. Smaller domain-adapted models outperformed large transformers and LLMs for sentiment analysis.

Conclusion: The pipeline serves as a template for short-form platforms and demonstrates how multimodal methods with qualitative interpretation can characterize sentiment patterns and visual cues in algorithmically driven video environments.

Abstract: YouTube Shorts have become central to news consumption on the platform, yet research on how geopolitical events are represented in this format remains limited. To address this gap, we present a multimodal pipeline that combines automatic transcription, aspect-based sentiment analysis (ABSA), and semantic scene classification. The pipeline is first assessed for feasibility and then applied to analyze short-form coverage of the Israel-Hamas war by state-funded outlets. Using over 2,300 conflict-related Shorts and more than 94,000 visual frames, we systematically examine war reporting across major international broadcasters. Our findings reveal that the sentiment expressed in transcripts regarding specific aspects differs across outlets and over time, whereas scene-type classifications reflect visual cues consistent with real-world events. Notably, smaller domain-adapted models outperform large transformers and even LLMs for sentiment analysis, underscoring the value of resource-efficient approaches for humanities research. The pipeline serves as a template for other short-form platforms, such as TikTok and Instagram, and demonstrates how multimodal methods, combined with qualitative interpretation, can characterize sentiment patterns and visual cues in algorithmically driven video environments.

Method: VRF uses variational distributions to represent user preferences in shared preference space, infers distributions via variational encoder, derives weights through Wasserstein distance matching with shared probabilistic bases, and downweights uncertain estimates through variance-attenuated loss.

Result: On three benchmarks, VRF outperforms all baselines across seen and unseen users, few-shot scenarios, and varying uncertainty levels, with gains extending to downstream alignment tasks.

Conclusion: VRF provides an uncertainty-aware framework for reward factorization that improves personalization of LLMs by better handling scarce user data and uncertainty in preference estimation.

Abstract: Reward factorization personalizes large language models (LLMs) by decomposing rewards into shared basis functions and user-specific weights. Yet, existing methods estimate user weights from scarce data in isolation and as deterministic points, leading to inaccurate and unreliable inference. We introduce Variational Reward Factorization (VRF), an uncertainty-aware framework that represents each user’s preferences as a variational distribution in a shared preference space. VRF infers user distributions via a variational encoder, derives weights through Wasserstein distance matching with shared probabilistic bases, and downweights uncertain estimates through a variance-attenuated loss. On three benchmarks, VRF outperforms all baselines across seen and unseen users, few-shot scenarios, and varying uncertainty levels, with gains extending to downstream alignment.

Method: Analyzed two large corpora (Books3 with 52,796 books/759 authors and PG-19 with 28,439 books/1,821 authors) using information-theoretic novelty curves. Examined both book-level scalar dynamics (mean novelty, speed, volume, circuitousness) and chapter-level SAX motif patterns in sliding windows for author attribution.

Result: Book-level scalar dynamics identified 43% of authors significantly above chance; chapter-level SAX motif patterns achieved 30x-above-chance attribution. Signals are complementary, not redundant. Fingerprint persists within-genre for ~25% of authors. Classical authors (Twain, Austen, Kipling) show comparable fingerprint strength to modern authors.

Conclusion: Authorial voice leaves measurable, multi-scale traces in how novelty unfolds across texts, detectable through information-theoretic analysis. The phenomenon is not an artifact of contemporary publishing conventions and persists across historical periods.

Abstract: We test whether authors have characteristic “fingerprints” in the information-theoretic novelty curves of their published works. Working with two corpora – Books3 (52,796 books, 759 qualifying authors) and PG-19 (28,439 books, 1,821 qualifying authors) – we find that authorial voice leaves measurable traces in how novelty unfolds across a text. The signal is multi-scale: at book level, scalar dynamics (mean novelty, speed, volume, circuitousness) identify 43% of authors significantly above chance; at chapter level, SAX motif patterns in sliding windows achieve 30x-above-chance attribution, far exceeding the scalar features that dominate at book level. These signals are complementary, not redundant. We show that the fingerprint is partly confounded with genre but persists within-genre for approximately one-quarter of authors. Classical authors (Twain, Austen, Kipling) show fingerprints comparable in strength to modern authors, suggesting the phenomenon is not an artifact of contemporary publishing conventions.

Method: Used free recall paradigm from cognitive science, conducted systematic ablation experiments on open-source LLMs, specifically targeting induction heads (attention heads that attend to tokens following previous occurrences), and compared effects of removing high-induction-score heads vs random heads.

Result: LLMs consistently display serial-recall-like patterns with peak probability for tokens immediately following repeated tokens; ablation of high-induction-score heads substantially reduces +1 lag bias and impairs serial recall performance more than random head removal.

Conclusion: Induction heads are mechanistically important for temporal context processing in transformers, playing a specific role in ordered retrieval and serial-recall-like behavior during in-context learning.

Abstract: Large language models (LLMs) exhibit strong in-context learning capabilities, but how they track and retrieve information from context remains underexplored. Drawing on the free recall paradigm in cognitive science (where participants recall list items in any order), we show that several open-source LLMs consistently display a serial-recall-like pattern, assigning peak probability to tokens that immediately follow a repeated token in the input sequence. Through systematic ablation experiments, we show that induction heads, specialized attention heads that attend to the token following a previous occurrence of the current token, play an important role in this phenomenon. Removing heads with a high induction score substantially reduces the +1 lag bias, whereas ablating random heads does not reproduce the same reduction. We also show that removing heads with high induction scores impairs the performance of models prompted to do serial recall using few-shot learning to a larger extent than removing random heads. Our findings highlight a mechanistically specific connection between induction heads and temporal context processing in transformers, suggesting that these heads are especially important for ordered retrieval and serial-recall-like behavior during in-context learning.

Method: Proposed CARE framework: remote LLM generates structured categories/transitions without accessing patient data, while local LLM uses these to guide evidence acquisition and decision-making. Uses MIMIC-DOS dataset derived from MIMIC-IV, focusing exclusively on cases with signs-symptoms discordance.

Result: CARE outperforms single-pass LLMs and agentic pipelines across all key metrics, demonstrating more robust handling of conflicting clinical evidence while preserving privacy.

Conclusion: Multi-stage privacy-compliant agentic reasoning can effectively handle conflicting evidence in medical decision-making, with separation of reasoning guidance from sensitive data processing proving effective.

Abstract: Large language model (LLM) systems are increasingly used to support high-stakes decision-making, but they typically perform worse when the available evidence is internally inconsistent. Such a scenario exists in real-world healthcare settings, with patient-reported symptoms contradicting medical signs. To study this problem, we introduce MIMIC-DOS, a dataset for short-horizon organ dysfunction worsening prediction in the intensive care unit (ICU) setting. We derive this dataset from the widely recognized MIMIC-IV, a publicly available electronic health record dataset, and construct it exclusively from cases in which discordance between signs and symptoms exists. This setting poses a substantial challenge for existing LLM-based approaches, with single-pass LLMs and agentic pipelines often struggling to reconcile such conflicting signals. To address this problem, we propose CARE: a multi-stage privacy-compliant agentic reasoning framework in which a remote LLM provides guidance by generating structured categories and transitions without accessing sensitive patient data, while a local LLM uses these categories and transitions to support evidence acquisition and final decision-making. Empirically, CARE achieves stronger performance across all key metrics compared to multiple baseline settings, showing that CARE can more robustly handle conflicting clinical evidence while preserving privacy.

Method: Introduces Paper Reconstruction Evaluation (PaperRecon): create overview from existing paper, have agent generate full paper based on overview + minimal resources, compare against original. Evaluates two orthogonal dimensions: Presentation (rubric-based) and Hallucination (agentic evaluation grounded in original source). Uses PaperWrite-Bench benchmark of 51 papers from top-tier venues post-2025.

Result: Experiments show clear trade-off: ClaudeCode achieves higher presentation quality but with >10 hallucinations per paper on average, while Codex produces fewer hallucinations but lower presentation quality. Both improve with model advances.

Conclusion: This work establishes first evaluation framework for AI-driven paper writing, improving understanding of risks within research community. Highlights quality-risk trade-offs in current coding agents.

Abstract: This paper introduces the first systematic evaluation framework for quantifying the quality and risks of papers written by modern coding agents. While AI-driven paper writing has become a growing concern, rigorous evaluation of the quality and potential risks of AI-written papers remains limited, and a unified understanding of their reliability is still lacking. We introduce Paper Reconstruction Evaluation (PaperRecon), an evaluation framework in which an overview (overview.md) is created from an existing paper, after which an agent generates a full paper based on the overview and minimal additional resources, and the result is subsequently compared against the original paper. PaperRecon disentangles the evaluation of the AI-written papers into two orthogonal dimensions, Presentation and Hallucination, where Presentation is evaluated using a rubric and Hallucination is assessed via agentic evaluation grounded in the original paper source. For evaluation, we introduce PaperWrite-Bench, a benchmark of 51 papers from top-tier venues across diverse domains published after 2025. Our experiments reveal a clear trade-off: while both ClaudeCode and Codex improve with model advances, ClaudeCode achieves higher presentation quality at the cost of more than 10 hallucinations per paper on average, whereas Codex produces fewer hallucinations but lower presentation quality. This work takes a first step toward establishing evaluation frameworks for AI-driven paper writing and improving the understanding of its risks within the research community.

Method: Five-component system: (1) MoE-LoRA with noisy top-2 routing across transformer projections under QLoRA quantization, (2) residual boosting by freezing trained stacks and adding new ones, (3) curriculum-ordered sequential domain training, (4) null-space projection via randomized SVD for zero forgetting, (5) outcome-based sigmoid meta-router for cross-domain composition.

Result: Achieves 2.5x faster convergence than parameter-matched single LoRA, breaks through single-stack ceiling with residual boosting, recovers generation quality destroyed by ungated stack accumulation. Outcome-based router discovers domain stacks encode transferable cognitive primitives rather than domain-specific knowledge.

Conclusion: Brainstacks enables efficient continual multi-domain fine-tuning with zero forgetting and discovers that domain expertise is encoded as transferable cognitive primitives that can be composed across domains.

Abstract: We present Brainstacks, a modular architecture for continual multi-domain fine-tuning of large language models that packages domain expertise as frozen adapter stacks composing additively on a shared frozen base at inference. Five interlocking components: (1) MoE-LoRA with Shazeer-style noisy top-2 routing across all seven transformer projections under QLoRA 4-bit quantization with rsLoRA scaling; (2) an inner loop performing residual boosting by freezing trained stacks and adding new ones; (3) an outer loop training sequential domain-specific stacks with curriculum-ordered dependencies; (4) null-space projection via randomized SVD constraining new stacks to subspaces orthogonal to prior directions, achieving zero forgetting in isolation; (5) an outcome-based sigmoid meta-router trained on empirically discovered domain-combination targets that selectively weights stacks, enabling cross-domain composition. Two boundary experiments: (6) PSN pretraining on a randomly initialized model; (7) per-domain RL (DPO/GRPO) validating compatibility with post-SFT alignment. Validated on TinyLlama-1.1B (4 domains, 9 stacks) and Gemma 3 12B IT (5 domains, 10 stacks), MoE-LoRA achieves 2.5x faster convergence than parameter-matched single LoRA, residual boosting breaks through the single-stack ceiling, and the routed system recovers generation quality destroyed by ungated stack accumulation. The central finding: the outcome-based router discovers that domain stacks encode transferable cognitive primitives (instruction-following clarity, numerical reasoning, procedural logic, chain-of-thought structure) rather than domain-specific knowledge, with medical prompts routing to chat+math stacks in 97% of cases despite zero medical data in those stacks.

Method: S0 tuning freezes all model weights and optimizes only one initial state matrix per recurrent layer. The method uses roughly 48 execution-verified HumanEval training solutions. It works with hybrid models like Qwen3.5-4B (GatedDeltaNet hybrid) and FalconH1-7B (Mamba-2 hybrid). A variant called per-step state-offset tuning is also explored but adds per-step inference cost.

Result: S0 tuning outperforms LoRA by +10.8 pp on HumanEval (p < 0.001). On Qwen3.5-4B, it improves greedy pass@1 by +23.6 +/- 1.7 pp. On FalconH1-7B, S0 reaches 71.8% +/- 1.3 vs LoRA’s 71.4% +/- 2.4. Cross-domain transfer shows significant improvements on MATH-500 (+4.8 pp) and GSM8K (+2.8 pp). The tuned state is only ~48 MB with no weight merging needed for task switching.

Conclusion: Recurrent state initialization is an effective zero-inference-overhead PEFT surface for hybrid language models when verified supervision is scarce. The method enables task switching without weight merging or model reloading, making it practical for deployment.

Abstract: Using roughly 48 execution-verified HumanEval training solutions, tuning a single initial state matrix per recurrent layer, with zero inference overhead, outperforms LoRA by +10.8 pp (p < 0.001) on HumanEval. The method, which we call S0 tuning, optimizes one state matrix per recurrent layer while freezing all model weights. On Qwen3.5-4B (GatedDeltaNet hybrid), S0 tuning improves greedy pass@1 by +23.6 +/- 1.7 pp (10 seeds). On FalconH1-7B (Mamba-2 hybrid), S0 reaches 71.8% +/- 1.3 and LoRA reaches 71.4% +/- 2.4 (3 seeds), statistically indistinguishable at this sample size while requiring no weight merging. Cross-domain transfer is significant on MATH-500 (+4.8 pp, p = 0.00002, 8 seeds) and GSM8K (+2.8 pp, p = 0.0003, 10 seeds); a text-to-SQL benchmark (Spider) shows no transfer, consistent with the trajectory-steering mechanism. A prefix-tuning control on a pure Transformer (Qwen2.5-3B) degrades performance by -13.9 pp under all nine configurations tested. On Qwen3.5, a per-step state-offset variant reaches +27.1 pp, above both S0 and LoRA but with per-step inference cost. Taken together, the results show that recurrent state initialization is a strong zero-inference-overhead PEFT surface for hybrid language models when verified supervision is scarce. The tuned state is a ~48 MB file; task switching requires no weight merging or model reload. Code and library: https://github.com/jackyoung27/s0-tuning.

Method: Simple self-distillation (SSD): sample multiple solutions from the LLM with specific temperature and truncation settings, then perform standard supervised fine-tuning on those self-generated samples.

Result: SSD improves Qwen3-30B-Instruct from 42.4% to 55.3% pass@1 on LiveCodeBench v6, with gains concentrated on harder problems. The method generalizes across Qwen and Llama models at 4B, 8B, and 30B scales, including both instruct and thinking variants.

Conclusion: SSD offers a complementary post-training direction for improving LLM code generation by reshaping token distributions to balance precision and exploration, suppressing distractor tails while preserving useful diversity.

Abstract: Can a large language model (LLM) improve at code generation using only its own raw outputs, without a verifier, a teacher model, or reinforcement learning? We answer in the affirmative with simple self-distillation (SSD): sample solutions from the model with certain temperature and truncation configurations, then fine-tune on those samples with standard supervised fine-tuning. SSD improves Qwen3-30B-Instruct from 42.4% to 55.3% pass@1 on LiveCodeBench v6, with gains concentrating on harder problems, and it generalizes across Qwen and Llama models at 4B, 8B, and 30B scale, including both instruct and thinking variants. To understand why such a simple method can work, we trace these gains to a precision-exploration conflict in LLM decoding and show that SSD reshapes token distributions in a context-dependent way, suppressing distractor tails where precision matters while preserving useful diversity where exploration matters. Taken together, SSD offers a complementary post-training direction for improving LLM code generation.

Method: A modular four-stage framework: 1) seed creation, 2) question-answer pair generation, 3) self-verification, and 4) external verification using complete web search. Generated 20K reasoning-intensive queries across 15 domains requiring 4-5 reasoning steps each.

Result: Trained Qwen3-4B on ORBIT using GRPO, achieving strong performance among sub-4B LLMs as search agents on Wikipedia question answering tasks, demonstrating utility of synthetic datasets.

Conclusion: ORBIT provides an effective synthetic dataset for training search agents, showing that frugal frameworks without paid APIs can generate high-quality training data for complex reasoning tasks.

Abstract: Search agents, which integrate language models (LMs) with web search, are becoming crucial for answering complex user queries. Constructing training datasets for deep research tasks, involving multi-step retrieval and reasoning, remains challenging due to expensive human annotation, or cumbersome prerequisites. In this work, we introduce ORBIT, a training dataset with 20K reasoning-intensive queries with short verifiable answers, generated using a frugal framework without relying on paid API services. The modular framework relies on four stages: seed creation, question–answer pair generation, and two stages of verification: self and external. ORBIT spans 15 domains and each training pair requires 4–5 reasoning steps, with external search verification required from the complete web. We train Qwen3-4B as the base model on ORBIT using GRPO and evaluate it on Wikipedia question answering tasks. Extensive experiment results demonstrate that ORBIT-4B achieves strong performance among sub-4B LLMs as search agents, proving the utility of synthetic datasets. Our framework, code and datasets are open-sourced and available publicly.

Method: RELISH uses a learned latent state that iteratively refines through cross-attention over token-level LLM representations. It employs a latent iterative state head that processes frozen LLM embeddings, then maps the final refined state to a scalar value with a linear regressor. The architecture is lightweight and works with frozen LLM backbones.

Result: Across five datasets, four LLM backbones, and two training regimes, RELISH consistently outperforms prior baselines from all three major LLM regression families. It achieves these gains with only 3.4-3.7M trainable parameters (0.01-0.04% overhead), far less than LoRA-based alternatives (0.26-0.42%).

Conclusion: RELISH provides an effective, parameter-efficient solution for text regression with LLMs, demonstrating superior performance over existing methods while maintaining minimal computational overhead through its innovative latent iterative state refinement approach.

Abstract: We present RELISH (REgression with a Latent Iterative State Head), a novel, lightweight architecture designed for text regression with large language models. Rather than decoding numeric targets as text or aggregating multiple generated outputs, RELISH predicts scalar values directly from frozen LLM representations by iteratively refining a learned latent state through cross-attention over token-level representations, and then mapping the final state to a point estimate with a linear regressor. Across five datasets, four LLM backbones, and two LLM training regimes, RELISH consistently outperforms prior baselines from all three major LLM regression families, including autoregressive decoding, regression-aware inference, and existing predictive head methods. Despite these gains, RELISH remains highly parameter-efficient, requiring only 3.4-3.7M trainable parameters across frozen LLM backbones (only 0.01-0.04% additional overhead), far less than LoRA-based alternatives that grow with model size (0.26-0.42%).

Method: Introduces YC-Bench, a benchmark where agents run a simulated startup over a one-year horizon spanning hundreds of turns. Agents must manage employees, select task contracts, and maintain profitability in a partially observable environment with adversarial clients and growing payroll. Evaluates 12 models (proprietary and open source) across 3 seeds each.

Result: Only three models consistently surpass the starting capital of $200K. Claude Opus 4.6 achieved highest average final funds at $1.27M, followed by GLM-5 at $1.21M at 11× lower inference cost. Scratchpad usage (for persisting information across context truncation) was the strongest predictor of success. Adversarial client detection was the primary failure mode, accounting for 47% of bankruptcies.

Conclusion: Frontier models still fail through distinct failure modes like over-parallelization, demonstrating capability gaps for long-horizon performance. YC-Bench is open-source, reproducible, and configurable for evaluating strategic coherence in LLM agents.

Abstract: As LLM agents tackle increasingly complex tasks, a critical question is whether they can maintain strategic coherence over long horizons: planning under uncertainty, learning from delayed feedback, and adapting when early mistakes compound. We introduce $\texttt{YC-Bench}$, a benchmark that evaluates these capabilities by tasking an agent with running a simulated startup over a one-year horizon spanning hundreds of turns. The agent must manage employees, select task contracts, and maintain profitability in a partially observable environment where adversarial clients and growing payroll create compounding consequences for poor decisions. We evaluate 12 models, both proprietary and open source, across 3 seeds each. Only three models consistently surpass the starting capital of $200K, with Claude Opus 4.6 achieving the highest average final funds at $1.27 M, followed by GLM-5 at $1.21 M at 11$\times$ lower inference cost. Scratchpad usage, the sole mechanism for persisting information across context truncation, is the strongest predictor of success, and adversarial client detection is the primary failure mode, accounting for $47%$ of bankruptcies. Our analysis reveals that frontier models still fail through distinct failure modes such as over-parallelization, demonstrating the capability gaps for long-horizon performance. $\texttt{YC-Bench}$ is open-source, reproducible, and configurable.

Method: Combines YOCO decoder-decoder architecture with recursive computation via a Universal Self-Decoder that performs multiple iterations through parameter sharing, confining iterative processes to shallow, efficient-attention layers.

Result: Achieves favorable capability-efficiency tradeoff, constant global KV cache, linear pre-filling, and improved token utility and scaling behavior while maintaining efficient inference. Remains competitive in general and long-context benchmarks.

Conclusion: Integration of efficient-attention architectures and recursive computation is a promising direction for scalable LLMs, demonstrating that YOCO-U provides better scaling behavior than either approach alone.

Abstract: The rise of test-time scaling has remarkably boosted the reasoning and agentic proficiency of Large Language Models (LLMs). Yet, standard Transformers struggle to scale inference-time compute efficiently, as conventional looping strategies suffer from high computational overhead and a KV cache that inflates alongside model depth. We present Universal YOCO (YOCO-U), which combines the YOCO decoder-decoder architecture with recursive computation to achieve a synergistic effect greater than either alone. Built on the YOCO framework, YOCO-U implements a Universal Self-Decoder that performs multiple iterations via parameter sharing, while confining the iterative process to shallow, efficient-attention layers. This combination yields a favorable capability-efficiency tradeoff that neither YOCO nor recursion achieves independently. The YOCO architecture provides a constant global KV cache and linear pre-filling, while partial recursion enhances representational depth with limited overhead. Together, YOCO-U improves token utility and scaling behavior while maintaining efficient inference. Empirical results confirm that YOCO-U remains highly competitive in general and long-context benchmarks, demonstrating that the integration of efficient-attention architectures and recursive computation is a promising direction for scalable LLMs.

Method: Supervised fine-tuning on curated (false claim, correction) pairs injected as small “vaccine doses” (5-10% of tokens) alongside truthful data. Introduces direct negative supervision on labeled falsehoods with key design requirements: dosage, labeling, quarantine, and diversity.

Result: Across four open weight model families: improves TruthfulQA accuracy by 12 points, increases misinformation rejection rates by 30 points, while preserving overall model capability.

Conclusion: Immunization is a practical and scalable component of responsible LLM development. Advocates for standardized vaccine corpora and benchmarks to evaluate generalization.

Abstract: Large language models (LLMs) reproduce misinformation not by memorizing false facts alone, but by learning the linguistic patterns that make falsehoods persuasive, such as hedging, false presuppositions, and fabricated citations. We propose model immunization, a training paradigm based on supervised fine-tuning over curated (false claim, correction) pairs, injected as small vaccine doses (5 to 10% of tokens) alongside truthful data. Unlike post-hoc filtering or preference-based alignment, immunization introduces direct negative supervision on labeled falsehoods. Across four open weight model families, this approach improves TruthfulQA accuracy by 12 points and increases misinformation rejection rates by 30 points, while preserving overall model capability. We further outline key design requirements, including dosage, labeling, quarantine, and diversity and advocate for standardized vaccine corpora and benchmarks to evaluate generalization. These findings position immunization as a practical and scalable component of responsible LLM development.

Method: Proposes LWF framework that uses Fisher Information Matrix to weight parameter updates, computes forgetting confidence to evaluate self-generated knowledge regarding forgetting tasks, and periodically unlearns high-confidence knowledge during fine-tuning.

Result: Experiments demonstrate that applying graceful forgetting can contribute to enhanced fine-tuning performance, though fully uncovering knowledge interaction mechanisms in pre-trained language models remains challenging.

Conclusion: LWF provides an effective approach for graceful forgetting in generative language models, addressing negative transfer issues in the pretrain-finetune paradigm and improving downstream task performance through selective knowledge removal.

Abstract: Recently, the pretrain-finetune paradigm has become a cornerstone in various deep learning areas. While in general the pre-trained model would promote both effectiveness and efficiency of downstream tasks fine-tuning, studies have shown that not all knowledge acquired during pre-training is beneficial. Some of the knowledge may actually bring detrimental effects to the fine-tuning tasks, which is also known as negative transfer. To address this problem, graceful forgetting has emerged as a promising approach. The core principle of graceful forgetting is to enhance the learning plasticity of the target task by selectively discarding irrelevant knowledge. However, this approach remains underexplored in the context of generative language models, and it is often challenging to migrate existing forgetting algorithms to these models due to architecture incompatibility. To bridge this gap, in this paper we propose a novel framework, Learning With Forgetting (LWF), to achieve graceful forgetting in generative language models. With Fisher Information Matrix weighting the intended parameter updates, LWF computes forgetting confidence to evaluate self-generated knowledge regarding the forgetting task, and consequently, knowledge with high confidence is periodically unlearned during fine-tuning. Our experiments demonstrate that, although thoroughly uncovering the mechanisms of knowledge interaction remains challenging in pre-trained language models, applying graceful forgetting can contribute to enhanced fine-tuning performance.

Method: Proposes a ternary neuron classification methodology with identification algorithm, analyzes neuron distributions, and divides multilingual inference into four parts: multilingual understanding, shared semantic reasoning, multilingual output transformation, and vocabulary output.

Result: Provides empirical analysis of models before/after alignment, identifies “Spontaneous Multilingual Alignment” phenomenon, and offers insights into how different neuron types contribute to multilingual capabilities.

Conclusion: The work offers a comprehensive neuron-based framework for understanding multilingual alignment in LLMs, providing valuable empirical results and insights for improving multilingual capabilities.

Abstract: Multilingual Alignment is an effective and representative paradigm to enhance LLMs’ multilingual capabilities, which transfers the capabilities from the high-resource languages to the low-resource languages. Meanwhile, some research on language-specific neurons provides a new perspective to analyze and understand LLMs’ mechanisms. However, we find that there are many neurons that are shared by multiple but not all languages and cannot be correctly classified. In this work, we propose a ternary classification methodology that categorizes neurons into three types, including language-specific neurons, language-related neurons, and general neurons. And we propose a corresponding identification algorithm to distinguish these different types of neurons. Furthermore, based on the distributional characteristics of different types of neurons, we divide the LLMs’ internal process for multilingual inference into four parts: (1) multilingual understanding, (2) shared semantic space reasoning, (3) multilingual output space transformation, and (4) vocabulary space outputting. Additionally, we systematically analyze the models before and after alignment with a focus on different types of neurons. We also analyze the phenomenon of “Spontaneous Multilingual Alignment”. Overall, our work conducts a comprehensive investigation based on different types of neurons, providing empirical results and valuable insights to better understand multilingual alignment and multilingual capabilities of LLMs.

Method: Evaluated VLMs and LLMs on over 14K handwritten answers from grade-4 classrooms in Indonesia covering Mathematics and English aligned with local curriculum. Dataset features naturally curly, diverse handwriting from real classrooms. Assessment tasks included grading and generating personalized Indonesian feedback guided by rubric-based evaluation.

Result: VLM struggled with handwriting recognition, causing error propagation in LLM grading. However, LLM feedback remained pedagogically useful despite imperfect visual inputs, though limitations were revealed in personalization and contextual relevance.

Conclusion: Current VLMs face significant challenges with real-world handwritten content in educational contexts, but LLMs can still provide valuable feedback despite visual recognition limitations, highlighting the need for improved multimodal understanding in underrepresented educational settings.

Abstract: Despite rapid progress in vision-language and large language models (VLMs and LLMs), their effectiveness for AI-driven educational assessment in real-world, underrepresented classrooms remains largely unexplored. We evaluate state-of-the-art VLMs and LLMs on over 14K handwritten answers from grade-4 classrooms in Indonesia, covering Mathematics and English aligned with the local national curriculum. Unlike prior work on clean digital text, our dataset features naturally curly, diverse handwriting from real classrooms, posing realistic visual and linguistic challenges. Assessment tasks include grading and generating personalized Indonesian feedback guided by rubric-based evaluation. Results show that the VLM struggles with handwriting recognition, causing error propagation in LLM grading, yet LLM feedback remains pedagogically useful despite imperfect visual inputs, revealing limits in personalization and contextual relevance.

Method: Constructed MemeMind dataset with detailed Chain-of-Thought reasoning annotations aligned with international standards. Proposed MemeGuard, a reasoning-oriented multimodal detection framework for harmful meme detection.

Result: MemeGuard outperforms existing state-of-the-art methods on the MemeMind dataset, improving both detection accuracy and interpretability of model decisions.

Conclusion: The work establishes a solid foundation for future research in harmful meme detection through a high-quality dataset and effective multimodal reasoning framework.

Abstract: As a multimodal medium combining images and text, memes frequently convey implicit harmful content through metaphors and humor, rendering the detection of harmful memes a complex and challenging task. Although recent studies have made progress in detection accuracy and interpretability, large-scale, high-quality datasets for harmful memes remain scarce, and current methods still struggle to capture implicit risks and nuanced semantics. Thus, we construct MemeMind, a large-scale harmful meme dataset. Aligned with the international standards and the context of internet, MemeMind provides detailed Chain-of-Thought (CoT) reasoning annotations to support fine-grained analysis of implicit intentions in memes. Based on this dataset, we further propose MemeGuard, a reasoning-oriented multimodal detection framework that significantly improves both the accuracy of harmful meme detection and the interpretability of model decisions. Extensive experimental results demonstrate that MemeGuard outperforms existing state-of-the-art methods on the MemeMind dataset, establishing a solid foundation for future research in harmful meme detection. The complete dataset and code will be released upon acceptance.

Method: Proposes Structured In-context Environment (SIE) framework that automatically constructs reasoning environments from large-scale structured data. Uses rich compositional patterns in structured data to support generalizable reasoning, with explicit schemas and reasoning chains enabling rule-based verifiability.

Result: SIE achieves substantial improvements in in-domain structured reasoning and enables learned compositional reasoning skills to generalize effectively to out-of-domain mathematical and logical reasoning tasks. LLMs can infer missing information in partial SIEs through exploration.

Conclusion: SIE provides a scalable framework for RL fine-tuning of LLMs that supports generalizable reasoning and verifiability, bridging the gap between specialized game environments and expert-dependent mathematical/coding environments.

Abstract: Large language models (LLMs) have achieved significant advancements in reasoning capabilities through reinforcement learning (RL) via environmental exploration. As the intrinsic properties of the environment determine the abilities that LLMs can learn, the environment plays a important role in the RL finetuning process. An ideal LLM reasoning environment should possess three core characteristics: scalability, generalizable reasoning, and verifiability. However, existing mathematical and coding environments are difficult to scale due to heavy reliance on expert annotation, while the skills learned in game-based environments are too specialized to generalize. To bridge this gap, we introduce the \textbf{S}tructured \textbf{I}n-context \textbf{E}nvironment (SIE) framework. SIE achieves scalability by automatically constructing reasoning environments from large-scale structured data, where the rich compositional patterns naturally support generalizable reasoning. Moreover, the explicit schemas and reasoning chains in structured data provide a foundation for rule-based verifiability. Experimental results show that SIE framework not only achieves substantial improvements in in-domain structured reasoning, but also enables the learned compositional reasoning skills to generalize effectively to out-of-domain mathematical and logical reasoning tasks. We further explored learning in information-limited partial SIEs and found that LLMs can infer the missing information through exploring the environment, leading to robust reasoning improvements and generalization performance.

Method: Used multilingual Wug Test adaptation to test six models across four languages (Catalan, English, Greek, Spanish) and compared with human speakers.

Result: Models generalize morphology with human-like accuracy but accuracy patterns align more with community size/data availability than structural complexity; languages with larger communities (Spanish, English) showed higher accuracy than less-resourced ones (Catalan, Greek).

Conclusion: Model behavior is driven by richness of linguistic resources rather than sensitivity to grammatical complexity, reflecting only superficial resemblance to human linguistic competence.

Abstract: The linguistic abilities of Large Language Models are a matter of ongoing debate. This study contributes to this discussion by investigating model performance in a morphological generalization task that involves novel words. Using a multilingual adaptation of the Wug Test, six models were tested across four partially unrelated languages (Catalan, English, Greek, and Spanish) and compared with human speakers. The aim is to determine whether model accuracy approximates human competence and whether it is shaped primarily by linguistic complexity or by the size of the linguistic community, which affects the quantity of available training data. Consistent with previous research, the results show that the models are able to generalize morphological processes to unseen words with human-like accuracy. However, accuracy patterns align more closely with community size and data availability than with structural complexity, refining earlier claims in the literature. In particular, languages with larger speaker communities and stronger digital representation, such as Spanish and English, revealed higher accuracy than less-resourced ones like Catalan and Greek. Overall, our findings suggest that model behavior is mainly driven by the richness of linguistic resources rather than by sensitivity to grammatical complexity, reflecting a form of performance that resembles human linguistic competence only superficially.

Method: Proposes PluriHopRAG which learns from synthetic examples to decompose queries according to corpus-specific document structure and employs a cross-encoder filter at document level to minimize costly LLM reasoning.

Result: On PluriHopWIND benchmark (48 pluri-hop questions over 191 wind-industry reports), shows 18-52% F1 score improvement across base LLMs. On Loong benchmark (financial/legal/scientific reports), shows 33% improvement over long-context reasoning and 52% improvement over naive RAG.

Conclusion: PluriHopRAG effectively addresses the challenging pluri-hop question answering scenario requiring exhaustive document checking, significantly outperforming existing RAG methods on real-world multilingual and domain-specific benchmarks.

Abstract: Retrieval-Augmented Generation (RAG) has been used in question answering (QA) systems to improve performance when relevant information is in one (single-hop) or multiple (multi-hop) passages. However, many real life scenarios (e.g. dealing with financial, legal, medical reports) require checking all documents for relevant information without a clear stopping condition. We term these pluri-hop questions, and formalize them by 3 conditions - recall sensitivity, exhaustiveness, and exactness. To study this setting, we introduce PluriHopWIND, a multilingual diagnostic benchmark of 48 pluri-hop questions over 191 real wind-industry reports, with high repetitiveness to reflect the challenge of distractors in real-world datasets. Naive, graph-based, and multimodal RAG methods only reach up to 40% statement-wise F1 on PluriHopWIND. Motivated by this, we propose PluriHopRAG, which learns from synthetic examples to decompose queries according to corpus-specific document structure, and employs a cross-encoder filter at the document level to minimize costly LLM reasoning. We test PluriHopRAG on PluriHopWIND and the Loong benchmark built on financial, legal and scientific reports. On PluriHopWIND, our method shows 18-52% F1 score improvement across base LLMs, while on Loong, we show 33% improvement over long-context reasoning and 52% improvement over naive RAG.

Method: Created multilingual corpus from TED-Multi and Europarl covering 8 language pairs, identified candidate explicitation cases through null alignments, refined using active learning with human annotation, and developed detection framework.

Result: Entity and system-level explicitations are most frequent; active learning improves classifier accuracy by 7-8 percentage points, achieving up to 0.88 accuracy and 0.82 F1 across languages.

Conclusion: PragExTra establishes pragmatic explicitation as a measurable, cross-linguistic phenomenon and takes a step towards building culturally aware machine translation.

Abstract: Translators often enrich texts with background details that make implicit cultural meanings explicit for new audiences. This phenomenon, known as pragmatic explicitation, has been widely discussed in translation theory but rarely modeled computationally. We introduce PragExTra, the first multilingual corpus and detection framework for pragmatic explicitation. The corpus covers eight language pairs from TED-Multi and Europarl and includes additions such as entity descriptions, measurement conversions, and translator remarks. We identify candidate explicitation cases through null alignments and refined using active learning with human annotation. Our results show that entity and system-level explicitations are most frequent, and that active learning improves classifier accuracy by 7-8 percentage points, achieving up to 0.88 accuracy and 0.82 F1 across languages. PragExTra establishes pragmatic explicitation as a measurable, cross-linguistic phenomenon and takes a step towards building culturally aware machine translation. Keywords: translation, multilingualism, explicitation

Method: 1) Automated pipeline linking ~100K papers to their actual used baselines/datasets; 2) Collective perception enhanced retriever using self-descriptions + aggregated citation contexts; 3) Reasoning-augmented reranker with explicit reasoning chains and LLM fine-tuning for interpretable justifications.

Result: Dataset covers 85% of datasets/baselines used at top AI conferences over 5 years. Method outperforms prior baselines by +5.85% in Recall@20 and +8.30% in HitRate@5.

Conclusion: The framework advances reliable, interpretable automation of experimental design by leveraging collective perception from citation networks.

Abstract: Large language model agents are becoming increasingly capable at web-centric tasks such as information retrieval, complex reasoning. These emerging capabilities have given rise to surge research interests in developing LLM agent for facilitating scientific quest. One key application in AI research is to automate experiment design through agentic dataset and baseline retrieval. However, prior efforts suffer from limited data coverage, as recommendation datasets primarily harvest candidates from public portals and omit many datasets actually used in published papers, and from an overreliance on content similarity that biases model toward superficial similarity and overlooks experimental suitability. Harnessing collective perception embedded in the baseline and dataset citation network, we present a comprehensive framework for baseline and dataset recommendation. First, we design an automated data-collection pipeline that links roughly one hundred thousand accepted papers to the baselines and datasets they actually used. Second, we propose a collective perception enhanced retriever. To represent the position of each dataset or baseline within the scholarly network, it concatenates self-descriptions with aggregated citation contexts. To achieve efficient candidate recall, we finetune an embedding model on these representations. Finally, we develop a reasoning-augmented reranker that exact interaction chains to construct explicit reasoning chains and finetunes a large language model to produce interpretable justifications and refined rankings. The dataset we curated covers 85% of the datasets and baselines used at top AI conferences over the past five years. On our dataset, the proposed method outperforms the strongest prior baseline with average gains of +5.85% in Recall@20, +8.30% in HitRate@5. Taken together, our results advance reliable, interpretable automation of experimental design.

Method: Dual approach: 1) Quantitative benchmark of unitary/toxic-bert model comparing performance on African-American English (AAE) vs Standard American English (SAE); 2) Development of interactive pedagogical tool with user-controlled sensitivity threshold to demonstrate how bias manifests in policy decisions.

Result: Clear systematic bias: AAE text scored 1.8 times more toxic and 8.8 times higher for “identity hate” compared to SAE. Interactive tool makes abstract biases tangible and shows how human-set policies operationalize discrimination.

Conclusion: Provides statistical evidence of disparate impact in AI moderation systems and creates public-facing tool to foster critical AI literacy about how algorithmic bias combines with human policies to create discrimination.

Abstract: Now that AI-driven moderation has become pervasive in everyday life, we often hear claims that “the AI is biased”. While this is often said jokingly, the light-hearted remark reflects a deeper concern. How can we be certain that an online post flagged as “inappropriate” was not simply the victim of a biased algorithm? This paper investigates this problem using a dual approach. First, I conduct a quantitative benchmark of a widely used toxicity model (unitary/toxic-bert) to measure performance disparity between text in African-American English (AAE) and Standard American English (SAE). The benchmark reveals a clear, systematic bias: on average, the model scores AAE text as 1.8 times more toxic and 8.8 times higher for “identity hate”. Second, I introduce an interactive pedagogical tool that makes these abstract biases tangible. The tool’s core mechanic, a user-controlled “sensitivity threshold,” demonstrates that the biased score itself is not the only harm; instead, the more-concerning harm is the human-set, seemingly neutral policy that ultimately operationalises discrimination. This work provides both statistical evidence of disparate impact and a public-facing tool designed to foster critical AI literacy.

Method: Used 600 authentic student essays from AES 2.0 corpus, created controlled counterfactuals with implicit (lexicon-based gendered term swaps) and explicit (gendered author background in prompt) cues. Tested six LLMs, quantified response divergence using cosine/Euclidean distances over sentence embeddings, assessed significance via permutation tests, visualized with dimensionality reduction.

Result: All models showed larger semantic shifts for male-female counterfactuals than female-male. Only GPT and Llama models were sensitive to explicit gender cues. Qualitative analysis revealed linguistic differences: more autonomy-supportive feedback under male cues vs. more controlling feedback under female cues.

Conclusion: State-of-the-art LLMs exhibit persistent gender biases in feedback generation, requiring fairness auditing, reporting standards for counterfactual evaluation in learning analytics, and practical guidance for equitable prompt design and deployment.

Abstract: As teachers increasingly turn to GenAI in their educational practice, we need robust methods to benchmark large language models (LLMs) for pedagogical purposes. This article presents an embedding-based benchmarking framework to detect bias in LLMs in the context of formative feedback. Using 600 authentic student essays from the AES 2.0 corpus, we constructed controlled counterfactuals along two dimensions: (i) implicit cues via lexicon-based swaps of gendered terms within essays, and (ii) explicit cues via gendered author background in the prompt. We investigated six representative LLMs (i.e. GPT-5 mini, GPT-4o mini, DeepSeek-R1, DeepSeek-R1-Qwen, Gemini 2.5 Pro, Llama-3-8B). We first quantified the response divergence with cosine and Euclidean distances over sentence embeddings, then assessed significance via permutation tests, and finally, visualised structure using dimensionality reduction. In all models, implicit manipulations reliably induced larger semantic shifts for male-female counterfactuals than for female-male. Only the GPT and Llama models showed sensitivity to explicit gender cues. These findings show that even state-of-the-art LLMs exhibit asymmetric semantic responses to gender substitutions, suggesting persistent gender biases in feedback they provide learners. Qualitative analyses further revealed consistent linguistic differences (e.g., more autonomy-supportive feedback under male cues vs. more controlling feedback under female cues). We discuss implications for fairness auditing of pedagogical GenAI, propose reporting standards for counterfactual evaluation in learning analytics, and outline practical guidance for prompt design and deployment to safeguard equitable feedback.

Method: AlphaResearch uses a dual-environment approach combining execution-based verification rewards with simulated peer review feedback. It iteratively: (1) proposes new ideas, (2) programs to verify them, and (3) optimizes research proposals. The system is benchmarked on a dataset of eight open-ended algorithmic problems.

Result: AlphaResearch outperforms other agentic discovery systems on six open-ended problems. On the “packing circles” problem, it discovered an algorithm achieving best-known performance, surpassing human researchers and baselines like AlphaEvolve.

Conclusion: Autonomous research agents like AlphaResearch show promise for algorithmic discovery, though challenges remain. The work provides valuable insights for future research in autonomous scientific discovery systems.

Abstract: LLMs have made significant progress in complex but easy-to-verify problems, yet they still struggle with discovering the unknown. In this paper, we present \textbf{AlphaResearch}, an autonomous research agent designed to discover new algorithms on open-ended problems by iteratively running the following steps: (1) propose new ideas (2) program to verify (3) optimize the research proposals. To synergize the feasibility and innovation of the discovery process, we construct a novel dual environment by combining the execution-based verifiable reward and reward from simulated real-world peer review environment in AlphaResearch. We construct \textbf{\dataset}, a set of questions that includes an eight open-ended algorithmic problems competition to benchmark AlphaResearch. Experimental results show that AlphaResearch achieves stronger discovery performance than other agentic discovery systems on six open-ended problems. Notably, the algorithm discovered by AlphaResearch on the \emph{``packing circles’’} problem achieves the best-of-known performance, surpassing the results of human researchers and strong baselines from recent work (e.g., AlphaEvolve). Additionally, we conduct a comprehensive analysis of the benefits and remaining challenges of autonomous research agent, providing valuable insights for future research.

Method: Predicting verbalized probability distributions instead of single answers, requiring LLMs to consider all possible options, with systematic experiments comparing different verbalization methods across multiple LLMs and tasks, including simple prompting and RL optimization

Result: Method consistently outperforms baselines, achieves higher reasoning efficacy during inference-time scaling, saves nearly 6× computation to reach best Brier score on MMLU-Pro, though limitations exist on specific tasks

Conclusion: Verbalized probability distribution prediction effectively promotes reasoning for confidence estimation by requiring consideration of all possible answers and eliciting more careful confidence assignment

Abstract: Knowing the reliability of a model’s response is essential in practical applications. Given the strong generation capabilities of large language models (LLMs), research has focused on generating verbalized confidence. This approach is further enhanced by integrating chain-of-thought reasoning, which provides logical and transparent estimates. However, how reasoning strategies affect the estimated confidence remains under-explored. In this work, we demonstrate that predicting a verbalized probability distribution effectively promotes reasoning for confidence estimation. It requires an LLM to consider all possible answers rather than relying on a single guess, and the requirement of producing a distribution elicits more careful confidence assignment. We conduct systematic experiments comparing different verbalization-based methods across multiple LLMs and tasks. Our method consistently shows advantages, whether in the simple prompting setup or after optimization via reinforcement learning (RL). Notably, it achieves higher reasoning efficacy during inference-time scaling, saving nearly 6$\times$ the computation to reach the best Brier score of the strongest baseline on MMLU-Pro. Additionally, we reveal its limitations on specific tasks and discuss possible solutions for broader applicability.

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 2511.20836: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.20836&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed API request

Result: Unable to determine results due to failed API request

Conclusion: Unable to determine conclusion due to failed API request

Abstract: Failed to fetch summary for 2512.00234: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.00234&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2601.06565: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.06565&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2601.09504: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.09504&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to draw conclusions due to failed paper fetch

Abstract: Failed to fetch summary for 2601.04823: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.04823&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 2602.02326: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.02326&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 2602.16080: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.16080&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 2602.17194: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.17194&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2602.23440: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.23440&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as abstract retrieval failed

Result: Cannot determine results as abstract retrieval failed

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2603.06865: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.06865&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as abstract retrieval failed

Result: Cannot determine results as abstract retrieval failed

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2603.10035: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.10035&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access restrictions

Result: Unable to determine results due to access restrictions

Conclusion: Unable to determine conclusion due to access restrictions

Abstract: Failed to fetch summary for 2603.14443: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.14443&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2603.17445: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.17445&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 2603.20209: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.20209&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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

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

Conclusion: Unable to determine conclusion as paper content could not be retrieved

Abstract: Failed to fetch summary for 2603.20895: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.20895&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to draw conclusions due to failed paper fetch

Abstract: Failed to fetch summary for 2603.21454: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.21454&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2603.25052: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.25052&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 2603.29493: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.29493&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to missing paper content

Result: Unable to determine results due to missing paper content

Conclusion: Unable to draw conclusions due to missing paper content

Abstract: Failed to fetch summary for 2405.15556: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2405.15556&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed API request

Result: Unable to determine results due to failed API request

Conclusion: Unable to determine conclusion due to failed API request

Abstract: Failed to fetch summary for 2508.07629: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.07629&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to the paper abstract

Result: Cannot determine results without access to the paper abstract

Conclusion: Cannot draw conclusions without access to the paper abstract

Abstract: Failed to fetch summary for 2510.22276: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.22276&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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

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

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

Abstract: Failed to fetch summary for 2602.22721: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.22721&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 2603.03823: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.03823&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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

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

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

Abstract: Failed to fetch summary for 2603.09697: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.09697&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 draw conclusions due to failed fetch

Abstract: Failed to fetch summary for 2603.17205: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.17205&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 2603.27982: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.27982&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Proposes HIVE framework with hierarchical cross-attention between vision encoder and LLM layers, using a three-stage training strategy for progressive alignment and stable optimization.

Result: Achieves superior performance in image classification and vision-language tasks, outperforming self-attention-based methods on MME, GQA, OK-VQA, and ScienceQA benchmarks.

Conclusion: Hierarchical feature integration improves vision-language models, paving way for more efficient and expressive multimodal systems.

Abstract: The field of computer vision has experienced significant advancements through scalable vision encoders and multimodal pre-training frameworks. However, existing approaches often treat vision encoders and large language models (LLMs) as independent modules, limiting the integration of hierarchical visual features. In this work, we propose HIVE (Hierarchical Pre-Training of Vision Encoders), a novel framework that enhances vision-language alignment by introducing hierarchical cross-attention between the vision encoder and LLM. Unlike conventional methods that flatten image embeddings, HIVE enables structured feature fusion across multiple layers, improving gradient flow and representation learning. To optimize this interaction, we introduce a three-stage training strategy that progressively aligns the vision encoder with the LLM, ensuring stable optimization and effective multimodal fusion. Empirical evaluations demonstrate that HIVE achieves superior performance not only in image classification but also on various vision-language tasks, outperforming self-attention-based methods in benchmarks such as MME, GQA, OK-VQA, and ScienceQA. Our results highlight the benefits of hierarchical feature integration, paving the way for more efficient and expressive vision-language models.

Method: Two-stage pipeline: 1) Visual segmentation to isolate virtual and physical layers, 2) Region-aware retrieval to avoid ambiguous semantic context from projection distortion. Introduces RGBP dataset with 65 scenes and 180k+ projections with decoupled annotations.

Result: ProCap provides robust semantic foundation for SAR, enabling accurate scene understanding by decoupling projected content from physical reality. Establishes dual-captioning evaluation protocol with task-specific tokens.

Conclusion: ProCap addresses the virtual-physical ambiguity problem in SAR, enabling more intelligent interaction capabilities. The framework, models, and RGBP dataset provide foundation for future SAR research.

Abstract: Spatial augmented reality (SAR) directly projects digital content onto physical scenes using projectors, creating immersive experience without head-mounted displays. However, for SAR to support intelligent interaction, such as reasoning about the scene or answering user queries, it must semantically distinguish between the physical scene and the projected content. Standard Vision Language Models (VLMs) struggle with this virtual-physical ambiguity, often confusing the two contexts. To address this issue, we introduce ProCap, a novel framework that explicitly decouples projected content from physical scenes. ProCap employs a two-stage pipeline: first it visually isolates virtual and physical layers via automated segmentation; then it uses region-aware retrieval to avoid ambiguous semantic context due to projection distortion. To support this, we present RGBP (RGB + Projections), the first large-scale SAR semantic benchmark dataset, featuring 65 diverse physical scenes and over 180,000 projections with dense, decoupled annotations. Finally, we establish a dual-captioning evaluation protocol using task-specific tokens to assess physical scene and projection descriptions independently. Our experiments show that ProCap provides a robust semantic foundation for future SAR research. The source code, pre-trained models and the RGBP dataset are available on the project page: https://ZimoCao.github.io/ProCap/.

Method: Leverages generative priors from large-scale sRGB diffusion models to synthesize linear outputs (CIE XYZ or camera-specific raw). Creates many-to-one inverse-ISP dataset with multiple sRGB renditions anchored to common scene-referred targets. Fine-tunes conditional denoiser and specialized decoder to invert diverse ISP pipelines.

Result: Superior performance over traditional inverse-ISP methods that assume fixed ISP. Enables scalable text-driven synthetic data generation that benefits downstream low-level vision tasks.

Conclusion: RawGen successfully bridges the gap between sRGB diffusion models and physically meaningful linear representations, enabling text-to-raw generation and inversion for arbitrary cameras.

Abstract: Cameras capture scene-referred linear raw images, which are processed by onboard image signal processors (ISPs) into display-referred 8-bit sRGB outputs. Although raw data is more faithful for low-level vision tasks, collecting large-scale raw datasets remains a major bottleneck, as existing datasets are limited and tied to specific camera hardware. Generative models offer a promising way to address this scarcity – however, existing diffusion frameworks are designed to synthesize photo-finished sRGB images rather than physically consistent linear representations. This paper presents RawGen, to our knowledge the first diffusion-based framework enabling text-to-raw generation for arbitrary target cameras, alongside sRGB-to-raw inversion. RawGen leverages the generative priors of large-scale sRGB diffusion models to synthesize physically meaningful linear outputs, such as CIE XYZ or camera-specific raw representations, via specialized processing in latent and pixel spaces. To handle unknown and diverse ISP pipelines and photo-finishing effects in diffusion-model training data, we build a many-to-one inverse-ISP dataset where multiple sRGB renditions of the same scene generated using diverse ISP parameters are anchored to a common scene-referred target. Fine-tuning a conditional denoiser and specialized decoder on this dataset allows RawGen to obtain camera-centric linear reconstructions that effectively invert the rendering pipeline. We demonstrate RawGen’s superior performance over traditional inverse-ISP methods that assume a fixed ISP. Furthermore, we show that augmenting training pipelines with RawGen’s scalable, text-driven synthetic data can benefit downstream low-level vision tasks.

Method: Paraphrase-Decomposition-Aggregation (PDA) framework with three components: 1) prompt paraphrasing to generate diverse text inputs, 2) question decomposition to break complex queries into simpler parts, and 3) consistency aggregation to combine multiple outputs. PDA operates entirely at test time without model modifications.

Result: Experiments on multiple VLM architectures and benchmarks (visual question answering, classification, captioning) show PDA achieves consistent robustness gains against various adversarial perturbations while maintaining competitive clean accuracy.

Conclusion: PDA establishes a generic, strong, and practical defense framework for VLMs during inference that balances robustness and efficiency without requiring model retraining.

Abstract: Vision-language models (VLMs) are vulnerable to adversarial image perturbations. Existing works based on adversarial training against task-specific adversarial examples are computationally expensive and often fail to generalize to unseen attack types. To address these limitations, we introduce Paraphrase-Decomposition-Aggregation (PDA), a training-free defense framework that leverages text augmentation to enhance VLM robustness under diverse adversarial image attacks. PDA performs prompt paraphrasing, question decomposition, and consistency aggregation entirely at test time, thus requiring no modification on the underlying models. To balance robustness and efficiency, we instantiate PDA as invariants that reduce the inference cost while retaining most of its robustness gains. Experiments on multiple VLM architectures and benchmarks for visual question answering, classification, and captioning show that PDA achieves consistent robustness gains against various adversarial perturbations while maintaining competitive clean accuracy, establishing a generic, strong and practical defense framework for VLMs during inference.

Method: Proposes Q-Mask framework with causal query-driven mask decoder (CQMD) that performs sequential query-conditioned visual mask generation before final OCR output, inspired by chain-of-thought reasoning. Uses TextAnchor-26M dataset with fine-grained masks for training.

Result: Q-Mask substantially improves text anchoring and understanding across diverse visual scenes, outperforming both general-purpose and OCR-specific vision-language models on the introduced TextAnchor-Bench benchmark.

Conclusion: The visual chain-of-thought paradigm disentangles text localization from recognition, enabling explicit text anchor construction and improving OCR reliability for vision-language models.

Abstract: Optical Character Recognition (OCR) is increasingly regarded as a foundational capability for modern vision-language models (VLMs), enabling them not only to read text in images but also to support downstream reasoning in real-world visual question answering (VQA). However, practical applications further require reliable text anchors, i.e., accurately grounding queried text to its corresponding spatial region. To systematically evaluate this capability, we introduce TextAnchor-Bench (TABench), a benchmark for fine-grained text-region grounding, which reveals that both general-purpose and OCR-specific VLMs still struggle to establish accurate and stable text anchors. To address this limitation, we propose Q-Mask, a precise OCR framework built upon a causal query-driven mask decoder (CQMD). Inspired by chain-of-thought reasoning, Q-Mask performs causal visual decoding that sequentially generates query-conditioned visual masks before producing the final OCR output. This visual CoT paradigm disentangles where the text is from what the text is, enforcing grounded evidence acquisition prior to recognition and enabling explicit text anchor construction during inference. To train CQMD, we construct TextAnchor-26M, a large-scale dataset of image-text pairs annotated with fine-grained masks corresponding to specific textual elements, encouraging stable text-region correspondences and injecting strong spatial priors into VLM training. Extensive experiments demonstrate that Q-Mask substantially improves text anchoring and understanding across diverse visual scenes.

Method: Developed a model-agnostic semantic invariance score using PCA on real and synthetic non-semantic images, then created SOAP - a post-hoc suppression method using orthogonal projection to remove non-semantic information from patch representations.

Result: SOAP leads to consistent improvements in zero-shot performance across various MIM-based models, requires zero training, and can be attached as a single linear head to any model.

Conclusion: SOAP effectively addresses the non-semantic information retention problem in MIM representations through a simple, training-free post-hoc solution that improves downstream performance.

Abstract: Masked Image Modeling (MIM) has become a ubiquitous self-supervised vision paradigm. In this work, we show that MIM objectives cause the learned representations to retain non-semantic information, which ultimately hurts performance during inference. We introduce a model-agnostic score for semantic invariance using Principal Component Analysis (PCA) on real and synthetic non-semantic images. Based on this score, we propose a simple method, Semantically Orthogonal Artifact Projection (SOAP), to directly suppress non-semantic information in patch representations, leading to consistent improvements in zero-shot performance across various MIM-based models. SOAP is a post-hoc suppression method, requires zero training, and can be attached to any model as a single linear head.

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

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

Conclusion: CoEvo effectively addresses cross-modal misalignment in zero-shot OOD detection through bidirectional, sample-conditioned adaptation of proxies without requiring training or annotations.

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

Method: UCell incorporates a recursive structure into the forward computation graph for parameter efficiency. The 10-30M parameter model is trained from scratch on microscopy imaging data without relying on massive pretraining on natural images, enabling domain-specific training.

Result: UCell matches performance of models 10-20 times larger on multiple single-cell segmentation benchmarks, with similar generalizability to unseen out-of-domain data. It demonstrates strong adaptability through one-shot and few-shot fine-tuning on diverse small datasets.

Conclusion: Small models can achieve competitive performance through architectural innovations like recursive structures, decoupling biomedical model development from commercial interests and massive pretraining requirements.

Abstract: The modern deep learning field is a scale-centric one. Larger models have been shown to consistently perform better than smaller models of similar architecture. In many sub-domains of biomedical research, however, the model scaling is bottlenecked by the amount of available training data, and the high cost associated with generating and validating additional high quality data. Despite the practical hurdle, the majority of the ongoing research still focuses on building bigger foundation models, whereas the alternative of improving the ability of small models has been under-explored. Here we experiment with building models with 10-30M parameters, tiny by modern standards, to perform the single-cell segmentation task. An important design choice is the incorporation of a recursive structure into the model’s forward computation graph, leading to a more parameter-efficient architecture. We found that for the single-cell segmentation, on multiple benchmarks, our small model, UCell, matches the performance of models 10-20 times its size, and with a similar generalizability to unseen out-of-domain data. More importantly, we found that ucell can be trained from scratch using only a set of microscopy imaging data, without relying on massive pretraining on natural images, and therefore decouples the model building from any external commercial interests. Finally, we examined and confirmed the adaptability of ucell by performing a wide range of one-shot and few-shot fine tuning experiments on a diverse set of small datasets. Implementation is available at https://github.com/jiyuuchc/ucell

Method: Proposes Unify-Agent, a unified multimodal agent that reframes image generation as an agentic pipeline: prompt understanding → multimodal evidence searching → grounded recaptioning → final synthesis. Constructs 143K high-quality agent trajectories for training and introduces FactIP benchmark for evaluation.

Result: Unify-Agent substantially improves over its base unified model across diverse benchmarks and real-world generation tasks, approaching the world knowledge capabilities of the strongest closed-source models. Shows strong performance on the FactIP benchmark covering culturally significant and long-tail factual concepts.

Conclusion: Agent-based modeling for world-grounded image synthesis demonstrates value of tightly coupling reasoning, searching, and generation for reliable open-world agentic image synthesis. Highlights potential of agentic approaches to overcome limitations of frozen parametric knowledge in multimodal models.

Abstract: Unified multimodal models provide a natural and promising architecture for understanding diverse and complex real-world knowledge while generating high-quality images. However, they still rely primarily on frozen parametric knowledge, which makes them struggle with real-world image generation involving long-tail and knowledge-intensive concepts. Inspired by the broad success of agents on real-world tasks, we explore agentic modeling to address this limitation. Specifically, we present Unify-Agent, a unified multimodal agent for world-grounded image synthesis, which reframes image generation as an agentic pipeline consisting of prompt understanding, multimodal evidence searching, grounded recaptioning, and final synthesis. To train our model, we construct a tailored multimodal data pipeline and curate 143K high-quality agent trajectories for world-grounded image synthesis, enabling effective supervision over the full agentic generation process. We further introduce FactIP, a benchmark covering 12 categories of culturally significant and long-tail factual concepts that explicitly requires external knowledge grounding. Extensive experiments show that our proposed Unify-Agent substantially improves over its base unified model across diverse benchmarks and real world generation tasks, while approaching the world knowledge capabilities of the strongest closed-source models. As an early exploration of agent-based modeling for world-grounded image synthesis, our work highlights the value of tightly coupling reasoning, searching, and generation for reliable open-world agentic image synthesis.

Method: Proposes MAGNet, a multimodal transformer-based model that jointly encodes spatial and semantic goal representations, integrating historical context with self-motion cues for memory-augmented goal reasoning in continuous 3D environments.

Result: MAGNet significantly outperforms state-of-the-art methods with up to 12.1% absolute improvement in success rate, demonstrating robustness to short-duration sounds and long-distance navigation scenarios.

Conclusion: SAVN-CE establishes a more realistic continuous audio-visual navigation setting, and MAGNet effectively addresses the challenge of intermittent sound targets through multimodal encoding and memory-augmented reasoning, advancing embodied AI navigation capabilities.

Abstract: Audio-visual navigation enables embodied agents to navigate toward sound-emitting targets by leveraging both auditory and visual cues. However, most existing approaches rely on precomputed room impulse responses (RIRs) for binaural audio rendering, restricting agents to discrete grid positions and leading to spatially discontinuous observations. To establish a more realistic setting, we introduce Semantic Audio-Visual Navigation in Continuous Environments (SAVN-CE), where agents can move freely in 3D spaces and perceive temporally and spatially coherent audio-visual streams. In this setting, targets may intermittently become silent or stop emitting sound entirely, causing agents to lose goal information. To tackle this challenge, we propose MAGNet, a multimodal transformer-based model that jointly encodes spatial and semantic goal representations and integrates historical context with self-motion cues to enable memory-augmented goal reasoning. Comprehensive experiments demonstrate that MAGNet significantly outperforms state-of-the-art methods, achieving up to a 12.1% absolute improvement in success rate. These results also highlight its robustness to short-duration sounds and long-distance navigation scenarios. The code is available at https://github.com/yichenzeng24/SAVN-CE.

Method: PRISM uses a differentiable analysis-by-synthesis framework with explicit multi-compartment forward modeling (CSF, gray matter, up to K white-matter fiber compartments, and restricted compartment). It includes soft model selection via repulsion and sparsity priors, supports both MSE and Rician negative log-likelihood objectives, and includes a lightweight nuisance calibration module for robustness.

Result: On synthetic crossing-fiber data, PRISM achieves 3.5° best-match angular error with 95% recall (1.9x lower than best baseline). With NLL mode and learned sigma, error drops to 2.3° with 99% recall, resolving crossings down to 20°. On DiSCo1 phantom, PRISM improves connectivity correlation over CSD baselines. Whole-brain HCP fitting completes in ~12 minutes on a single GPU.

Conclusion: PRISM provides an effective differentiable framework for diffusion MRI microstructure fitting that significantly improves fiber orientation recovery in crossing regions while being computationally efficient and robust.

Abstract: Diffusion MRI microstructure fitting is nonconvex and often performed voxelwise, which limits fiber peak recovery in narrow crossings. This work introduces PRISM, a differentiable analysis-by-synthesis framework that fits an explicit multi-compartment forward model end-to-end over spatial patches. The model combines cerebrospinal fluid (CSF), gray matter, up to K white-matter fiber compartments (stick-and-zeppelin), and a restricted compartment, with explicit fiber directions and soft model selection via repulsion and sparsity priors. PRISM supports a fast MSE objective and a Rician negative log-likelihood (NLL) that jointly learns sigma without oracle information. A lightweight nuisance calibration module (smooth bias field and per-measurement scale/offset) is included for robustness and regularized to identity in clean-data tests. On synthetic crossing-fiber data (SNR=30; five methods, 16 crossing angles), PRISM achieves 3.5 degrees best-match angular error with 95% recall, which is 1.9x lower than the best baseline (MSMT-CSD, 6.8 degrees, 83% recall); in NLL mode with learned sigma, error drops to 2.3 degrees with 99% recall, resolving crossings down to 20 degrees. On the DiSCo1 phantom (NLL mode), PRISM improves connectivity correlation over CSD baselines at all four tracking angles (best r=.934 at 25 degrees vs. .920 for MSMT-CSD). Whole-brain HCP fitting (~741k voxels, MSE mode) completes in ~12 min on a single GPU with near-identical results across random seeds.

Method: Proposes QAsk-Nav benchmark with: 1) lightweight question-asking protocol scored independently of navigation, 2) enhanced navigation protocol with realistic target descriptions, and 3) open-source dataset with 28,000 quality-checked reasoning and question-asking traces.

Result: Developed Light-CoNav, a lightweight unified model that is 3x smaller and 70x faster than existing modular methods while outperforming state-of-the-art CoIN approaches in generalization to unseen objects and environments.

Conclusion: QAsk-Nav enables explicit, separate assessment of embodied navigation and collaborative question asking, providing a reproducible benchmark for evaluating interactive capabilities in collaborative navigation tasks.

Abstract: We propose Question-Asking Navigation (QAsk-Nav), the first reproducible benchmark for Collaborative Instance Object Navigation (CoIN) that enables an explicit, separate assessment of embodied navigation and collaborative question asking. CoIN tasks an embodied agent with reaching a target specified in free-form natural language under partial observability, using only egocentric visual observations and interactive natural-language dialogue with a human, where the dialogue can help to resolve ambiguity among visually similar object instances. Existing CoIN benchmarks are primarily focused on navigation success and offer no support for consistent evaluation of collaborative interaction. To address this limitation, QAsk-Nav provides (i) a lightweight question-asking protocol scored independently of navigation, (ii) an enhanced navigation protocol with realistic, diverse, high-quality target descriptions, and (iii) an open-source dataset, that includes 28,000 quality-checked reasoning and question-asking traces for training and analysis of interactive capabilities of CoIN models. Using the proposed QAsk-Nav benchmark, we develop Light-CoNav, a lightweight unified model for collaborative navigation that is 3x smaller and 70x faster than existing modular methods, while outperforming state-of-the-art CoIN approaches in generalization to unseen objects and environments. Project page at https://benchmarking-interaction.github.io/

Method: Proposes Omni-MMSI-R, a reference-guided pipeline that uses tools to produce identity-attributed social cues and conducts chain-of-thought social reasoning, with participant-level reference pairs and curated reasoning annotations built on existing datasets.

Result: Omni-MMSI-R outperforms advanced LLMs and other counterparts on the Omni-MMSI task, demonstrating superior performance in comprehensive social interaction understanding from raw multimodal inputs.

Conclusion: The Omni-MMSI task and proposed pipeline address critical limitations in multimodal AI systems for social interaction understanding, enabling more realistic and effective AI assistants that can perceive and reason about human interactions from raw data.

Abstract: We introduce Omni-MMSI, a new task that requires comprehensive social interaction understanding from raw audio, vision, and speech input. The task involves perceiving identity-attributed social cues (e.g., who is speaking what) and reasoning about the social interaction (e.g., whom the speaker refers to). This task is essential for developing AI assistants that can perceive and respond to human interactions. Unlike prior studies that operate on oracle-preprocessed social cues, Omni-MMSI reflects realistic scenarios where AI assistants must perceive and reason from raw data. However, existing pipelines and multi-modal LLMs perform poorly on Omni-MMSI because they lack reliable identity attribution capabilities, which leads to inaccurate social interaction understanding. To address this challenge, we propose Omni-MMSI-R, a reference-guided pipeline that produces identity-attributed social cues with tools and conducts chain-of-thought social reasoning. To facilitate this pipeline, we construct participant-level reference pairs and curate reasoning annotations on top of the existing datasets. Experiments demonstrate that Omni-MMSI-R outperforms advanced LLMs and counterparts on Omni-MMSI. Project page: https://sampson-lee.github.io/omni-mmsi-project-page.

Method: Created OmniSch benchmark with 1,854 real-world schematic diagrams and four tasks: visual grounding of schematic entities, diagram-to-graph reasoning, geometric reasoning for layout-dependent weights, and tool-augmented agentic reasoning for visual search.

Result: Results show substantial gaps in current LMMs: unreliable fine-grained grounding, brittle layout-to-graph parsing, inconsistent global connectivity reasoning, and inefficient visual exploration of schematic engineering artifacts.

Conclusion: The benchmark reveals significant limitations in current multimodal models for interpreting complex engineering schematics, highlighting the need for improved visual understanding and reasoning capabilities in specialized domains.

Abstract: Recent large multimodal models (LMMs) have made rapid progress in visual grounding, document understanding, and diagram reasoning tasks. However, their ability to convert Printed Circuit Board (PCB) schematic diagrams into machine-readable spatially weighted netlist graphs, jointly capturing component attributes, connectivity, and geometry, remains largely underexplored, despite such graph representations are the backbone of practical electronic design automation (EDA) workflows. To bridge this gap, we introduce OmniSch, the first comprehensive benchmark designed to assess LMMs on schematic understanding and spatial netlist graph construction. OmniSch contains 1,854 real-world schematic diagrams and includes four tasks: (1) visual grounding for schematic entities, with 109.9K grounded instances aligning 423.4K diagram semantic labels to their visual regions; (2) diagram-to-graph reasoning, understanding topological relationship among diagram elements; (3) geometric reasoning, constructing layout-dependent weights for each connection; and (4) tool-augmented agentic reasoning for visual search, invoking external tools to accomplish (1)-(3). Our results reveal substantial gaps of current LMMs in interpreting schematic engineering artifacts, including unreliable fine-grained grounding, brittle layout-to-graph parsing, inconsistent global connectivity reasoning and inefficient visual exploration.

Method: ProOOD combines prototype-guided semantic imputation (filling occluded regions with class-consistent features), prototype-guided tail mining (strengthening rare-class representations), and EchoOOD (fusing local logit coherence with local/global prototype matching for voxel-level OOD scoring).

Result: Achieves state-of-the-art performance on five datasets: +3.57% mIoU overall and +24.80% tail-class mIoU on SemanticKITTI, and +19.34 points AuPRCr improvement on VAA-KITTI, with consistent gains across benchmarks.

Conclusion: ProOOD provides more calibrated occupancy estimates and reliable OOD detection for safety-critical urban driving, addressing key limitations in current 3D semantic occupancy prediction systems.

Abstract: 3D semantic occupancy prediction is central to autonomous driving, yet current methods are vulnerable to long-tailed class bias and out-of-distribution (OOD) inputs, often overconfidently assigning anomalies to rare classes. We present ProOOD, a lightweight, plug-and-play method that couples prototype-guided refinement with training-free OOD scoring. ProOOD comprises (i) prototype-guided semantic imputation that fills occluded regions with class-consistent features, (ii) prototype-guided tail mining that strengthens rare-class representations to curb OOD absorption, and (iii) EchoOOD, which fuses local logit coherence with local and global prototype matching to produce reliable voxel-level OOD scores. Extensive experiments on five datasets demonstrate that ProOOD achieves state-of-the-art performance on both in-distribution 3D occupancy prediction and OOD detection. On SemanticKITTI, it surpasses baselines by +3.57% mIoU overall and +24.80% tail-class mIoU; on VAA-KITTI, it improves AuPRCr by +19.34 points, with consistent gains across benchmarks. These improvements yield more calibrated occupancy estimates and more reliable OOD detection in safety-critical urban driving. The source code is publicly available at https://github.com/7uHeng/ProOOD.

Method: EASe introduces SAUCE to excite low-resolution FM feature channels for selective calibration and attends across spatially-encoded image and FM features. It then uses CAFE, a training-free feature aggregator that leverages SAUCE attention scores as semantic grouping signals to segment aggregated features into multi-granularity masks at pixel-level representations.

Result: EASe demonstrates superior performance over previous state-of-the-art methods across major standard benchmarks and diverse datasets with complex morphologies.

Conclusion: EASe enables accurate fine-grained dense semantic mask discovery by operating directly at pixel-level feature representations, overcoming limitations of coarse patch-level approaches in complex scenes.

Abstract: Unsupervised segmentation approaches have increasingly leveraged foundation models (FM) to improve salient object discovery. However, these methods often falter in scenes with complex, multi-component morphologies, where fine-grained structural detail is indispensable. Many state-of-the-art unsupervised segmentation pipelines rely on mask discovery approaches that utilize coarse, patch-level representations. These coarse representations inherently suppress the fine-grained detail required to resolve such complex morphologies. To overcome this limitation, we propose Excite, Attend and Segment (EASe), an unsupervised domain-agnostic semantic segmentation framework for easy fine-grained mask discovery across challenging real-world scenes. EASe utilizes novel Semantic-Aware Upsampling with Channel Excitation (SAUCE) to excite low-resolution FM feature channels for selective calibration and attends across spatially-encoded image and FM features to recover full-resolution semantic representations. Finally, EASe segments the aggregated features into multi-granularity masks using a novel training-free Cue-Attentive Feature Aggregator (CAFE) which leverages SAUCE attention scores as a semantic grouping signal. EASe, together with SAUCE and CAFE, operate directly at pixel-level feature representations to enable accurate fine-grained dense semantic mask discovery. Our evaluation demonstrates superior performance of EASe over previous state-of-the-arts (SOTAs) across major standard benchmarks and diverse datasets with complex morphologies. Code is available at https://ease-project.github.io

Method: Proposes TPC-CMA (Three-Phase Curriculum for Cross-Modal Alignment) that decomposes modality gap into Centroid Gap and Distribution Gap. Uses a fine-tuning framework with gradient-aware scheduling that progressively introduces alignment during training to reduce both components simultaneously.

Result: Significantly improves cross-modal alignment: reduces modality gap by 66.6% with only 4.84% accuracy drop at α_target=0.05, and by 82.3% at α_target=0.5. Improves clustering ARI from 0.318 to 0.516 and captioning CIDEr by 57.1% over original model.

Conclusion: The Distribution Gap is the true predictor of cross-modal task quality, not the Raw Gap. TPC-CMA effectively addresses both centroid and distribution gaps, enabling better cross-modal interchangeability in VLMs.

Abstract: Vision-Language Models (VLMs) such as CLIP learn a shared embedding space for images and text, yet their representations remain geometrically separated, a phenomenon known as the modality gap. This gap limits tasks requiring cross-modal interchangeability, such as captioning and joint clustering. Existing post-processing approaches can partially improve cross-modal compatibility; however, we show through geometric analysis that they primarily reduce the global centroid offset while leaving the underlying distributional mismatch intact. We decompose the modality gap into a Centroid Gap and a Distribution Gap, and demonstrate that the Distribution Gap is the true predictor of cross-modal task quality ($R^2 = 0.986$), whereas the commonly used Raw Gap is misleading ($R^2 = 0.691$). Motivated by this observation, we propose TPC-CMA (Three-Phase Curriculum for Cross-Modal Alignment), a fine-tuning framework that explicitly reduces both components. The proposed CMA jointly mitigates centroid offsets and reshapes the distributional structure, while a three-phase curriculum with gradient-aware scheduling progressively introduces alignment during training to enable stable optimization. Experiments demonstrate that our method significantly improves cross-modal alignment. With $α_{\text{target}}{=}0.05$, the modality gap is reduced by 66.6% with only 4.84% accuracy drop. Under stronger alignment ($α_{\text{target}}{=}0.5$), the gap is reduced by 82.3%, clustering ARI improves from 0.318 to 0.516, and captioning CIDEr increases by 57.1% over the original model. Our code and pre-trained models will be made publicly available upon acceptance.

Method: Develops an invertible mixing-unmixing process via bi-directional GAN framework with cycle consistency constraints and linkage between linear and nonlinear mixtures, called LCGU net.

Result: Experimental results show LCGU net exhibits stable and competitive performance across different datasets compared to state-of-the-art model-based HNU methods.

Conclusion: The proposed approach successfully addresses hyperspectral nonlinear unmixing without explicit mixing model knowledge, demonstrating effectiveness through generative modeling principles.

Abstract: Due to the large footprint of pixels in remote sensing imagery, hyperspectral unmixing (HU) has become an important and necessary procedure in hyperspectral image analysis. Traditional HU methods rely on a prior spectral mixing model, especially for nonlinear mixtures, which has largely limited the performance and generalization capacity of the unmixing approach. In this paper, we address the challenging problem of hyperspectral nonlinear unmixing (HNU) without explicit knowledge of the mixing model. Inspired by the principle of generative models, where images of the same distribution can be generated as that of the training images without knowing the exact probability distribution function of the image, we develop an invertible mixing-unmixing process via a bi-directional GAN framework, constrained by both the cycle consistency and the linkage between linear and nonlinear mixtures. The combination of cycle consistency and linear linkage provides powerful constraints without requiring an explicit mixing model. We refer to the proposed approach as the linearly-constrained CycleGAN unmixing net, or LCGU net. Experimental results indicate that the proposed LCGU net exhibits stable and competitive performance across different datasets compared with other state-of-the-art model-based HNU methods.

Method: Extends the SANA family by removing textual conditioning entirely, using paired source-target images to learn a conditional flow-matching model in latent space. Learns a conditional velocity field mapping target image distributions, with synthetic data generation for training when real paired data is unavailable.

Result: Effectively suppresses motion artifacts while preserving anatomical structure in fetal MRI, achieving competitive performance with few inference steps.

Conclusion: Flow-based, text-free generative models are efficient and suitable for supervised image-to-image tasks in medical imaging, particularly when language prompts are unavailable or irrelevant.

Abstract: We propose SANA-I2I, a text-free high-resolution image-to-image generation framework that extends the SANA family by removing textual conditioning entirely. In contrast to SanaControlNet, which combines text and image-based control, SANA-I2I relies exclusively on paired source-target images to learn a conditional flow-matching model in latent space. The model learns a conditional velocity field that maps a target image distribution to another one, enabling supervised image translation without reliance on language prompts. We evaluate the proposed approach on the challenging task of fetal MRI motion artifact reduction. To enable paired training in this application, where real paired data are difficult to acquire, we adopt a synthetic data generation strategy based on the method proposed by Duffy et al., which simulates realistic motion artifacts in fetal magnetic resonance imaging (MRI). Experimental results demonstrate that SANA-I2I effectively suppresses motion artifacts while preserving anatomical structure, achieving competitive performance few inference steps. These results highlight the efficiency and suitability of our proposed flow-based, text-free generative models for supervised image-to-image tasks in medical imaging.

Method: Use frozen DINOv3 ViT-B embeddings without fine-tuning, propagate small labeled seed sets through unlabeled data via nearest-neighbor self-training, evaluate on AQUA20 benchmark (20 marine species).

Result: With <5% labels, self-training on frozen embeddings closes most gap to fully supervised baseline; at full supervision, gap narrows to few percentage points, some species exceed supervised baseline. High class separability in embedding space even with extreme label scarcity.

Conclusion: Frozen foundation model embeddings capture discriminative structure for marine species recognition, enabling practical label-efficient classification without training, domain engineering, or underwater-adapted models.

Abstract: Automated species classification from underwater imagery is bottlenecked by the cost of expert annotation, and supervised models trained on one dataset rarely transfer to new conditions. We investigate whether semi-supervised methods operating on frozen foundation model embeddings can close this annotation gap with minimal labeling effort. Using DINOv3 ViT-B embeddings with no fine-tuning, we propagate a small set of labeled seeds through unlabeled data via nearest-neighbor-based self-training and evaluate on the AQUA20 benchmark (20 marine species). With fewer than 5% of the training labels, self-training on frozen embeddings closes much of the gap to a fully supervised ConvNeXt baseline trained on the entire labeled dataset; at full supervision, the gap narrows to a few percentage points, with several species exceeding the supervised baseline. Class separability in the embedding space, measured by ROC-AUC, is high even at extreme label scarcity, indicating that the frozen representations capture discriminative structure well before decision boundaries can be reliably estimated. Our approach requires no training, no domain-specific data engineering, and no underwater-adapted models, establishing a practical, immediately deployable baseline for label-efficient marine species recognition. All results are reported on the held-out test set over 100 random seed initializations.

Method: Proposes VADMamba++ with Gray-to-RGB paradigm that reconstructs grayscale frames into RGB space, forcing inference of color from structure. Uses hybrid Mamba-CNN-Transformer backbone to capture diverse normal patterns. Implements intra-task fusion scoring combining explicit future-frame prediction errors with implicit quantized feature errors.

Result: Extensive experiments on three benchmark datasets show VADMamba++ outperforms state-of-the-art methods while maintaining performance and efficiency, especially under strict single-task settings with only frame-level inputs.

Conclusion: VADMamba++ successfully addresses limitations of previous methods by eliminating need for optical flow and multi-task fusion, achieving superior performance through Gray-to-RGB reconstruction and hybrid modeling approach.

Abstract: VADMamba pioneered the introduction of Mamba to Video Anomaly Detection (VAD), achieving high accuracy and fast inference through hybrid proxy tasks. Nevertheless, its heavy reliance on optical flow as auxiliary input and inter-task fusion scoring constrains its applicability to a single proxy task. In this paper, we introduce VADMamba++, an efficient VAD method based on the Gray-to-RGB paradigm that enforces a Single-Channel to Three-Channel reconstruction mapping, designed for a single proxy task and operating without auxiliary inputs. This paradigm compels inferring color appearances from grayscale structures, allowing anomalies to be more effectively revealed through dual inconsistencies between structure and chromatic cues. Specifically, VADMamba++ reconstructs grayscale frames into the RGB space to simultaneously discriminate structural geometry and chromatic fidelity, thereby enhancing sensitivity to explicit visual anomalies. We further design a hybrid modeling backbone that integrates Mamba, CNN, and Transformer modules to capture diverse normal patterns while suppressing the appearance of anomalies. Furthermore, an intra-task fusion scoring strategy integrates explicit future-frame prediction errors with implicit quantized feature errors, further improving accuracy under a single task setting. Extensive experiments on three benchmark datasets demonstrate that VADMamba++ outperforms state-of-the-art methods while meeting performance and efficiency, especially under a strict single-task setting with only frame-level inputs.

Method: Proposes PULSAR-Net with U-Net architecture and axial spatial attention to identify attack signals in full-waveform data, plus a physics-aware synthetic dataset generation pipeline for training.

Result: Achieves 92% and 73% reconstruction rates for vehicles obscured by jamming attacks in real-world static and driving scenarios, despite training only on synthetic data.

Conclusion: Full-waveform data contains distinguishable signatures that enable reconstruction of authentic point clouds under jamming attacks, enhancing LiDAR security for autonomous driving.

Abstract: LiDAR sensors are critical for autonomous driving perception, yet remain vulnerable to spoofing attacks. Jamming attacks inject high-frequency laser pulses that completely blind LiDAR sensors by overwhelming authentic returns with malicious signals. We discover that while point clouds become randomized, the underlying full-waveform data retains distinguishable signatures between attack and legitimate signals. In this work, we propose PULSAR-Net, capable of reconstructing authentic point clouds under jamming attacks by leveraging previously underutilized intermediate full-waveform representations and simultaneous laser sensing in modern LiDAR systems. PULSAR-Net adopts a novel U-Net architecture with axial spatial attention mechanisms specifically designed to identify attack-induced signals from authentic object returns in the full-waveform representation. To address the lack of full-waveform representations in existing LiDAR datasets under jamming attacks, we introduce a physics-aware dataset generation pipeline that synthesizes realistic full-waveform representations under jamming attacks. Despite being trained exclusively on synthetic data, PULSAR-Net achieves reconstruction rates of 92% and 73% for vehicles obscured by jamming attacks in real-world static and driving scenarios, respectively.

Method: Proposes a dynamic graph model with adaptive node selection to capture key local features from RGB and depth. Nodes are grouped by three distance levels (near/far relations), and the graph is updated dynamically using attention weights. Features from both modalities are fused for recognition.

Result: Superior performance on SUN RGB-D and NYU Depth v2 datasets compared to state-of-the-art methods, demonstrating effective exploitation of crucial local features from both RGB and depth modalities.

Conclusion: The dynamic graph model with adaptive node selection successfully addresses the problem of selecting and exploiting key local features in RGB-D indoor scene recognition, achieving improved performance through effective multimodal feature fusion.

Abstract: Multi-modality of color and depth, i.e., RGB-D, is of great importance in recent research of indoor scene recognition. In this kind of data representation, depth map is able to describe the 3D structure of scenes and geometric relations among objects. Previous works showed that local features of both modalities are vital for promotion of recognition accuracy. However, the problem of adaptive selection and effective exploitation on these key local features remains open in this field. In this paper, a dynamic graph model is proposed with adaptive node selection mechanism to solve the above problem. In this model, a dynamic graph is built up to model the relations among objects and scene, and a method of adaptive node selection is proposed to take key local features from both modalities of RGB and depth for graph modeling. After that, these nodes are grouped by three different levels, representing near or far relations among objects. Moreover, the graph model is updated dynamically according to attention weights. Finally, the updated and optimized features of RGB and depth modalities are fused together for indoor scene recognition. Experiments are performed on public datasets including SUN RGB-D and NYU Depth v2. Extensive results demonstrate that our method has superior performance when comparing to state-of-the-arts methods, and show that the proposed method is able to exploit crucial local features from both modalities of RGB and depth.

Method: Proposes UCMNet with uncertainty-aware adaptive processing: learns uncertainty maps through uncertainty-driven loss to quantify spatial uncertainty, uses these maps to guide a Memory Bank to retrieve region-adaptive context from a Context Bank, enabling effective modeling of non-uniform degradation characteristics.

Result: Achieves state-of-the-art performance on multiple benchmarks with 30% fewer parameters than previous models.

Conclusion: UCMNet effectively addresses spatially varying degradations in UDC imaging through uncertainty-aware adaptive processing, enabling superior detail recovery with reduced computational complexity.

Abstract: Under-display cameras (UDCs) allow for full-screen designs by positioning the imaging sensor underneath the display. Nonetheless, light diffraction and scattering through the various display layers result in spatially varying and complex degradations, which significantly reduce high-frequency details. Current PSF-based physical modeling techniques and frequency-separation networks are effective at reconstructing low-frequency structures and maintaining overall color consistency. However, they still face challenges in recovering fine details when dealing with complex, spatially varying degradation. To solve this problem, we propose a lightweight \textbf{U}ncertainty-aware \textbf{C}ontext-\textbf{M}emory \textbf{Network} (\textbf{UCMNet}), for UDC image restoration. Unlike previous methods that apply uniform restoration, UCMNet performs uncertainty-aware adaptive processing to restore high-frequency details in regions with varying degradations. The estimated uncertainty maps, learned through an uncertainty-driven loss, quantify spatial uncertainty induced by diffraction and scattering, and guide the Memory Bank to retrieve region-adaptive context from the Context Bank. This process enables effective modeling of the non-uniform degradation characteristics inherent to UDC imaging. Leveraging this uncertainty as a prior, UCMNet achieves state-of-the-art performance on multiple benchmarks with 30% fewer parameters than previous models. Project page: \href{https://kdhrick2222.github.io/projects/UCMNet/}{https://kdhrick2222.github.io/projects/UCMNet}.

Method: Multi-modal framework combining mmWave radar with RGBD input. Uses ResNet-based classifier to extract visual semantic cues, conditional latent diffusion model to synthesize expected mmWave spectrum, and dual-input comparison module to identify spatial deviations between real and generated spectra.

Result: Achieves up to 94% F1 score and sub-meter localization error across three applications: concealed weapon localization, through-wall intruder localization, and through-wall fall localization. Demonstrates robust generalization across clothing, occlusions, and cluttered environments.

Conclusion: mmAnomaly establishes an accurate and interpretable framework for context-aware anomaly detection in mmWave sensing by effectively combining visual context with radar data.

Abstract: mmWave radar enables human sensing in non-visual scenarios-e.g., through clothing or certain types of walls-where traditional cameras fail due to occlusion or privacy limitations. However, robust anomaly detection with mmWave remains challenging, as signal reflections are influenced by material properties, clutter, and multipath interference, producing complex, non-Gaussian distortions. Existing methods lack contextual awareness and misclassify benign signal variations as anomalies. We present mmAnomaly, a multi-modal anomaly detection framework that combines mmWave radar with RGBD input to incorporate visual context. Our system extracts semantic cues-such as scene geometry and material properties-using a fast ResNet-based classifier, and uses a conditional latent diffusion model to synthesize the expected mmWave spectrum for the given visual context. A dual-input comparison module then identifies spatial deviations between real and generated spectra to localize anomalies. We evaluate mmAnomaly on two multi-modal datasets across three applications: concealed weapon localization, through-wall intruder localization, and through-wall fall localization. The system achieves up to 94% F1 score and sub-meter localization error, demonstrating robust generalization across clothing, occlusions, and cluttered environments. These results establish mmAnomaly as an accurate and interpretable framework for context-aware anomaly detection in mmWave sensing.

Method: Proposes Mine-JEPA, an in-domain SSL pipeline using SIGReg (regularization-based SSL loss) to pretrain on only 1,170 unlabeled sonar images. Uses a compact ViT-Tiny backbone and compares against fine-tuned DINOv3 foundation model. Evaluates on binary mine vs. non-mine and 3-class mine-like object classification tasks.

Result: Mine-JEPA achieves F1 score of 0.935 in binary classification (outperforming DINOv3’s 0.922) and 0.820 in 3-class classification with synthetic augmentation (outperforming DINOv3’s 0.810). In-domain SSL on foundation models degrades performance by 10-13 percentage points. Compact ViT-Tiny achieves competitive performance with 4x fewer parameters than DINOv3.

Conclusion: Carefully designed in-domain self-supervised learning is a viable alternative to much larger foundation models in data-scarce maritime sonar imagery. Stronger pretrained models don’t always benefit from additional domain adaptation in specialized domains.

Abstract: Side-scan sonar (SSS) mine classification is a challenging maritime vision problem characterized by extreme data scarcity and a large domain gap from natural images. While self-supervised learning (SSL) and general-purpose vision foundation models have shown strong performance in general vision and several specialized domains, their use in SSS remains largely unexplored. We present Mine-JEPA, the first in-domain SSL pipeline for SSS mine classification, using SIGReg, a regularization-based SSL loss, to pretrain on only 1,170 unlabeled sonar images. In the binary mine vs. non-mine setting, Mine-JEPA achieves an F1 score of 0.935, outperforming fine-tuned DINOv3 (0.922), a foundation model pretrained on 1.7B images. For 3-class mine-like object classification, Mine-JEPA reaches 0.820 with synthetic data augmentation, again outperforming fine-tuned DINOv3 (0.810). We further observe that applying in-domain SSL to foundation models degrades performance by 10–13 percentage points, suggesting that stronger pretrained models do not always benefit from additional domain adaptation. In addition, Mine-JEPA with a compact ViT-Tiny backbone achieves competitive performance while using 4x fewer parameters than DINOv3. These results suggest that carefully designed in-domain self-supervised learning is a viable alternative to much larger foundation models in data-scarce maritime sonar imagery.

Method: TEP is a training-free approach that leverages external tracking models and Multimodal Large Language Models (MLLMs) to generate tracking-enhanced prompts. These prompts help SAM3 better understand challenging targets like tiny and semantic-dominated objects.

Result: Achieved first place (56.91%) on the test set of the PVUW Challenge 2026: Complex Video Object Segmentation Track, demonstrating improved performance over SAM3 on challenging targets.

Conclusion: Tracking-enhanced prompts from external models can effectively enhance SAM3’s capability for complex video object segmentation, particularly for difficult targets like tiny and semantic-dominated objects, without requiring retraining.

Abstract: In the Complex Video Object Segmentation task, researchers are required to track and segment specific targets within cluttered environments, which rigorously tests a method’s capability for target comprehension and environmental adaptability. Although SAM3, the current state-of-the-art solution, exhibits unparalleled segmentation performance and robustness on conventional targets, it underperforms on tiny and semantic-dominated objects. The root cause of this limitation lies in SAM3’s insufficient comprehension of these specific target types. To address this issue, we propose TEP: Advancing Complex Video Object Segmentation via Tracking-Enhanced Prompts. As a training-free approach, TEP leverages external tracking models and Multimodal Large Language Models to introduce tracking-enhanced prompts, thereby alleviating the difficulty SAM3 faces in understanding these challenging targets. Our method achieved first place (56.91%) on the test set of the PVUW Challenge 2026: Complex Video Object Segmentation Track.

Method: Proposes SKETCH as a new visual modality using edge-based detectors to extract high-frequency structural cues (limb articulations, self-occlusion contours) directly from RGB images without semantic labels. Introduces SKETCHGAIT framework with two independent streams for modality-specific learning and lightweight early-stage fusion to capture structural complementarity between label-guided parsing and label-free sketch.

Result: Achieves 92.9% Rank-1 on SUSTech1K and 93.1% mean Rank-1 on CCPG datasets, demonstrating superior performance over existing approaches.

Conclusion: Sketch provides a valuable complementary modality to parsing for gait recognition, offering dense structural information without semantic labels, and the multi-modal framework effectively leverages both representations for improved performance.

Abstract: Gait recognition is a non-intrusive biometric technique for security applications, yet existing studies are dominated by silhouette- and parsing-based representations. Silhouettes are sparse and miss internal structural details, limiting discriminability. Parsing enriches silhouettes with part-level structures, but relies heavily on upstream human parsers (e.g., label granularity and boundary precision), leading to unstable performance across datasets and sometimes even inferior results to silhouettes. We revisit gait representations from a structural perspective and describe a design space defined by edge density and supervision form: silhouettes use sparse boundary edges with weak single-label supervision, while parsing uses denser cues with strong semantic priors. In this space, we identify an underexplored paradigm: dense part-level structure without explicit semantic labels, and introduce SKETCH as a new visual modality for gait recognition. Sketch extracts high-frequency structural cues (e.g., limb articulations and self-occlusion contours) directly from RGB images via edge-based detectors in a label-free manner. We further show that label-guided parsing and label-free sketch are semantically decoupled and structurally complementary. Based on this, we propose SKETCHGAIT, a hierarchically disentangled multi-modal framework with two independent streams for modality-specific learning and a lightweight early-stage fusion branch to capture structural complementarity. Extensive experiments on SUSTech1K and CCPG validate the proposed modality and framework: SketchGait achieves 92.9% Rank-1 on SUSTech1K and 93.1% mean Rank-1 on CCPG.

Method: Proposes VLM-in-the-Loop, a plug-in module that wraps any digitization backend with closed-loop VLM feedback via standardized interface. Uses tool grounding - anchoring VLM assessment in quantitative evidence from domain-specific signal analysis tools.

Result: Tool grounding raises verdict consistency from 71% to 89% and doubles fidelity separation. Improves all four tested backends: 29.4% borderline leads improved, 41.2% failed limb leads recovered, valid leads per image doubled from 2.5 to 5.8. On 428 real clinical HCM images, reaches 98.0% Excellent quality.

Conclusion: The plug-in architecture and tool-grounding mechanism are domain-parametric and broadly applicable wherever quality criteria are objectively measurable, suggesting wider applications beyond ECG digitization.

Abstract: ECG digitization could unlock billions of archived clinical records, yet existing methods collapse on real-world images despite strong benchmark numbers. We introduce \textbf{VLM-in-the-Loop}, a plug-in quality assurance module that wraps any digitization backend with closed-loop VLM feedback via a standardized interface, requiring no modification to the underlying digitizer. The core mechanism is \textbf{tool grounding}: anchoring VLM assessment in quantitative evidence from domain-specific signal analysis tools. In a controlled ablation on 200 records with paired ground truth, tool grounding raises verdict consistency from 71% to 89% and doubles fidelity separation ($Δ$PCC 0.03 $\rightarrow$ 0.08), with the effect replicating across three VLMs (Claude Opus4, GPT-4o, Gemini2.5 Pro), confirming a pattern-level rather than model-specific gain. Deployed across four backends, the module improves every one: 29.4% of borderline leads improved on our pipeline; 41.2% of failed limb leads recovered on ECG-Digitiser; valid leads per image doubled on Open-ECG-Digitizer (2.5 $\rightarrow$ 5.8). On 428 real clinical HCM images, the integrated system reaches 98.0% Excellent quality. Both the plug-in architecture and tool-grounding mechanism are domain-parametric, suggesting broader applicability wherever quality criteria are objectively measurable.

Method: Proposes VAE-MMD preprocessing pipeline combining variational autoencoders with maximum mean discrepancy loss, incorporating skip connections and self-attention mechanisms alongside nnU-Net segmentation for domain adaptation across four public databases.

Result: VAE-MMD reduced domain classifier accuracy from 0.91 to 0.50 (successful feature alignment), maintained anatomical accuracy with PSNR >36 dB, and improved segmentation metrics: mean F1 increased 11.1%, mean surface Dice increased 7.93%, and mean 95th percentile Hausdorff distance reduced 65.5%.

Conclusion: VAE-MMD effectively reduces cross-institutional data heterogeneity and enhances brain metastases segmentation generalization across volumetric, detection, and boundary-level metrics without requiring target-domain labels, overcoming a key obstacle to clinical AI implementation.

Abstract: Background: Deep learning has demonstrated significant potential for automated brain metastases (BM) segmentation; however, models trained at a singular institution often exhibit suboptimal performance at various sites due to disparities in scanner hardware, imaging protocols, and patient demographics. The goal of this work is to create a domain adaptation framework that will allow for BM segmentation to be used across multiple institutions. Methods: We propose a VAE-MMD preprocessing pipeline that combines variational autoencoders (VAE) with maximum mean discrepancy (MMD) loss, incorporating skip connections and self-attention mechanisms alongside nnU-Net segmentation. The method was tested on 740 patients from four public databases: Stanford, UCSF, UCLM, and PKG, evaluated by domain classifier’s accuracy, sensitivity, precision, F1/F2 scores, surface Dice (sDice), and 95th percentile Hausdorff distance (HD95). Results: VAE-MMD reduced domain classifier accuracy from 0.91 to 0.50, indicating successful feature alignment across institutions. Reconstructed volumes attained a PSNR greater than 36 dB, maintaining anatomical accuracy. The combined method raised the mean F1 by 11.1% (0.700 to 0.778), the mean sDice by 7.93% (0.7121 to 0.7686), and reduced the mean HD95 by 65.5% (11.33 to 3.91 mm) across all four centers compared to the baseline nnU-Net. Conclusions: VAE-MMD effectively diminishes cross-institutional data heterogeneity and enhances BM segmentation generalization across volumetric, detection, and boundary-level metrics without necessitating target-domain labels, thereby overcoming a significant obstacle to the clinical implementation of AI-assisted segmentation.

Method: Investigates adaptability of Query-Centric Trajectory Prediction (QCNet) transferred from US to Korean data using four training strategies: zero-shot transfer, training from scratch, full fine-tuning, and encoder freezing with selective decoder fine-tuning.

Result: Leveraging pretrained knowledge significantly improves prediction performance. Selectively fine-tuning decoder while freezing encoder yields best trade-off between accuracy and training efficiency, reducing prediction error by over 66% compared to training from scratch.

Conclusion: Provides practical insights into effective transfer learning strategies for deploying trajectory prediction models in new geographic domains, demonstrating the importance of domain adaptation for autonomous driving systems.

Abstract: Developing robust models to accurately predict the trajectories of surrounding agents is fundamental to autonomous driving safety. However, most public datasets, such as the Waymo Open Motion Dataset and Argoverse, are collected in Western road environments and do not reflect the unique traffic patterns, infrastructure, and driving behaviors of other regions, including South Korea. This domain discrepancy leads to performance degradation when state-of-the-art models trained on Western data are deployed in different geographic contexts. In this work, we investigate the adaptability of Query-Centric Trajectory Prediction (QCNet) when transferred from U.S.-based data to Korean road environments. Using a Korean autonomous driving dataset, we compare four training strategies: zero-shot transfer, training from scratch, full fine-tuning, and encoder freezing. Experimental results demonstrate that leveraging pretrained knowledge significantly improves prediction performance. Specifically, selectively fine-tuning the decoder while freezing the encoder yields the best trade-off between accuracy and training efficiency, reducing prediction error by over 66% compared to training from scratch. This study provides practical insights into effective transfer learning strategies for deploying trajectory prediction models in new geographic domains.

Method: Three-stage pipeline: 1) Gemini-3.1 Pro decomposes target events into instance-level grounding targets, selects optimal frames, and generates discriminative descriptions; 2) SAM3-agent produces seed masks and SAM3 tracker propagates through video; 3) Qwen3.5-Plus with behavior-level verification refines predictions.

Result: Ranked first on PVUW 2026 MeViS-Text test set with Final score of 0.909064 and J&F score of 0.7897, without task-specific fine-tuning.

Conclusion: Demonstrates that combining strong multimodal LLMs with specialized vision models (SAM3) in a training-free pipeline can achieve state-of-the-art performance on complex video understanding tasks requiring joint appearance, temporal, and interaction understanding.

Abstract: This report presents our winning solution to the 5th PVUW MeViS-Text Challenge. The track studies referring video object segmentation under motion-centric language expressions, where the model must jointly understand appearance, temporal behavior, and object interactions. To address this problem, we build a fully training-free pipeline that combines strong multimodal large language models with SAM3. Our method contains three stages. First, Gemini-3.1 Pro decomposes each target event into instance-level grounding targets, selects the frame where the target is most clearly visible, and generates a discriminative description. Second, SAM3-agent produces a precise seed mask on the selected frame, and the official SAM3 tracker propagates the mask through the whole video. Third, a refinement stage uses Qwen3.5-Plus and behavior-level verification to correct ambiguous or semantically inconsistent predictions. Without task-specific fine-tuning, our method ranks first on the PVUW 2026 MeViS-Text test set, achieving a Final score of 0.909064 and a J&F score of 0.7897. The code is available at https://github.com/Moujuruo/MeViSv2_Track_Solution_2026.

Method: FADE employs two attack strategies: 1) Temporal Query Flooding generates spurious temporally consistent track queries to exhaust query budget, and 2) Temporal Memory Corruption attacks the query updater’s memory by severing temporal links and erasing feature identity. Also includes differentiable pipeline for physical-world realizability using sensor spoofing simulations.

Result: Experiments on MOT17 and MOT20 benchmarks show FADE is highly effective against state-of-the-art TBP trackers, causing significant identity switches and track terminations.

Conclusion: The paper demonstrates serious vulnerabilities in TBP tracking architectures and introduces a novel attack framework that exploits temporal propagation mechanisms, highlighting security concerns in modern MOT systems.

Abstract: Recent Tracking-by-Query-Propagation (TBP) methods have advanced Multi-Object Tracking (MOT) by enabling end-to-end (E2E) pipelines with long-range temporal modeling. However, this reliance on query propagation introduces unexplored architectural vulnerabilities to adversarial attacks. We present FADE, a novel attack framework designed to exploit these specific vulnerabilities. FADE employs two attack strategies targeting core TBP mechanisms: (i) Temporal Query Flooding: Generates spurious temporally consistent track queries to exhaust the tracker’s limited query budget, forcing it to terminate valid tracks. (ii) Temporal Memory Corruption: Directly attacks the query updater’s memory by severing temporal links via state de-correlation and erasing the learned feature identity of matched tracks. Furthermore, we introduce a differentiable pipeline to optimize these attacks for physical-world realizability by leveraging simulations of advanced perception sensor spoofing. Experiments on MOT17 and MOT20 benchmarks demonstrate that FADE is highly effective against state-of-the-art TBP trackers, causing significant identity switches and track terminations.

Method: First Logit Boosting (FLB) stores the logit of the first generated token and adds it to subsequent token predictions. This sustains visual information from the first token throughout generation and suppresses hallucinated words through the stabilizing effect of the “The” token.

Result: FLB significantly reduces object hallucination across various tasks, benchmarks, and backbone models. It causes negligible inference overhead, making it highly applicable to real-time multimodal systems.

Conclusion: FLB is a simple yet effective training-free technique that effectively mitigates long-term decay in LVLMs, reducing object hallucination with minimal computational cost.

Abstract: Recent Large Vision-Language Models (LVLMs) have demonstrated remarkable performance across various multimodal tasks that require understanding both visual and linguistic inputs. However, object hallucination – the generation of nonexistent objects in answers – remains a persistent challenge. Although several approaches such as retraining and external grounding methods have been proposed to mitigate this issue, they still suffer from high data costs or structural complexity. Training-free methods such as Contrastive Decoding (CD) are more cost-effective, avoiding additional training or external models, but still suffer from long-term decay, where visual grounding weakens and language priors dominate as the generation progresses. In this paper, we propose First Logit Boosting (FLB), a simple yet effective training-free technique designed to alleviate long-term decay in LVLMs. FLB stores the logit of the first generated token and adds it to subsequent token predictions, effectively mitigating long-term decay of visual information. We observe that FLB (1) sustains the visual information embedded in the first token throughout generation, and (2) suppresses hallucinated words through the stabilizing effect of the ``The’’ token. Experimental results show that FLB significantly reduces object hallucination across various tasks, benchmarks, and backbone models. Notably, it causes negligible inference overhead, making it highly applicable to real-time multimodal systems. Code is available at https://github.com/jiwooha20/FLB

Method: Trained 3D UNETR and SegFormer transformer models on 7T FLAIR scans at multiple resolutions, compared them against classical LST-LPA and LST-AI tools using BraTS 2023 metrics, and evaluated performance on held-out test sets.

Result: 7T-trained transformers achieved competitive overlap with LST-AI while detecting additional small lesions missed by classical methods, with SegFormer achieving voxel-wise Dice of 0.61 and lesion-wise Dice of 0.20, outperforming LST-LPA (Dice 0.39, lesion-wise Dice 0.02).

Conclusion: Native 7T resolution is crucial for small-lesion detection in MS, and transformer-based models trained on ultra-high field data provide superior performance compared to classical tools developed for lower-field MRI.

Abstract: Ultra-high field 7-tesla (7T) MRI improves visualization of multiple sclerosis (MS) white matter lesions (WML) but differs sufficiently in contrast and artifacts from 1.5-3T imaging - suggesting that widely used automated segmentation tools may not translate directly. We analyzed 7T FLAIR scans and generated reference WML masks from Lesion Segmentation Tool (LST) outputs followed by expert manual revision. As external comparators, we applied LST-LPA and the more recent LST-AI ensemble, both originally developed on lower-field data. We then trained 3D UNETR and SegFormer transformer-based models on 7T FLAIR at multiple resolutions (0.5x0.5x0.5^3, 1.0x1.0x1.0^3, and 1.5x1.5x2.0^3) and evaluated all methods using voxel-wise and lesion-wise metrics from the BraTS 2023 framework. On the held-out test set at native 0.5x0.5x0.5^3 resolution, 7T-trained transformers achieved competitive overlap with LST-AI while recovering additional small lesions that were missed by classical methods, at the cost of some boundary variability and occasional artifact-related false positives. On a held-out 7 T test set, our best transformer model (SegFormer) achieved a voxel-wise Dice of 0.61 and lesion-wise Dice of 0.20, improving on the classical LST-LPA tool (Dice 0.39, lesion-wise Dice 0.02). Performance decreased for models trained on downsampled images, underscoring the value of native 7T resolution for small-lesion detection. By releasing our 7T-trained models, we aim to provide a reproducible, ready-to-use resource for automated lesion quantification in ultra-high field MS research (https://github.com/maynord/7T-MS-lesion-segmentation).

Method: Proposes Multi-Group Policy Optimization (MUPO), which incentivizes divergent thinking across multiple solutions to maintain diverse reasoning strategies during training.

Result: Demonstrates effectiveness on established benchmarks, showing improved performance over GRPO by maintaining diverse reasoning patterns.

Conclusion: MUPO successfully addresses the diversity collapse problem in RL-trained VLMs, enabling broader reasoning capabilities while maintaining the benefits of RL optimization.

Abstract: Recent studies have demonstrated that Reinforcement Learning (RL), notably Group Relative Policy Optimization (GRPO), can intrinsically elicit and enhance the reasoning capabilities of Vision-Language Models (VLMs). However, despite the promise, the underlying mechanisms that drive the effectiveness of RL models as well as their limitations remain underexplored. In this paper, we highlight a fundamental behavioral distinction between RL and base models, where the former engages in deeper yet narrow reasoning, while base models, despite less refined along individual path, exhibit broader and more diverse thinking patterns. Through further analysis of training dynamics, we show that GRPO is prone to diversity collapse, causing models to prematurely converge to a limited subset of reasoning strategies while discarding the majority of potential alternatives, leading to local optima and poor scalability. To address this, we propose Multi-Group Policy Optimization (MUPO), a simple yet effective approach designed to incentivize divergent thinking across multiple solutions, and demonstrate its effectiveness on established benchmarks. Project page: https://xytian1008.github.io/MUPO/

Method: Two-stage training framework combining instruction tuning with reinforcement learning on 14.7M instruction/reasoning samples from 30 public datasets covering 36 CXR tasks.

Result: Outperforms existing medical and general-domain foundation models across 17 evaluation settings; clinical study shows CheXOne reports comparable/better than resident reports in 55% of cases with high clinical factuality.

Conclusion: Explicit reasoning improves model performance, interpretability, and clinical utility in AI-assisted CXR interpretation by providing clinically grounded explanations.

Abstract: Chest X-rays (CXRs) are among the most frequently performed imaging examinations worldwide, yet rising imaging volumes increase radiologist workload and the risk of diagnostic errors. Although artificial intelligence (AI) systems have shown promise for CXR interpretation, most generate only final predictions, without making explicit how visual evidence is translated into radiographic findings and diagnostic predictions. We present CheXOne, a reasoning-enabled vision-language model for CXR interpretation. CheXOne jointly generates diagnostic predictions and explicit, clinically grounded reasoning traces that connect visual evidence, radiographic findings, and these predictions. The model is trained on 14.7 million instruction and reasoning samples curated from 30 public datasets spanning 36 CXR interpretation tasks, using a two-stage framework that combines instruction tuning with reinforcement learning to improve reasoning quality. We evaluate CheXOne in zero-shot settings across visual question answering, report generation, visual grounding and reasoning assessment, covering 17 evaluation settings. CheXOne outperforms existing medical and general-domain foundation models and achieves strong performance on independent public benchmarks. A clinical reader study demonstrates that CheXOne-drafted reports are comparable to or better than resident-written reports in 55% of cases, while effectively addressing clinical indications and enhancing both report writing and CXR interpretation efficiency. Further analyses involving radiologists reveal that the generated reasoning traces show high clinical factuality and provide causal support for the final predictions, offering a plausible explanation for the performance gains. These results suggest that explicit reasoning can improve model performance, interpretability and clinical utility in AI-assisted CXR interpretation.

Method: Proposes auto-regressive Gaussian splatting (ARGS) with Gaussian simplification strategy and reverse simplification for next-scale generation. Uses hierarchical trees for O(log n) generation steps and tree-based transformer to predict structure auto-regressively with attention from leaf nodes to ancestors.

Result: Extensive experiments show the approach effectively generates multi-scale Gaussian representations with controllable detail levels, visual fidelity, and manageable time consumption.

Conclusion: ARGS successfully extends auto-regressive frameworks to 3D generation using Gaussian splatting, enabling efficient multi-scale 3D object generation with structural consistency.

Abstract: Auto-regressive frameworks for next-scale prediction of 2D images have demonstrated strong potential for producing diverse and sophisticated content by progressively refining a coarse input. However, extending this paradigm to 3D object generation remains largely unexplored. In this paper, we introduce auto-regressive Gaussian splatting (ARGS), a framework for making next-scale predictions in parallel for generation according to levels of detail. We propose a Gaussian simplification strategy and reverse the simplification to guide next-scale generation. Benefiting from the use of hierarchical trees, the generation process requires only (\mathcal{O}(\log n)) steps, where (n) is the number of points. Furthermore, we propose a tree-based transformer to predict the tree structure auto-regressively, allowing leaf nodes to attend to their internal ancestors to enhance structural consistency. Extensive experiments demonstrate that our approach effectively generates multi-scale Gaussian representations with controllable levels of detail, visual fidelity, and a manageable time consumption budget.

Method: PC-SAM integrates fully automatic and interactive segmentation in a unified framework with a fine-tuning strategy that constrains point prompt influence to corresponding patches, enabling fine local corrections and mask refinement.

Result: Extensive experiments on remote sensing road segmentation datasets show PC-SAM with point prompts significantly outperforms state-of-the-art fully automatic models while providing flexible local mask refinement and segmentation.

Conclusion: PC-SAM successfully addresses limitations of both automatic segmentation and SAM for remote sensing road segmentation, offering improved performance and fine-grained interactive refinement capabilities.

Abstract: Road masks obtained from remote sensing images effectively support a wide range of downstream tasks. In recent years, most studies have focused on improving the performance of fully automatic segmentation models for this task, achieving significant gains. However, current fully automatic methods are still insufficient for identifying certain challenging road segments and often produce false positive and false negative regions. Moreover, fully automatic segmentation does not support local segmentation of regions of interest or refinement of existing masks. Although the SAM model is widely used as an interactive segmentation model and performs well on natural images, it shows poor performance in remote sensing road segmentation and cannot support fine-grained local refinement. To address these limitations, we propose PC-SAM, which integrates fully automatic road segmentation and interactive segmentation within a unified framework. By carefully designing a fine-tuning strategy, the influence of point prompts is constrained to their corresponding patches, overcoming the inability of the original SAM to perform fine local corrections and enabling fine-grained interactive mask refinement. Extensive experiments on several representative remote sensing road segmentation datasets demonstrate that, when combined with point prompts, PC-SAM significantly outperforms state-of-the-art fully automatic models in road mask segmentation, while also providing flexible local mask refinement and local road segmentation. The code will be available at https://github.com/Cyber-CCOrange/PC-SAM.

Method: Proposes PET-DINO with Alignment-Friendly Visual Prompt Generation (AFVPG) module built on advanced text-prompted detector. Introduces two training strategies: Intra-Batch Parallel Prompting (IBP) at iteration level and Dynamic Memory-Driven Prompting (DMD) at overall training level for simultaneous modeling of multiple prompt routes.

Result: PET-DINO exhibits competitive zero-shot object detection capabilities across various prompt-based detection protocols, attributed to inheritance-based philosophy and prompt-enriched training strategies.

Conclusion: The approach effectively addresses open-set detection challenges through universal prompt support and novel training strategies, building an effective generic object detector.

Abstract: Open-Set Object Detection (OSOD) enables recognition of novel categories beyond fixed classes but faces challenges in aligning text representations with complex visual concepts and the scarcity of image-text pairs for rare categories. This results in suboptimal performance in specialized domains or with complex objects. Recent visual-prompted methods partially address these issues but often involve complex multi-modal designs and multi-stage optimizations, prolonging the development cycle. Additionally, effective training strategies for data-driven OSOD models remain largely unexplored. To address these challenges, we propose PET-DINO, a universal detector supporting both text and visual prompts. Our Alignment-Friendly Visual Prompt Generation (AFVPG) module builds upon an advanced text-prompted detector, addressing the limitations of text representation guidance and reducing the development cycle. We introduce two prompt-enriched training strategies: Intra-Batch Parallel Prompting (IBP) at the iteration level and Dynamic Memory-Driven Prompting (DMD) at the overall training level. These strategies enable simultaneous modeling of multiple prompt routes, facilitating parallel alignment with diverse real-world usage scenarios. Comprehensive experiments demonstrate that PET-DINO exhibits competitive zero-shot object detection capabilities across various prompt-based detection protocols. These strengths can be attributed to inheritance-based philosophy and prompt-enriched training strategies, which play a critical role in building an effective generic object detector. Project page: https://fuweifuvtoo.github.io/pet-dino.

Method: Proposes Relational Grounding Transformer (RegFormer) that leverages spatially grounded signals as guidance for interaction reasoning. It promotes locality-aware interaction learning to distinguish humans, objects, and their interactions, enabling direct transfer from image-level to instance-level reasoning.

Result: Extensive experiments show RegFormer effectively learns spatial cues for instance-level interaction reasoning, operates with high efficiency, and achieves performance comparable to fully supervised models.

Conclusion: RegFormer provides an efficient and accurate solution for weakly-supervised HOI detection by learning localized interaction cues that enable direct transfer to instance-level reasoning without additional training.

Abstract: Weakly-supervised Human-Object Interaction (HOI) detection is essential for scalable scene understanding, as it learns interactions from only image-level annotations. Due to the lack of localization signals, prior works typically rely on an external object detector to generate candidate pairs and then infer their interactions through pairwise reasoning. However, this framework often struggles to scale due to the substantial computational cost incurred by enumerating numerous instance pairs. In addition, it suffers from false positives arising from non-interactive combinations, which hinder accurate instance-level HOI reasoning. To address these issues, we introduce Relational Grounding Transformer (RegFormer), a versatile interaction recognition module for efficient and accurate HOI reasoning. Under image-level supervision, RegFormer leverages spatially grounded signals as guidance for the reasoning process and promotes locality-aware interaction learning. By learning localized interaction cues, our module distinguishes humans, objects, and their interactions, enabling direct transfer from image-level interaction reasoning to precise and efficient instance-level reasoning without additional training. Our extensive experiments and analyses demonstrate that RegFormer effectively learns spatial cues for instance-level interaction reasoning, operates with high efficiency, and even achieves performance comparable to fully supervised models. Our code is available at https://github.com/mlvlab/RegFormer.

Method: Proposes MAESIL framework using ‘superpatch’ 3D chunk-based input units to balance 3D context preservation with computational efficiency. Employs 3D masked autoencoder with dual-masking strategy to learn comprehensive spatial representations from partitioned volumes.

Result: Validated on three diverse large-scale public CT datasets. MAESIL demonstrates significant improvements over AE, VAE, and VQ-VAE in reconstruction metrics (PSNR and SSIM).

Conclusion: MAESIL establishes itself as a robust and practical pre-training solution for 3D medical imaging tasks by effectively capturing 3D structural information.

Abstract: Training deep learning models for three-dimensional (3D) medical imaging, such as Computed Tomography (CT), is fundamentally challenged by the scarcity of labeled data. While pre-training on natural images is common, it results in a significant domain shift, limiting performance. Self-Supervised Learning (SSL) on unlabeled medical data has emerged as a powerful solution, but prominent frameworks often fail to exploit the inherent 3D nature of CT scans. These methods typically process 3D scans as a collection of independent 2D slices, an approach that fundamentally discards critical axial coherence and the 3D structural context. To address this limitation, we propose the autoencoder for enhanced self-supervised medical image learning(MAESIL), a novel self-supervised learning framework designed to capture 3D structural information efficiently. The core innovation is the ‘superpatch’, a 3D chunk-based input unit that balances 3D context preservation with computational efficiency. Our framework partitions the volume into superpatches and employs a 3D masked autoencoder strategy with a dual-masking strategy to learn comprehensive spatial representations. We validated our approach on three diverse large-scale public CT datasets. Our experimental results show that MAESIL demonstrates significant improvements over existing methods such as AE, VAE and VQ-VAE in key reconstruction metrics such as PSNR and SSIM. This establishes MAESIL as a robust and practical pre-training solution for 3D medical imaging tasks.

Method: Proposes Individual-Behavior-Aware Network (IBA-Net) with two modules: 1) Mixture-of-Experts-based Feature Customization (MFC) that adaptively fuses data from multiple sampling rates to capture customized features for different behaviors, and 2) Neural Collapse-driven Classifier Calibration (NC3) that uses a fixed equiangular tight frame (ETF) classifier to maximize angles between classifier vectors and improve minority class performance.

Result: Experiments on three public datasets (goat, cattle, and horse activity recognition) demonstrate that IBA-Net consistently outperforms existing approaches across all datasets.

Conclusion: IBA-Net effectively addresses the challenges of suboptimal sampling rates and class imbalance in animal activity recognition, achieving high classification accuracy across all individual behaviors in farm animals.

Abstract: With the rapid advancements in deep learning techniques, wearable sensor-aided animal activity recognition (AAR) has demonstrated promising performance, thereby improving livestock management efficiency as well as animal health and welfare monitoring. However, existing research often prioritizes overall performance, overlooking the fact that classification accuracies for specific animal behavioral categories may remain unsatisfactory. This issue typically stems from suboptimal sampling rates or class imbalance problems. To address these challenges and achieve high classification accuracy across all individual behaviors in farm animals, we propose a novel Individual-Behavior-Aware Network (IBA-Net). This network enhances the recognition of each specific behavior by simultaneously customizing features and calibrating the classifier. Specifically, considering that different behaviors require varying sampling rates to achieve optimal performance, we design a Mixture-of-Experts (MoE)-based Feature Customization (MFC) module. This module adaptively fuses data from multiple sampling rates, capturing customized features tailored to various animal behaviors. Additionally, to mitigate classifier bias toward majority classes caused by class imbalance, we develop a Neural Collapse-driven Classifier Calibration (NC3) module. This module introduces a fixed equiangular tight frame (ETF) classifier during the classification stage, maximizing the angles between pair-wise classifier vectors and thereby improving the classification performance for minority classes. To validate the effectiveness of IBA-Net, we conducted experiments on three public datasets covering goat, cattle, and horse activity recognition. The results demonstrate that our method consistently outperforms existing approaches across all datasets.

Method: Proposes learnability-driven dataset distillation that constructs synthetic datasets incrementally through successive stages. Starting from a small set, trains a model and generates new samples guided by learnability scores that identify what the current model can learn from. Introduces Learnability-Guided Diffusion (LGD) which balances training utility for the current model with validity under a reference model to generate curriculum-aligned samples.

Result: Reduces redundancy by 39.1%, promotes specialization across training stages, and achieves state-of-the-art results on ImageNet-1K (60.1%), ImageNette (87.2%), and ImageWoof (72.9%).

Conclusion: Learnability-guided approach effectively addresses redundancy in dataset distillation by creating curriculum-aligned synthetic datasets that provide complementary rather than overlapping training signals, leading to improved performance on vision benchmarks.

Abstract: Training machine learning models on massive datasets is expensive and time-consuming. Dataset distillation addresses this by creating a small synthetic dataset that achieves the same performance as the full dataset. Recent methods use diffusion models to generate distilled data, either by promoting diversity or matching training gradients. However, existing approaches produce redundant training signals, where samples convey overlapping information. Empirically, disjoint subsets of distilled datasets capture 80-90% overlapping signals. This redundancy stems from optimizing visual diversity or average training dynamics without accounting for similarity across samples, leading to datasets where multiple samples share similar information rather than complementary knowledge. We propose learnability-driven dataset distillation, which constructs synthetic datasets incrementally through successive stages. Starting from a small set, we train a model and generate new samples guided by learnability scores that identify what the current model can learn from, creating an adaptive curriculum. We introduce Learnability-Guided Diffusion (LGD), which balances training utility for the current model with validity under a reference model to generate curriculum-aligned samples. Our approach reduces redundancy by 39.1%, promotes specialization across training stages, and achieves state-of-the-art results on ImageNet-1K (60.1%), ImageNette (87.2%), and ImageWoof (72.9%). Our code is available on our project page https://jachansantiago.github.io/learnability-guided-distillation/.

Method: Proposes TAB framework: 1) VLM agent dynamically invokes visual tools to track targets across 2D frames, 2) Semantic-Anchored Geometric Expansion anchors target in reference video and propagates spatial location across unobserved frames using multi-view geometry, 3) Builds 3D representation by aggregating multi-view features via camera parameters.

Result: Extensive experiments on ScanRefer and Nr3D show TAB significantly outperforms previous zero-shot methods and surpasses fully supervised baselines, using only open-source models.

Conclusion: The TAB framework successfully decouples 3D visual grounding into semantic understanding via 2D VLMs and geometric reconstruction, enabling dynamic processing of raw RGB-D streams without preprocessed 3D data.

Abstract: 3D Visual Grounding (3D-VG) aims to localize objects in 3D scenes via natural language descriptions. While recent advancements leveraging Vision-Language Models (VLMs) have explored zero-shot possibilities, they typically suffer from a static workflow relying on preprocessed 3D point clouds, essentially degrading grounding into proposal matching. To bypass this reliance, our core motivation is to decouple the task: leveraging 2D VLMs to resolve complex spatial semantics, while relying on deterministic multi-view geometry to instantiate the 3D structure. Driven by this insight, we propose “Think, Act, Build (TAB)”, a dynamic agentic framework that reformulates 3D-VG tasks as a generative 2D-to-3D reconstruction paradigm operating directly on raw RGB-D streams. Specifically, guided by a specialized 3D-VG skill, our VLM agent dynamically invokes visual tools to track and reconstruct the target across 2D frames. Crucially, to overcome the multi-view coverage deficit caused by strict VLM semantic tracking, we introduce the Semantic-Anchored Geometric Expansion, a mechanism that first anchors the target in a reference video clip and then leverages multi-view geometry to propagate its spatial location across unobserved frames. This enables the agent to “Build” the target’s 3D representation by aggregating these multi-view features via camera parameters, directly mapping 2D visual cues to 3D coordinates. Furthermore, to ensure rigorous assessment, we identify flaws such as reference ambiguity and category errors in existing benchmarks and manually refine the incorrect queries. Extensive experiments on ScanRefer and Nr3D demonstrate that our framework, relying entirely on open-source models, significantly outperforms previous zero-shot methods and even surpasses fully supervised baselines.

Method: 1) Formulates grading as generative color transformation producing 3D-LUTs conditioned on text or images; 2) Develops VQ-VAE tokenizer compressing 3×32³ LUT vectors to 64 discrete tokens; 3) Builds AceTone-800K dataset; 4) Trains vision-language model to predict LUT tokens; 5) Uses reinforcement learning to align outputs with perceptual fidelity and aesthetics.

Result: Achieves state-of-the-art performance on both text-guided and reference-guided grading tasks, improving LPIPS by up to 50% over existing methods. Human evaluations confirm visually pleasing and stylistically coherent results.

Conclusion: Demonstrates a new pathway toward language-driven, aesthetic-aligned color grading through multimodal conditioning within a unified framework.

Abstract: Color affects how we interpret image style and emotion. Previous color grading methods rely on patch-wise recoloring or fixed filter banks, struggling to generalize across creative intents or align with human aesthetic preferences. In this study, we propose AceTone, the first approach that supports multimodal conditioned color grading within a unified framework. AceTone formulates grading as a generative color transformation task, where a model directly produces 3D-LUTs conditioned on text prompts or reference images. We develop a VQ-VAE based tokenizer which compresses a $3\times32^3$ LUT vector to 64 discrete tokens with $ΔE<2$ fidelity. We further build a large-scale dataset, AceTone-800K, and train a vision-language model to predict LUT tokens, followed by reinforcement learning to align outputs with perceptual fidelity and aesthetics. Experiments show that AceTone achieves state-of-the-art performance on both text-guided and reference-guided grading tasks, improving LPIPS by up to 50% over existing methods. Human evaluations confirm that AceTone’s results are visually pleasing and stylistically coherent, demonstrating a new pathway toward language-driven, aesthetic-aligned color grading.

Method: 1) Physiological Bandpass Filtering to suppress out-of-band interference; 2) Physiological Spectrum Modulation with adaptive spectral selection to emphasize pulse-related frequencies; 3) Cross-domain Representation Learning to fuse spectral priors with deep time-domain features; 4) Frequency-aware conditional diffusion process for progressive rPPG signal reconstruction.

Result: Extensive experiments on six benchmarks demonstrate significant improvements over state-of-the-art approaches, particularly under challenging motion conditions. The method shows robustness to motion artifacts and illumination fluctuations.

Conclusion: FreqPhys highlights the importance of explicitly modeling physiological frequency priors for robust rPPG signal recovery. The frequency-guided approach with diffusion modeling effectively handles challenging real-world conditions.

Abstract: Remote photoplethysmography (rPPG) enables contactless physiological monitoring by capturing subtle skin-color variations from facial videos. However, most existing methods predominantly rely on time-domain modeling, making them vulnerable to motion artifacts and illumination fluctuations, where weak physiological clues are easily overwhelmed by noise. To address these challenges, we propose FreqPhys, a frequency-guided rPPG framework that explicitly leverages physiological frequency priors for robust signal recovery. Specifically, FreqPhys first applies a Physiological Bandpass Filtering module to suppress out-of-band interference, and then performs Physiological Spectrum Modulation together with adaptive spectral selection to emphasize pulse-related frequency components while suppress residual in-band noise. A Cross-domain Representation Learning module further fuses these spectral priors with deep time-domain features to capture informative spatial–temporal dependencies. Finally, a frequency-aware conditional diffusion process progressively reconstructs high-fidelity rPPG signals. Extensive experiments on six benchmarks demonstrate that FreqPhys yields significant improvements over state-of-the-art approaches, particularly under challenging motion conditions. It highlights the importance of explicitly modeling physiological frequency priors. The source code will be released.

Method: Proposes MATHENA with two main components: MATHE (multi-resolution SSM-driven detector with four-directional VSS blocks for O(N) global context modeling) generates per-tooth crops, and HENA (lightweight Mamba-UNet with triple-head architecture and Global Context State Token) processes crops. Uses upstream CarSeg training to establish shared representations, then freezes them for downstream AD fine-tuning and DDS classification via linear probing.

Result: Achieves 93.78% mAP@50 in tooth detection, 90.11% Dice for caries segmentation, 88.35% for anomaly detection, and 72.40% ACC for dental developmental staging. Also curates PARTHENON benchmark with 15,062 annotated instances from ten datasets.

Conclusion: MATHENA provides a unified, efficient framework for comprehensive dental diagnosis from OPGs, demonstrating strong performance across all four tasks while leveraging Mamba’s linear-complexity advantages for medical imaging.

Abstract: Dental diagnosis from Orthopantomograms (OPGs) requires coordination of tooth detection, caries segmentation (CarSeg), anomaly detection (AD), and dental developmental staging (DDS). We propose Mamba-based Architectural Tooth Hierarchical Estimator and Holistic Evaluation Network for Anatomy (MATHENA), a unified framework leveraging Mamba’s linear-complexity State Space Models (SSM) to address all four tasks. MATHENA integrates MATHE, a multi-resolution SSM-driven detector with four-directional Vision State Space (VSS) blocks for O(N) global context modeling, generating per-tooth crops. These crops are processed by HENA, a lightweight Mamba-UNet with a triple-head architecture and Global Context State Token (GCST). In the triple-head architecture, CarSeg is first trained as an upstream task to establish shared representations, which are then frozen and reused for downstream AD fine-tuning and DDS classification via linear probing, enabling stable, efficient learning. We also curate PARTHENON, a benchmark comprising 15,062 annotated instances from ten datasets. MATHENA achieves 93.78% mAP@50 in tooth detection, 90.11% Dice for CarSeg, 88.35% for AD, and 72.40% ACC for DDS.

Method: Proposes TRiGS with unified continuous geometric transformations: SE(3) transformations for rigid motions, hierarchical Bezier residuals for smooth deformation, and learnable local anchors for geometric consistency, preserving temporal identity of primitives.

Result: Achieves high-fidelity rendering on standard benchmarks while uniquely scaling to extended sequences (600-1200 frames) without severe memory bottlenecks, significantly outperforming prior works in temporal stability.

Conclusion: TRiGS provides a continuous 4D representation that effectively addresses temporal fragmentation and memory scalability issues in dynamic scene reconstruction, enabling longer video sequence processing with stable temporal identity.

Abstract: Recent 4D Gaussian Splatting (4DGS) methods achieve impressive dynamic scene reconstruction but often rely on piecewise linear velocity approximations and short temporal windows. This disjointed modeling leads to severe temporal fragmentation, forcing primitives to be repeatedly eliminated and regenerated to track complex nonlinear dynamics. This makeshift approximation eliminates the long-term temporal identity of objects and causes an inevitable proliferation of Gaussians, hindering scalability to extended video sequences. To address this, we propose TRiGS, a novel 4D representation that utilizes unified, continuous geometric transformations. By integrating $SE(3)$ transformations, hierarchical Bezier residuals, and learnable local anchors, TRiGS models geometrically consistent rigid motions for individual primitives. This continuous formulation preserves temporal identity and effectively mitigates unbounded memory growth. Extensive experiments demonstrate that TRiGS achieves high fidelity rendering on standard benchmarks while uniquely scaling to extended video sequences (e.g., 600 to 1200 frames) without severe memory bottlenecks, significantly outperforming prior works in temporal stability.

Method: A 3D convolutional neural network was trained on cognitively stable participants from the Alzheimer’s Disease Neuroimaging Initiative to learn the mapping between brain anatomy (structural MRI) and Neuropsychiatric Inventory Questionnaire (NPIQ) scores. Deviations between predicted and observed scores defined the Divergence from NPIQ scores (DNPI).

Result: Higher DNPI was significantly associated with future AD conversion (adjusted OR=2.5; p < 0.01) and achieved predictive accuracy comparable to cerebrospinal fluid AB42 biomarker (AUC=0.74 vs 0.75).

Conclusion: The deep learning-based normative modelling framework provides a scalable, non-invasive strategy for early Alzheimer’s disease detection by identifying atypical neuropsychiatric symptom burden from structural MRI.

Abstract: Neuropsychiatric symptoms (NPS) such as depression and apathy are common in Alzheimer’s disease (AD) and often precede cognitive decline. NPS assessments hold promise as early detection markers due to their correlation with disease progression and their non-invasive nature. Yet current tools cannot distinguish whether NPS are part of aging or early signs of AD, limiting their utility. We present a deep learning-based normative modelling framework to identify atypical NPS burden from structural MRI. A 3D convolutional neural network was trained on cognitively stable participants from the Alzheimer’s Disease Neuroimaging Initiative, learning the mapping between brain anatomy and Neuropsychiatric Inventory Questionnaire (NPIQ) scores. Deviations between predicted and observed scores defined the Divergence from NPIQ scores (DNPI). Higher DNPI was associated with future AD conversion (adjusted OR=2.5; p < 0.01) and achieved predictive accuracy comparable to cerebrospinal fluid AB42 (AUC=0.74 vs 0.75). Our approach supports scalable, non-invasive strategies for early AD detection.

Method: Proposes Reliev3R, a weakly-supervised paradigm that uses zero-shot predictions from pretrained models: monocular relative depths and image sparse correspondences. Introduces ambiguity-aware relative depth loss and trigonometry-based reprojection loss to enforce multi-view geometric consistency without explicit 3D annotations.

Result: Reliev3R achieves comparable performance to fully-supervised sibling models when trained from scratch with less data, demonstrating effectiveness of the weakly-supervised approach for 3D reconstruction.

Conclusion: Reliev3R represents a step toward low-cost 3D reconstruction supervision and scalable Feed-forward Reconstruction Models by reducing dependency on expensive geometric annotations while maintaining reconstruction quality.

Abstract: With recent advances, Feed-forward Reconstruction Models (FFRMs) have demonstrated great potential in reconstruction quality and adaptiveness to multiple downstream tasks. However, the excessive reliance on multi-view geometric annotations, e.g. 3D point maps and camera poses, makes the fully-supervised training scheme of FFRMs difficult to scale up. In this paper, we propose Reliev3R, a weakly-supervised paradigm for training FFRMs from scratch without cost-prohibitive multi-view geometric annotations. Relieving the reliance on geometric sensory data and compute-exhaustive structure-from-motion preprocessing, our method draws 3D knowledge directly from monocular relative depths and image sparse correspondences given by zero-shot predictions of pretrained models. At the core of Reliev3R, we design an ambiguity-aware relative depth loss and a trigonometry-based reprojection loss to facilitate supervision for multi-view geometric consistency. Training from scratch with the less data, Reliev3R catches up with its fully-supervised sibling models, taking a step towards low-cost 3D reconstruction supervisions and scalable FFRMs.

Method: TF-SSD combines SAM and DINO: 1) SAM generates raw proposals as candidate masks, 2) quality mask generator filters redundant masks, 3) intra-image saliency filter uses DINO’s attention maps to identify salient masks within individual images, 4) inter-image prototype selector computes similarity scores across images to select the most consistent masks as final predictions.

Result: Extensive experiments show TF-SSD outperforms existing methods, achieving 13.7% gains over recent training-free methods. The method demonstrates strong generalization without requiring training.

Conclusion: TF-SSD successfully leverages Vision Foundation Models (SAM and DINO) for Co-salient Object Detection, providing a training-free solution with strong generalization capabilities that outperforms existing approaches.

Abstract: Co-salient Object Detection (CoSOD) aims to segment salient objects that consistently appear across a group of related images. Despite the notable progress achieved by recent training-based approaches, they still remain constrained by the closed-set datasets and exhibit limited generalization. However, few studies explore the potential of Vision Foundation Models (VFMs) to address CoSOD, which demonstrate a strong generalized ability and robust saliency understanding. In this paper, we investigate and leverage VFMs for CoSOD, and further propose a novel training-free method, TF-SSD, through the synergy between SAM and DINO. Specifically, we first utilize SAM to generate comprehensive raw proposals, which serve as a candidate mask pool. Then, we introduce a quality mask generator to filter out redundant masks, thereby acquiring a refined mask set. Since this generator is built upon SAM, it inherently lacks semantic understanding of saliency. To this end, we adopt an intra-image saliency filter that employs DINO’s attention maps to identify visually salient masks within individual images. Moreover, to extend saliency understanding across group images, we propose an inter-image prototype selector, which computes similarity scores among cross-image prototypes to select masks with the highest score. These selected masks serve as final predictions for CoSOD. Extensive experiments show that our TF-SSD outperforms existing methods (e.g., 13.7% gains over the recent training-free method). Codes are available at https://github.com/hzz-yy/TF-SSD.

Method: Two-stage framework: 1) Supervised fine-tuning to internalize spatial semantics and prune redundant paths, 2) Spatial-aware Segment-level Direct Preference Optimization (SDPO) for self-correction in long-horizon navigation. Uses topological anchors and the RedMaze-23K dataset with human-inspired turnpoint annotations.

Result: STAR achieves state-of-the-art performance among open-source models: 32B variant outperforms DeepSeek-V3 (29.27% vs. 25.00%) and reaches 82.4% of GPT-4’s performance.

Conclusion: STAR demonstrates effective structured spatial navigation through topological anchors and spatial-aware preference optimization, significantly improving LLM performance on complex spatial reasoning tasks.

Abstract: Structured spatial navigation is a core benchmark for Large Language Models (LLMs) spatial reasoning. Existing paradigms like Visualization-of-Thought (VoT) are prone to cascading errors in complex topologies. To solve this, we propose STAR, a two-stage framework grounded on topological anchors, and introduce the RedMaze-23K dataset with human-inspired turnpoint annotations. The first stage uses supervised fine-tuning to help models internalize spatial semantics and prune redundant paths. The second adopts Spatial-aware Segment-level Direct Preference Optimization (SDPO) to refine self-correction in long-horizon navigation. Experiments show STAR achieves state-of-the-art performance among open-source models: its 32B variant outperforms DeepSeek-V3 (29.27% vs. 25.00%) and reaches 82.4% of GPT-4’s performance.

Method: Introduces FecalFed framework with poultry-fecal-fl dataset (8,770 unique images, 46.89% deduplication), evaluates under non-IID conditions (Dirichlet α=0.5), uses federated learning with server-side adaptive optimization (FedAdam) and Swin-Small/Tiny architectures.

Result: Federated approach achieves 90.31% accuracy (Swin-Small) vs. 64.86% for isolated single-farm training, approaching centralized upper bound of 95.10%. Edge-optimized Swin-Tiny maintains 89.74% accuracy.

Conclusion: FecalFed establishes an efficient, privacy-first blueprint for on-farm avian disease monitoring that addresses data privacy concerns while maintaining competitive performance.

Abstract: Early detection of highly pathogenic avian influenza (HPAI) and endemic poultry diseases is critical for global food security. While computer vision models excel at classifying diseases from fecal imaging, deploying these systems at scale is bottlenecked by farm data privacy concerns and institutional data silos. Furthermore, existing open-source agricultural datasets frequently suffer from severe, undocumented data contamination. In this paper, we introduce $\textbf{FecalFed}$, a privacy-preserving federated learning framework for poultry disease classification. We first curate and release $\texttt{poultry-fecal-fl}$, a rigorously deduplicated dataset of 8,770 unique images across four disease classes, revealing and eliminating a 46.89$%$ duplication rate in popular public repositories. To simulate realistic agricultural environments, we evaluate FecalFed under highly heterogeneous, non-IID conditions (Dirichlet $α=0.5$). While isolated single-farm training collapses under this data heterogeneity, yielding only 64.86$%$ accuracy, our federated approach recovers performance without centralizing sensitive data. Specifically, utilizing server-side adaptive optimization (FedAdam) with a Swin-Small architecture achieves 90.31$%$ accuracy, closely approaching the centralized upper bound of 95.10%. Furthermore, we demonstrate that an edge-optimized Swin-Tiny model maintains highly competitive performance at 89.74$%$, establishing a highly efficient, privacy-first blueprint for on-farm avian disease monitoring.

Method: Developed a vision-language model (VLM) based system using only visual input. Constructed an IRB-approved harassment vision dataset, applied prompt engineering, and fine-tuned VLMs to detect harassment behavior by considering contextual information in social VR.

Result: Achieved competitive performance: 88.09% accuracy in binary classification and 68.85% in multi-class classification. Matched state-of-the-art baselines (LSTM/CNN, Transformer) while using significantly fewer fine-tuning samples (200 vs. 1,115).

Conclusion: HarassGuard demonstrates that VLMs can effectively detect physical harassment in social VR with visual-only input, offering advantages in contextual reasoning and privacy preservation compared to traditional methods.

Abstract: Social Virtual Reality (VR) platforms provide immersive social experiences but also expose users to serious risks of online harassment. Existing safety measures are largely reactive, while proactive solutions that detect harassment behavior during an incident often depend on sensitive biometric data, raising privacy concerns. In this paper, we present HarassGuard, a vision-language model (VLM) based system that detects physical harassment in social VR using only visual input. We construct an IRB-approved harassment vision dataset, apply prompt engineering, and fine-tune VLMs to detect harassment behavior by considering contextual information in social VR. Experimental results demonstrate that HarassGuard achieves competitive performance compared to state-of-the-art baselines (i.e., LSTM/CNN, Transformer), reaching an accuracy of up to 88.09% in binary classification and 68.85% in multi-class classification. Notably, HarassGuard matches these baselines while using significantly fewer fine-tuning samples (200 vs. 1,115), offering unique advantages in contextual reasoning and privacy-preserving detection.

Method: Leverages geometric priors from a 3D foundation model by injecting per-pixel 3D positions derived from depth estimates as positional embeddings and fusing intermediate geometric features through cross-attention mechanisms.

Result: Experiments on VR-Drive camera viewpoint perturbation benchmark show reduced performance degradation under most perturbation conditions, with clear improvements under pitch and height perturbations, though gains under longitudinal translation are smaller.

Conclusion: The approach demonstrates improved robustness to camera viewpoint changes but suggests more viewpoint-agnostic integration is needed for complete robustness, especially for longitudinal translation perturbations.

Abstract: Robust trajectory planning under camera viewpoint changes is important for scalable end-to-end autonomous driving. However, existing models often depend heavily on the camera viewpoints seen during training. We investigate an augmentation-free approach that leverages geometric priors from a 3D foundation model. The method injects per-pixel 3D positions derived from depth estimates as positional embeddings and fuses intermediate geometric features through cross-attention. Experiments on the VR-Drive camera viewpoint perturbation benchmark show reduced performance degradation under most perturbation conditions, with clear improvements under pitch and height perturbations. Gains under longitudinal translation are smaller, suggesting that more viewpoint-agnostic integration is needed for robustness to camera viewpoint changes.

Method: Proposes KG-CMI framework with four modules: fine-grained cross-modal feature alignment (FCFA), knowledge graph embedding (KGE), cross-modal interaction representation (CMIR), and free-form answer enhanced multi-task learning (FAMT). Integrates medical knowledge graphs with Mamba architecture for cross-modal interaction.

Result: Outperforms state-of-the-art methods on three Med-VQA datasets (VQA-RAD, SLAKE, OVQA). Interpretability experiments validate framework effectiveness.

Conclusion: KG-CMI effectively integrates medical knowledge through graphs, establishes lesion-disease associations, and improves open-ended Med-VQA through multi-task learning with free-form answers.

Abstract: Medical visual question answering (Med-VQA) is a crucial multimodal task in clinical decision support and telemedicine. Recent methods fail to fully leverage domain-specific medical knowledge, making it difficult to accurately associate lesion features in medical images with key diagnostic criteria. Additionally, classification-based approaches typically rely on predefined answer sets. Treating Med-VQA as a simple classification problem limits its ability to adapt to the diversity of free-form answers and may overlook detailed semantic information in those answers. To address these challenges, we propose a knowledge graph enhanced cross-Mamba interaction (KG-CMI) framework, which consists of a fine-grained cross-modal feature alignment (FCFA) module, a knowledge graph embedding (KGE) module, a cross-modal interaction representation (CMIR) module, and a free-form answer enhanced multi-task learning (FAMT) module. The KG-CMI learns cross-modal feature representations for images and texts by effectively integrating professional medical knowledge through a graph, establishing associations between lesion features and disease knowledge. Moreover, FAMT leverages auxiliary knowledge from open-ended questions, improving the model’s capability for open-ended Med-VQA. Experimental results demonstrate that KG-CMI outperforms existing state-of-the-art methods on three Med-VQA datasets, i.e., VQA-RAD, SLAKE, and OVQA. Additionally, we conduct interpretability experiments to further validate the framework’s effectiveness.

Method: Tested four SNN architectures under two threat models (l-infinity and l-2 PGD attacks), focusing on EMS-YOLO’s behavior where detection count remained high while mAP collapsed dramatically.

Result: EMS-YOLO retained >70% detections while mAP collapsed from 0.528 to 0.042 under PGD attacks. This Quality Corruption phenomenon was unique to EMS-YOLO among four tested architectures, and standard defenses failed to detect or mitigate it.

Conclusion: Adversarial failure modes can be substrate-dependent, challenging assumptions in defense mechanisms. The Quality Corruption phenomenon reveals vulnerabilities in current defense approaches that may be calibrated on limited architectures.

Abstract: The primary tools used to monitor and defend object detectors under adversarial attack assume that when accuracy degrades, detection count drops in tandem. This coupling was assumed, not measured. We report a counterexample observed on a single model: under standard PGD, EMS-YOLO, a spiking neural network (SNN) object detector, retains more than 70% of its detections while mAP collapses from 0.528 to 0.042. We term this count-preserving accuracy collapse Quality Corruption (QC), to distinguish it from the suppression that dominates untargeted evaluation. Across four SNN architectures and two threat models (l-infinity and l-2), QC appears only in one of the four detectors tested (EMS-YOLO). On this model, all five standard defense components fail to detect or mitigate QC, suggesting the defense ecosystem may rely on a shared assumption calibrated on a single substrate. These results provide, to our knowledge, the first evidence that adversarial failure modes can be substrate-dependent.

Method: Proposes TALENT framework with Rectified Cost Aggregator (RCA) for efficient text-referred feature aggregation, and Target-aware Learning Mechanism (TLM) including contextual pairwise consistency learning and target-centric contrastive learning.

Result: Outperforms existing methods across various metrics, achieving 2.5% mIoU gains on G-Ref validation set.

Conclusion: TALENT effectively addresses the non-target activation issue in parameter-efficient tuning for referring image segmentation through target-aware learning mechanisms.

Abstract: Referring image segmentation aims to segment specific targets based on a natural text expression. Recently, parameter-efficient tuning (PET) has emerged as a promising paradigm. However, existing PET-based methods often suffer from the fact that visual features can’t emphasize the text-referred target instance but activate co-category yet unrelated objects. We analyze and quantify this problem, terming it the non-target activation' (NTA) issue. To address this, we propose a novel framework, TALENT, which utilizes target-aware efficient tuning for PET-based RIS. Specifically, we first propose a Rectified Cost Aggregator (RCA) to efficiently aggregate text-referred features. Then, to calibrate NTA’ into accurate target activation, we adopt a Target-aware Learning Mechanism (TLM), including contextual pairwise consistency learning and target-centric contrastive learning. The former uses the sentence-level text feature to achieve a holistic understanding of the referent and constructs a text-referred affinity map to optimize the semantic association of visual features. The latter further enhances target localization to discover the distinct instance while suppressing associations with other unrelated ones. The two objectives work in concert and address `NTA’ effectively. Extensive evaluations show that TALENT outperforms existing methods across various metrics (e.g., 2.5% mIoU gains on G-Ref val set). Our codes will be released at: https://github.com/Kimsure/TALENT.

Method: Integrates fisheye camera model into 3DGS framework for native fisheye training, plus introduces feature-overlap-driven cross-view joint optimization to handle distortion at edges and prevent extreme Gaussian shapes.

Result: DirectFisheye-GS matches or surpasses state-of-the-art performance on public datasets, effectively handling fisheye distortion without preprocessing artifacts.

Conclusion: Native fisheye integration with cross-view optimization enables high-quality 3D reconstruction from fisheye images while preserving their wide FOV advantage.

Abstract: 3D Gaussian Splatting (3DGS) has enabled efficient 3D scene reconstruction from everyday images with real-time, high-fidelity rendering, greatly advancing VR/AR applications. Fisheye cameras, with their wider field of view (FOV), promise high-quality reconstructions from fewer inputs and have recently attracted much attention. However, since 3DGS relies on rasterization, most subsequent works involving fisheye camera inputs first undistort images before training, which introduces two problems: 1) Black borders at image edges cause information loss and negate the fisheye’s large FOV advantage; 2) Undistortion’s stretch-and-interpolate resampling spreads each pixel’s value over a larger area, diluting detail density – causes 3DGS overfitting these low-frequency zones, producing blur and floating artifacts. In this work, we integrate fisheye camera model into the original 3DGS framework, enabling native fisheye image input for training without preprocessing. Despite correct modeling, we observed that the reconstructed scenes still exhibit floaters at image edges: Distortion increases toward the periphery, and 3DGS’s original per-iteration random-selecting-view optimization ignores the cross-view correlations of a Gaussian, leading to extreme shapes (e.g., oversized or elongated) that degrade reconstruction quality. To address this, we introduce a feature-overlap-driven cross-view joint optimization strategy that establishes consistent geometric and photometric constraints across views-a technique equally applicable to existing pinhole-camera-based pipelines. Our DirectFisheye-GS matches or surpasses state-of-the-art performance on public datasets.

Method: Used multiple CNN architectures to identify systematically misclassified dermatoscopic images, then had expert dermatologists evaluate these challenging cases alongside control images to compare human and AI performance on the same difficult samples.

Result: Both AI and human experts performed poorly on the same challenging images: human agreement with ground truth dropped dramatically (kappa 0.08 vs 0.61), and expert consensus deteriorated (Fleiss kappa 0.275 vs 0.456), with image quality identified as the primary cause.

Conclusion: AI failures on dermatoscopic images often reflect inherent visual ambiguity rather than algorithmic deficiencies, as human experts also struggle with the same challenging cases, highlighting the importance of image quality in diagnostic reliability.

Abstract: The integration of artificial intelligence (AI), particularly Convolutional Neural Networks (CNNs), into dermatological diagnosis demonstrates substantial clinical potential. While existing literature predominantly benchmarks algorithmic performance against human experts, our study adopts a novel perspective by investigating the intrinsic complexity of dermatoscopic images. Through rigorous experimentation with multiple CNN architectures, we isolated a subset of images systematically misclassified across all models-a phenomenon statistically proven to exceed random chance. To determine if these failures stem from algorithmic biases or inherent visual ambiguity, expert dermatologists independently evaluated these challenging cases alongside a control group. The results revealed a collapse in human diagnostic performance on the AI-misclassified images. First, agreement with ground-truth labels plummeted, with Cohen’s kappa dropping to a mere 0.08 for the difficult images, compared to a 0.61 for the control group. Second, we observed a severe deterioration in expert consensus; inter-rater reliability among physicians fell from moderate concordance (Fleiss kappa = 0.456) on control images to only modest agreement (Fleiss kappa = 0.275) on difficult cases. We identified image quality as a primary driver of these dual systematic failures. To promote transparency and reproducibility, all data, code, and trained models have been made publicly available

Method: Created CL-VISTA benchmark with 8 diverse tasks spanning perception, understanding, and reasoning domains. Established comprehensive evaluation framework with 6 protocols across 3 dimensions: performance, computational efficiency, and memory footprint. Benchmarked 10 mainstream continual learning methods.

Result: Benchmarking revealed a fundamental trade-off: no single continual learning approach achieves universal superiority across all dimensions. Methods that successfully mitigate catastrophic forgetting tend to compromise generalization or incur prohibitive computational and memory overheads.

Conclusion: CL-VISTA provides critical insights for advancing continual learning in multimodal foundation models by exposing the inherent trade-offs between forgetting mitigation, generalization, and resource efficiency in Video-LLMs.

Abstract: Video Large Language Models (Video-LLMs) require continual learning to adapt to non-stationary real-world data. However, existing benchmarks fall short of evaluating modern foundation models: many still rely on models without large-scale pre-training, and prevailing benchmarks typically partition a single dataset into sub-tasks, resulting in high task redundancy and negligible forgetting on pre-trained Video-LLMs. To address these limitations, we propose CL-VISTA, a benchmark tailored for continual video understanding of Video-LLMs. By curating 8 diverse tasks spanning perception, understanding, and reasoning, CL-VISTA induces substantial distribution shifts that effectively expose catastrophic forgetting. To systematically assess CL methods, we establish a comprehensive evaluation framework comprising 6 distinct protocols across 3 critical dimensions: performance, computational efficiency, and memory footprint. Notably, the performance dimension incorporates a general video understanding assessment to assess whether CL methods genuinely enhance foundational intelligence or merely induce task-specific overfitting. Extensive benchmarking of 10 mainstream CL methods reveals a fundamental trade-off: no single approach achieves universal superiority across all dimensions. Methods that successfully mitigate catastrophic forgetting tend to compromise generalization or incur prohibitive computational and memory overheads. We hope CL-VISTA provides critical insights for advancing continual learning in multimodal foundation models.

Method: Created a unified benchmark with two complementary sub-datasets: 1) LunarGeo provides stereo images with dense depth maps and camera calibration for 3D reconstruction and pose estimation; 2) LunarPhoto provides photorealistic images using spatially-varying BRDF model with multi-illumination renderings under real solar configurations for reflectance estimation and illumination-robust perception.

Result: The benchmark comprises over 130K samples with comprehensive supervision. It offers a unique setting for algorithms under low-textured, high-contrast conditions and applies to other airless celestial bodies. Baselines were established using state-of-the-art methods.

Conclusion: MoonAnything provides the first comprehensive geometric and photometric supervision for lunar perception systems under diverse illumination at large scale. It serves as a challenging testbed for algorithms and could generalize beyond lunar applications to other airless celestial bodies.

Abstract: Accurate perception of lunar surfaces is critical for modern lunar exploration missions. However, developing robust learning-based perception systems is hindered by the lack of datasets that provide both geometric and photometric supervision. Existing lunar datasets typically lack either geometric ground truth, photometric realism, illumination diversity, or large-scale coverage. In this paper, we introduce MoonAnything, a unified benchmark built on real lunar topography with physically-based rendering, providing the first comprehensive geometric and photometric supervision under diverse illumination with large scale. The benchmark comprises two complementary sub-datasets : i) LunarGeo provides stereo images with corresponding dense depth maps and camera calibration enabling 3D reconstruction and pose estimation; ii) LunarPhoto provides photorealistic images using a spatially-varying BRDF model, along with multi-illumination renderings under real solar configurations, enabling reflectance estimation and illumination-robust perception. Together, these datasets offer over 130K samples with comprehensive supervision. Beyond lunar applications, MoonAnything offers a unique setting and challenging testbed for algorithms under low-textured, high-contrast conditions and applies to other airless celestial bodies and could generalize beyond. We establish baselines using state-of-the-art methods and release the complete dataset along with generation tools to support community extension: https://github.com/clementinegrethen/MoonAnything.

Method: TP-Seg uses: 1) Task-conditioned adapter with dual-path expert structure to balance shared and task-specific representations, enabling adaptive feature extraction across modalities; 2) Prototype-guided task decoder with learnable task prototypes as semantic anchors and cross-attention mechanism for fine-grained modeling of task-specific foreground/background semantics.

Result: TP-Seg consistently outperforms specialized, general, and unified segmentation methods across 8 different medical lesion segmentation tasks covering multiple imaging modalities, demonstrating strong generalization, scalability, and clinical applicability.

Conclusion: TP-Seg provides an effective unified framework for medical lesion segmentation that addresses feature entanglement issues in existing approaches and shows superior performance across diverse medical imaging tasks and modalities.

Abstract: Building a unified model with a single set of parameters to efficiently handle diverse types of medical lesion segmentation has become a crucial objective for AI-assisted diagnosis. Existing unified segmentation approaches typically rely on shared encoders across heterogeneous tasks and modalities, which often leads to feature entanglement, gradient interference, and suboptimal lesion discrimination. In this work, we propose TP-Seg, a task-prototype framework for unified medical lesion segmentation. On one hand, the task-conditioned adapter effectively balances shared and task-specific representations through a dual-path expert structure, enabling adaptive feature extraction across diverse medical imaging modalities and lesion types. On the other hand, the prototype-guided task decoder introduces learnable task prototypes as semantic anchors and employs a cross-attention mechanism to achieve fine-grained modeling of task-specific foreground and background semantics. Without bells and whistles, TP-Seg consistently outperforms specialized, general and unified segmentation methods across 8 different medical lesion segmentation tasks covering multiple imaging modalities, demonstrating strong generalization, scalability and clinical applicability.

Method: Combines test-time adaptation with step-by-step reasoning on multiple frame subsets, using batch-aware frequency-based rewards as pseudo ground truth to update the model. Includes multi-armed bandit strategy for adaptive frame selection that prioritizes informative frames guided by the same reward formulation.

Result: TTA-Vid yields consistent improvements across various video reasoning tasks and outperforms current state-of-the-art methods trained on large-scale data, demonstrating generalization from single batch/sample to entire datasets and across datasets.

Conclusion: The approach highlights the potential of test-time reinforcement learning for temporal multimodal understanding, requiring no ground-truth annotations or dedicated training splits while enabling efficient adaptation at inference time.

Abstract: Recent video reasoning models have shown strong results on temporal and multimodal understanding, yet they depend on large-scale supervised data and multi-stage training pipelines, making them costly to train and difficult to adapt to new domains. In this work, we leverage the paradigm of Test-Time Reinforcement Learning on video-language data to allow for adapting a pretrained model to incoming video samples at test-time without explicit labels. The proposed test-time adaptation for video approach (TTA-Vid) combines two components that work simultaneously: (1) a test-time adaptation that performs step-by-step reasoning at inference time on multiple frame subsets. We then use a batch-aware frequency-based reward computed across different frame subsets as pseudo ground truth to update the model. It shows that the resulting model trained on a single batch or even a single sample from a dataset, is able to generalize at test-time to the whole dataset and even across datasets. Because the adaptation occurs entirely at test time, our method requires no ground-truth annotations or dedicated training splits. Additionally, we propose a multi-armed bandit strategy for adaptive frame selection that learns to prioritize informative frames, guided by the same reward formulation. Our evaluation shows that TTA-Vid yields consistent improvements across various video reasoning tasks and is able to outperform current state-of-the-art methods trained on large-scale data. This highlights the potential of test-time reinforcement learning for temporal multimodal understanding.

Method: Proposes the first SSM-based OCR architecture combining CNN visual encoder with bi-directional and autoregressive Mamba sequence modeling. Compares SSMs with Transformer- and BiLSTM-based recognizers using multiple decoding strategies (CTC, autoregressive, non-autoregressive) under identical training conditions. Evaluates against strong neural baselines (VAN, DAN, DANIEL) and off-the-shelf OCR engines (PERO-OCR, Tesseract OCR, TrOCR, Gemini).

Result: All neural models achieve low error rates (~2% CER) on historical newspapers. Mamba-based models maintain competitive accuracy while halving inference time and showing superior memory scaling (1.26x vs 2.30x growth at 1000 chars). At severely degraded paragraph level, Mamba reaches 6.07% CER vs 5.24% for DAN while remaining 2.05x faster.

Conclusion: SSMs (specifically Mamba) provide a scalable alternative to Transformers for OCR, offering competitive accuracy with significantly better computational efficiency and memory scaling, making them suitable for large-scale cultural heritage OCR deployment.

Abstract: End-to-end OCR for historical newspapers remains challenging, as models must handle long text sequences, degraded print quality, and complex layouts. While Transformer-based recognizers dominate current research, their quadratic complexity limits efficient paragraph-level transcription and large-scale deployment. We investigate linear-time State-Space Models (SSMs), specifically Mamba, as a scalable alternative to Transformer-based sequence modeling for OCR. We present to our knowledge, the first OCR architecture based on SSMs, combining a CNN visual encoder with bi-directional and autoregressive Mamba sequence modeling, and conduct a large-scale benchmark comparing SSMs with Transformer- and BiLSTM-based recognizers. Multiple decoding strategies (CTC, autoregressive, and non-autoregressive) are evaluated under identical training conditions alongside strong neural baselines (VAN, DAN, DANIEL) and widely used off-the-shelf OCR engines (PERO-OCR, Tesseract OCR, TrOCR, Gemini). Experiments on historical newspapers from the Bibliothèque nationale du Luxembourg, with newly released >99% verified gold-standard annotations, and cross-dataset tests on Fraktur and Antiqua lines, show that all neural models achieve low error rates (~2% CER), making computational efficiency the main differentiator. Mamba-based models maintain competitive accuracy while halving inference time and exhibiting superior memory scaling (1.26x vs 2.30x growth at 1000 chars), reaching 6.07% CER at the severely degraded paragraph level compared to 5.24% for DAN, while remaining 2.05x faster. We release code, trained models, and standardized evaluation protocols to enable reproducible research and guide practitioners in large-scale cultural heritage OCR.

Method: Reformulates attention as implicit linear layer with rank-1 outer products, derives metric for token information magnitude and duplication, uses Progressive Chunked Maximal Marginal Relevance for efficient subset selection

Result: Achieves better trade-off between performance and efficiency compared to existing methods, provides new perspective on pruning approaches

Conclusion: The proposed training-free token pruning framework effectively reduces computational cost while maintaining performance by leveraging attention mechanism insights

Abstract: Large Vision Language Models show impressive performance across image and video understanding tasks, yet their computational cost grows rapidly with the number of visual tokens. Existing token pruning methods mitigate this issue through empirical approaches while overlooking the internal mechanism of attention. In this paper, we propose a novel training free token pruning framework grounded in the dual form perspective of attention. We reformulate attention as an implicit linear layer whose weight matrix is the sum of rank 1 outer products, each generated by a single token’s key value pair. Token pruning thus reduces to selecting an optimal subset of these rank 1 updates that best approximates the original dual weight matrix. Extending this perspective to standard softmax attention in LVLMs, we derive a novel metric quantifying both a token’s information magnitude and information duplication. To efficiently select the subset with the proposed metric, we introduce Progressive Chunked Maximal Marginal Relevance. Extensive experiments demonstrate that our method achieves a better trade off between performance and efficiency, while providing another perspective on existing pruning approaches.

Method: Created a benchmark dataset with 1,932 source videos across 15 activity classes, applying 9 parallel tiers of increasing privacy strength (from lightweight spatial obfuscation to cryptographic block permutation), including background-removed variants to isolate human motion features.

Result: Empirical validation using R3D-18 shows measurable degradation curve: within-tier accuracy declines from 88.8% (clear) to 53.5% (encrypted, background-removed), and cross-domain accuracy collapses to 4.8%, demonstrating controlled privacy-utility trade-off evaluation.

Conclusion: PrivHAR-Bench provides a standardized benchmark for comparing privacy-preserving HAR methods under controlled conditions, enabling systematic evaluation of privacy-utility trade-offs in video-based action recognition.

Abstract: Existing research on privacy-preserving Human Activity Recognition (HAR) typically evaluates methods against a binary paradigm: clear video versus a single privacy transformation. This limits cross-method comparability and obscures the nuanced relationship between privacy strength and recognition utility. We introduce \textit{PrivHAR-Bench}, a multi-tier benchmark dataset designed to standardize the evaluation of the \textit{Privacy-Utility Trade-off} in video-based action recognition. PrivHAR-Bench applies a graduated spectrum of visual privacy transformations: from lightweight spatial obfuscation to cryptographic block permutation, to a curated subset of 15 activity classes selected for human articulation diversity. Each of the 1,932 source videos is distributed across 9 parallel tiers of increasing privacy strength, with additional background-removed variants to isolate the contribution of human motion features from contextual scene bias. We provide lossless frame sequences, per-frame bounding boxes, estimated pose keypoints with joint-level confidence scores, standardized group-based train/test splits, and an evaluation toolkit computing recognition accuracy and privacy metrics. Empirical validation using R3D-18 demonstrates a measurable and interpretable degradation curve across tiers, with within-tier accuracy declining from 88.8% (clear) to 53.5% (encrypted, background-removed) and cross-domain accuracy collapsing to 4.8%, establishing PrivHAR-Bench as a controlled benchmark for comparing privacy-preserving HAR methods under standardized conditions. The dataset, generation pipeline, and evaluation code are publicly available.

Method: Developed SurgSTU-Pipeline with temporal and spatial continuity filtering to reliably generate surgical datasets. Applied to public surgical datasets to create SurgSTU dataset with 7515 video clips and 150k fine-grained spatial-temporal QA samples.

Result: State-of-the-art generalist VLMs struggle in zero-shot settings but improve with in-context learning. Fine-tuned VLM on SurgSTU training dataset achieves highest performance on spatial-temporal tasks.

Conclusion: SurgSTU dataset effectively improves spatial-temporal understanding of VLMs in surgical videos, addressing a critical gap in surgical video understanding for computer-assisted surgery.

Abstract: Surgical video understanding is a crucial prerequisite for advancing Computer-Assisted Surgery. While vision-language models (VLMs) have recently been applied to the surgical domain, existing surgical vision-language datasets lack in capturing and evaluating complex, interleaved spatial-temporal dynamics. Creating large scale datasets that accurately represent fine-grained spatial-temporal relationships in surgical videos is challenging due to costly manual annotations or error-prone generation using large language models. To address this gap, we introduce the SurgSTU-Pipeline, a deterministic generation pipeline featuring temporal and spatial continuity filtering to reliably create surgical datasets for fine-grained spatial-temporal multimodal understanding. Applying this pipeline to publicly available surgical datasets, we create the SurgSTU dataset, comprising 7515 video clips densely extended with 150k fine-grained spatial-temporal question-answer samples. Our comprehensive evaluation shows that while state-of-the-art generalist VLMs struggle in zero-shot settings, their spatial-temporal capabilities can be improved through in-context learning. A fine-tuned VLM on the SurgSTU training dataset achieves highest performance among all spatial-temporal tasks, validating the dataset’s efficacy to improve spatial-temporal understanding of VLMs in surgical videos. Code will be made publicly available.

Method: Two-stage framework: 1) Video diffusion model generates geometrically consistent multi-view projections from single panoramic image, 2) Ray-based dynamic attention network with X-ray sampling strategy reconstructs high-fidelity CBCT from projections. Built XCT dataset with 500 paired PX-CBCT cases.

Result: Extensive experiments show HiCT achieves state-of-the-art performance, delivering accurate and geometrically consistent reconstructions suitable for clinical use.

Conclusion: HiCT provides effective solution for 3D dental reconstruction from single panoramic X-ray, addressing radiation and cost limitations of CBCT while maintaining clinical accuracy.

Abstract: Accurate 3D dental imaging is vital for diagnosis and treatment planning, yet CBCT’s high radiation dose and cost limit its accessibility. Reconstructing 3D volumes from a single low-dose panoramic X-ray is a promising alternative but remains challenging due to geometric inconsistencies and limited accuracy. We propose HiCT, a two-stage framework that first generates geometrically consistent multi-view projections from a single panoramic image using a video diffusion model, and then reconstructs high-fidelity CBCT from the projections using a ray-based dynamic attention network and an X-ray sampling strategy. To support this, we built XCT, a large-scale dataset combining public CBCT data with 500 paired PX-CBCT cases. Extensive experiments show that HiCT achieves state-of-the-art performance, delivering accurate and geometrically consistent reconstructions for clinical use.

Method: Introduce a task where models must identify objects violating 3D motion consistency across two views of the same scene. Develop a scalable method for generating realistic, spatially inconsistent image pairs from multi-view scenes for systematic evaluation.

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

Conclusion: Current MLLMs lack robust 3D spatial reasoning capabilities, highlighting the need for approaches that develop more deeply grounded understanding of the physical world.

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

Method: Formulates retrieval as binary classification using pre-trained ViT representations, explores key design questions including object instance selection, annotation forms, active selection strategies, and representation approaches that balance global context with local object details.

Result: Compares several representation strategies across multi-object datasets, highlighting trade-offs between capturing global context and focusing on fine-grained local object details, providing practical insights for interactive retrieval pipelines.

Conclusion: Offers design guidelines for effective interactive retrieval systems based on active learning for object class retrieval, particularly in challenging multi-object scenarios where localized descriptors are essential.

Abstract: Building on existing approaches, we revisit Human-in-the-Loop Object Retrieval, a task that consists of iteratively retrieving images containing objects of a class-of-interest, specified by a user-provided query. Starting from a large unlabeled image collection, the aim is to rapidly identify diverse instances of an object category relying solely on the initial query and the user’s Relevance Feedback, with no prior labels. The retrieval process is formulated as a binary classification task, where the system continuously learns to distinguish between relevant and non-relevant images to the query, through iterative user interaction. This interaction is guided by an Active Learning loop: at each iteration, the system selects informative samples for user annotation, thereby refining the retrieval performance. This task is particularly challenging in multi-object datasets, where the object of interest may occupy only a small region of the image within a complex, cluttered scene. Unlike object-centered settings where global descriptors often suffice, multi-object images require more adapted, localized descriptors. In this work, we formulate and revisit the Human-in-the-Loop Object Retrieval task by leveraging pre-trained ViT representations, and addressing key design questions, including which object instances to consider in an image, what form the annotations should take, how Active Selection should be applied, and which representation strategies best capture the object’s features. We compare several representation strategies across multi-object datasets highlighting trade-offs between capturing the global context and focusing on fine-grained local object details. Our results offer practical insights for the design of effective interactive retrieval pipelines based on Active Learning for object class retrieval.

Method: Proposes DVGT-2, a streaming Driving Visual Geometry Transformer that processes inputs online using temporal causal attention and caches historical features. Implements a sliding-window streaming strategy to avoid repetitive computations while maintaining efficiency.

Result: DVGT-2 achieves superior geometry reconstruction performance on various datasets while being faster than previous methods. The same trained model can be directly applied to planning across diverse camera configurations without fine-tuning, performing well on both closed-loop NAVSIM and open-loop nuScenes benchmarks.

Conclusion: The Vision-Geometry-Action (VGA) paradigm with streaming dense geometry reconstruction provides an effective alternative to vision-language-action models for autonomous driving, offering comprehensive 3D information for decision-making with efficient online inference capabilities.

Abstract: End-to-end autonomous driving has evolved from the conventional paradigm based on sparse perception into vision-language-action (VLA) models, which focus on learning language descriptions as an auxiliary task to facilitate planning. In this paper, we propose an alternative Vision-Geometry-Action (VGA) paradigm that advocates dense 3D geometry as the critical cue for autonomous driving. As vehicles operate in a 3D world, we think dense 3D geometry provides the most comprehensive information for decision-making. However, most existing geometry reconstruction methods (e.g., DVGT) rely on computationally expensive batch processing of multi-frame inputs and cannot be applied to online planning. To address this, we introduce a streaming Driving Visual Geometry Transformer (DVGT-2), which processes inputs in an online manner and jointly outputs dense geometry and trajectory planning for the current frame. We employ temporal causal attention and cache historical features to support on-the-fly inference. To further enhance efficiency, we propose a sliding-window streaming strategy and use historical caches within a certain interval to avoid repetitive computations. Despite the faster speed, DVGT-2 achieves superior geometry reconstruction performance on various datasets. The same trained DVGT-2 can be directly applied to planning across diverse camera configurations without fine-tuning, including closed-loop NAVSIM and open-loop nuScenes benchmarks.

Method: Uses UpAttLLSTM - an attention-based recurrent segmentation network (2.5D) that aggregates context across slices via recurrent units and attention gates to fuse complementary cues across available sequences. Includes progressive training with gradual modality dropout to simulate site heterogeneity, noise, and missing modalities.

Result: On monocentric cohort: >90% detection rate with Dice score of 0.65. On multi-center setting with missing modalities: ~80% detection rate with Dice score around 0.35. Shows transferability to other small-lesion tasks in 3D medical imaging.

Conclusion: The methodology effectively addresses challenges of tiny target detection in 3D medical imaging with robustness to domain shifts and missing modalities, with demonstrated applicability beyond stroke to other small-lesion tasks.

Abstract: Detecting and delineating tiny targets in 3D brain scans is a central yet under-addressed challenge in medical imaging.In ischemic stroke, for instance, the culprit thrombus is small, low-contrast, and variably expressed across modalities(e.g., susceptibility-weighted T2 blooming, diffusion restriction on DWI/ADC), while real-world multi-center dataintroduce domain shifts, anisotropy, and frequent missing sequences. We introduce a methodology that couples an attention-based recurrent segmentation network (UpAttLLSTM), a training schedule that progressively increases the difficulty of hetero-modal learning, with gradual modality dropout, UpAttLLSTM aggregates context across slices via recurrent units (2.5D) and uses attention gates to fuse complementary cues across available sequences, making it robust to anisotropy and class imbalance. Gradual modality dropout systematically simulates site heterogeneity,noise, and missing modalities during training, acting as both augmentation and regularization to improve multi-center generalization. On a monocentric cohort, our approach detects thrombi in >90% of cases with a Dice score of 0.65. In a multi-center setting with missing modalities, it achieves-80% detection with a Dice score around 0.35. Beyond stroke, the proposed methodology directly transfers to other small-lesion tasks in 3D medical imaging where targets are scarce, subtle, and modality-dependent

Method: PixelPrune uses predictive-coding-based compression to prune redundant patches before the Vision Transformer encoder. It operates in pixel space prior to neural computation, supports both pixel-lossless (τ=0) and controlled lossy compression (τ>0), is training-free, and requires no learnable parameters.

Result: Experiments across three model scales and document/GUI benchmarks show PixelPrune maintains competitive task accuracy while delivering up to 4.2× inference speedup and 1.9× training acceleration.

Conclusion: PixelPrune effectively exploits pixel-level redundancy in document/GUI images to significantly accelerate VLM inference and training while preserving accuracy, addressing the computational burden of high-resolution vision-language tasks.

Abstract: Document understanding and GUI interaction are among the highest-value applications of Vision-Language Models (VLMs), yet they impose exceptionally heavy computational burden: fine-grained text and small UI elements demand high-resolution inputs that produce tens of thousands of visual tokens. We observe that this cost is largely wasteful – across document and GUI benchmarks, only 22–71% of image patches are pixel-unique, the rest being exact duplicates of another patch in the same image. We propose \textbf{PixelPrune}, which exploits this pixel-level redundancy through predictive-coding-based compression, pruning redundant patches \emph{before} the Vision Transformer (ViT) encoder. Because it operates in pixel space prior to any neural computation, PixelPrune accelerates both the ViT encoder and the downstream LLM, covering the full inference pipeline. The method is training-free, requires no learnable parameters, and supports pixel-lossless compression ($τ{=}0$) as well as controlled lossy compression ($τ{>}0$). Experiments across three model scales and document and GUI benchmarks show that PixelPrune maintains competitive task accuracy while delivering up to 4.2$\times$ inference speedup and 1.9$\times$ training acceleration. Code is available at https://github.com/OPPO-Mente-Lab/PixelPrune.

Method: Created CLeaRS benchmark with 10 curated subsets (207k+ image-text pairs) spanning diverse interpretation tasks, sensing modalities, and application scenarios. Defined three evaluation protocols: long-horizon, modality-incremental, and task-incremental settings.

Result: Benchmarking revealed catastrophic forgetting across all settings in diverse vision-language models. Existing continual learning methods adapted to RS VLMs showed limited effectiveness in handling task, instruction, and modality transitions.

Conclusion: There is a critical need for developing continual learning methods specifically tailored to RS VLMs to enable adaptation to emerging sensing modalities and tasks without catastrophic forgetting.

Abstract: Current remote sensing vision-language models (RS VLMs) demonstrate impressive performance in image interpretation but rely on static training data, limiting their ability to accommodate continuously emerging sensing modalities and downstream tasks. This exposes a fundamental challenge: enabling RS VLMs to continually adapt without catastrophic forgetting. Despite its practical importance, the continual learning capability of RS VLMs remains underexplored, and no dedicated benchmark currently exists. In this work, we present CLeaRS, a comprehensive benchmark for continual vision-language learning in remote sensing. CLeaRS comprises 10 curated subsets with over 207k image-text pairs, spanning diverse interpretation tasks, sensing modalities, and application scenarios. We further define three evaluation protocols: long-horizon, modality-incremental, and task-incremental settings, to systematically assess continual adaptation. Extensive benchmarking of diverse vision-language models reveals catastrophic forgetting across all settings. Moreover, representative continual learning methods, when adapted to RS VLMs, exhibit limited effectiveness in handling task, instruction, and modality transitions. Our findings underscore the need for developing continual learning methods tailored to RS VLMs.

Method: Proposes a fully differentiable temporal mapping module that leverages prior features from deeper layers (which exhibit strong foreground selectivity) to accurately select the most relevant patches in early network stages for video processing.

Result: Achieves up to 60% patch reduction in dense prediction tasks, exceeding conventional image-based patch pruning (typically ~30% sparsity). Maintains stable performance with only 0.6% drop on Youtube-VIS 2021 dataset even with patch usage reduced below 55%.

Conclusion: VPP enables efficient sparsity in early ViT layers for video understanding, significantly improving computational efficiency while maintaining performance, particularly excelling in high-sparsity regimes.

Abstract: Vision Transformers (ViTs) have demonstrated state-ofthe-art performance in several benchmarks, yet their high computational costs hinders their practical deployment. Patch Pruning offers significant savings, but existing approaches restrict token reduction to deeper layers, leaving early-stage compression unexplored. This limits their potential for holistic efficiency. In this work, we present a novel Video Patch Pruning framework (VPP) that integrates temporal prior knowledge to enable efficient sparsity within early ViT layers. Our approach is motivated by the observation that prior features extracted from deeper layers exhibit strong foreground selectivity. Therefore we propose a fully differentiable module for temporal mapping to accurately select the most relevant patches in early network stages. Notably, the proposed method enables a patch reduction of up to 60% in dense prediction tasks, exceeding the capabilities of conventional image-based patch pruning, which typically operate around a 30% patch sparsity. VPP excels the high-sparsity regime, sustaining remarkable performance even when patch usage is reduced below 55%. Specifically, it preserves stable results with a maximal performance drop of 0.6% on the Youtube-VIS 2021 dataset.

Method: Post-hoc canonical correlation analysis (CCA) operator that leverages shared structure between representations from two pretrained image encoders. Finds linear projections that serve as principled representation selection and dimensionality reduction, retaining shared semantic content while discarding redundant dimensions.

Result: Representations can be reduced by >75% dimensionality with improved downstream performance, or enhanced at fixed dimensionality via post-hoc representation transfer from larger/fine-tuned models. Achieves up to 12.6% accuracy gains on ImageNet-1k, CIFAR-100, MNIST, and other benchmarks over baseline and PCA-projected representations.

Conclusion: CCA-based post-hoc operator provides effective training-free method for improving image representation efficiency by leveraging cross-model agreement, outperforming single-model techniques like PCA and enabling flexible representation refinement.

Abstract: Modern vision pipelines increasingly rely on pretrained image encoders whose representations are reused across tasks and models, yet these representations are often overcomplete and model-specific. We propose a simple, training-free method to improve the efficiency of image representations via a post-hoc canonical correlation analysis (CCA) operator. By leveraging the shared structure between representations produced by two pre-trained image encoders, our method finds linear projections that serve as a principled form of representation selection and dimensionality reduction, retaining shared semantic content while discarding redundant dimensions. Unlike standard dimensionality reduction techniques such as PCA, which operate on a single embedding space, our approach leverages cross-model agreement to guide representation distillation and refinement. The technique allows representations to be reduced by more than 75% in dimensionality with improved downstream performance, or enhanced at fixed dimensionality via post-hoc representation transfer from larger or fine-tuned models. Empirical results on ImageNet-1k, CIFAR-100, MNIST, and additional benchmarks show consistent improvements over both baseline and PCA-projected representations, with accuracy gains of up to 12.6%.

Method: Proposes LinguDistill with layer-wise KV-cache sharing to expose the frozen language model teacher to the student’s multimodal representations, enabling selective distillation on language-intensive data while preserving visual grounding.

Result: Recovers ~10% of performance lost on language and knowledge benchmarks while maintaining comparable performance on vision-heavy tasks, without adding parameters or architectural complexity.

Conclusion: Linguistic capability can be recovered in multimodal models without additional modules, providing an efficient solution to modality-specific degradation through adapter-free distillation.

Abstract: Adapting pretrained language models (LMs) into vision-language models (VLMs) can degrade their native linguistic capability due to representation shift and cross-modal interference introduced during multimodal adaptation. Such loss is difficult to recover, even with targeted task-specific fine-tuning using standard objectives. Prior recovery approaches typically introduce additional modules that act as intermediate alignment layers to maintain or isolate modality-specific subspaces, which increases architectural complexity, adds parameters at inference time, and limits flexibility across models and settings. We propose LinguDistill, an adapter-free distillation method that restores linguistic capability by utilizing the original frozen LM as a teacher. We overcome the key challenge of enabling vision-conditioned teacher supervision by introducing layer-wise KV-cache sharing, which exposes the teacher to the student’s multimodal representations without modifying the architecture of either model. We then selectively distill the teacher’s strong linguistic signal on language-intensive data to recover language capability, while preserving the student’s visual grounding on multimodal tasks. As a result, LinguDistill recovers $\sim$10% of the performance lost on language and knowledge benchmarks, while maintaining comparable performance on vision-heavy tasks. Our findings demonstrate that linguistic capability can be recovered without additional modules, providing an efficient and practical solution to modality-specific degradation in multimodal models.

Method: Combines Skeleton Motion Quantisation (SMQ) with Spatio-Temporal Transformers (STT) to encode human poses into structured motion vocabulary. Uses DANCE RETRIEVAL ENGINE (DRE) with histogram-based indexing for sub-linear retrieval and re-ranking for refined matching.

Result: Demonstrates robust retrieval across diverse dance styles and strong generalization to unseen choreographies. Releases DANCETYPESBENCHMARK dataset with pose-aligned annotations and quantised motion tokens for reproducible research.

Conclusion: Establishes a foundation for scalable motion fingerprinting and quantitative choreographic analysis through discrete motion signatures that enable efficient large-scale dance retrieval.

Abstract: We present DANCEMATCH, an end-to-end framework for motion-based dance retrieval, the task of identifying semantically similar choreographies directly from raw video, defined as DANCE FINGERPRINTING. While existing motion analysis and retrieval methods can compare pose sequences, they rely on continuous embeddings that are difficult to index, interpret, or scale. In contrast, DANCEMATCH constructs compact, discrete motion signatures that capture the spatio-temporal structure of dance while enabling efficient large-scale retrieval. Our system integrates Skeleton Motion Quantisation (SMQ) with Spatio-Temporal Transformers (STT) to encode human poses, extracted via Apple CoMotion, into a structured motion vocabulary. We further design DANCE RETRIEVAL ENGINE (DRE), which performs sub-linear retrieval using a histogram-based index followed by re-ranking for refined matching. To facilitate reproducible research, we release DANCETYPESBENCHMARK, a pose-aligned dataset annotated with quantised motion tokens. Experiments demonstrate robust retrieval across diverse dance styles and strong generalisation to unseen choreographies, establishing a foundation for scalable motion fingerprinting and quantitative choreographic analysis.

Method: Proposes DisCo framework with: 1) textual-visual decoupling module that isolates subject identity from reference image and modification commands from simplified text prompts, 2) reinforcement learning with dedicated reward signal to recouple subject and context naturally.

Result: Achieves state-of-the-art performance with simultaneous high-fidelity subject preservation and precise textual control, producing highly realistic and coherent images.

Conclusion: DisCo effectively resolves the similarity-controllability paradox by disentangling and recoupling visual and textual information, enabling better subject-driven T2I generation.

Abstract: Subject-Driven Text-to-Image (T2I) Generation aims to preserve a subject’s identity while editing its context based on a text prompt. A core challenge in this task is the “similarity-controllability paradox”, where enhancing textual control often degrades the subject’s fidelity, and vice-versa. We argue this paradox stems from the ambiguous role of text prompts, which are often tasked with describing both the subject and the desired modifications, leading to conflicting signals for the model. To resolve this, we propose DisCo, a novel framework that first Disntangles and then re-Couples visual and textual information. First, our textual-visual decoupling module isolates the sources of information: subject identity is extracted exclusively from the reference image with the entity word of the subject, while the text prompt is simplified to contain only the modification command, where the subject refers to general pronouns, eliminating descriptive ambiguity. However, this strict separation can lead to unnatural compositions between the subject and its contexts. We address this by designing a dedicated reward signal and using reinforcement learning to seamlessly recouple the visually-defined subject and the textually-generated context. Our approach effectively resolves the paradox, enabling simultaneous high-fidelity subject preservation and precise textual control. Extensive experiments demonstrate that our method achieves state-of-the-art performance, producing highly realistic and coherent images.

Method: Proposes MotionGrounder with: 1) Flow-based Motion Signal (FMS) for stable motion prior, 2) Object-Caption Alignment Loss (OCAL) to ground object captions to spatial regions, and 3) Object Grounding Score (OGS) for evaluation.

Result: MotionGrounder consistently outperforms recent baselines across quantitative, qualitative, and human evaluations, demonstrating effective multi-object motion transfer.

Conclusion: MotionGrounder successfully addresses the limitation of single-object motion transfer by enabling multi-object controllability through flow-based motion signals and object-caption alignment.

Abstract: Motion transfer enables controllable video generation by transferring temporal dynamics from a reference video to synthesize a new video conditioned on a target caption. However, existing Diffusion Transformer (DiT)-based methods are limited to single-object videos, restricting fine-grained control in real-world scenes with multiple objects. In this work, we introduce MotionGrounder, a DiT-based framework that firstly handles motion transfer with multi-object controllability. Our Flow-based Motion Signal (FMS) in MotionGrounder provides a stable motion prior for target video generation, while our Object-Caption Alignment Loss (OCAL) grounds object captions to their corresponding spatial regions. We further propose a new Object Grounding Score (OGS), which jointly evaluates (i) spatial alignment between source video objects and their generated counterparts and (ii) semantic consistency between each generated object and its target caption. Our experiments show that MotionGrounder consistently outperforms recent baselines across quantitative, qualitative, and human evaluations.

Method: The P&R framework has two components: (1) Structure Perturbation and Restoration Simulation - generates synthetic abnormal chromosomes by perturbing chromosomal banding patterns of normal chromosomes, then uses a restoration diffusion network to reconstruct continuous chromosome content and edges; (2) Energy-guided Adaptive Sampling - an energy score-based online selection strategy that dynamically prioritizes high-quality synthetic samples by referencing the energy distribution of real samples.

Result: The method was evaluated on a comprehensive structural anomaly dataset of over 260,000 chromosome images including 4,242 abnormal samples across 24 categories. P&R achieved state-of-the-art performance with average improvements of 8.92% in sensitivity, 8.89% in precision, and 13.79% in F1-score across all categories.

Conclusion: The proposed P&R framework effectively addresses data imbalance in chromosome anomaly detection through simulation-driven structural augmentation, eliminating reliance on rare abnormal samples and achieving superior performance compared to existing methods.

Abstract: Detecting structural chromosomal abnormalities is crucial for accurate diagnosis and management of genetic disorders. However, collecting sufficient structural abnormality data is extremely challenging and costly in clinical practice, and not all abnormal types can be readily collected. As a result, deep learning approaches face significant performance degradation due to the severe imbalance and scarcity of abnormal chromosome data. To address this challenge, we propose a Perturb-and-Restore (P&R), a simulation-driven structural augmentation framework that effectively alleviates data imbalance in chromosome anomaly detection. The P&R framework comprises two key components: (1) Structure Perturbation and Restoration Simulation, which generates synthetic abnormal chromosomes by perturbing chromosomal banding patterns of normal chromosomes followed by a restoration diffusion network that reconstructs continuous chromosome content and edges, thus eliminating reliance on rare abnormal samples; and (2) Energy-guided Adaptive Sampling, an energy score-based online selection strategy that dynamically prioritizes high-quality synthetic samples by referencing the energy distribution of real samples. To evaluate our method, we construct a comprehensive structural anomaly dataset consisting of over 260,000 chromosome images, including 4,242 abnormal samples spanning 24 categories. Experimental results demonstrate that the P&R framework achieves state-of-the-art (SOTA) performance, surpassing existing methods with an average improvement of 8.92% in sensitivity, 8.89% in precision, and 13.79% in F1-score across all categories.

Method: Proposes Sparkle, a structured representation that unifies skeletal joints and surface anchors with explicit kinematic-geometric factorization. The SparkleMotion framework learns this representation through hierarchical modules that embed geometric continuity and kinematic constraints, explicitly disentangling internal kinematic structure from external surface geometry.

Result: Achieves state-of-the-art performance in accuracy, robustness, and generalization under severe domain shifts, noise, and occlusion. Demonstrates superiority across diverse sensor types and challenging real-world scenarios.

Conclusion: The Sparkle representation effectively balances expressiveness and robustness for point cloud-based motion capture by unifying skeletal and surface information through explicit kinematic-geometric factorization, enabling superior performance in challenging conditions.

Abstract: Point cloud-based motion capture leverages rich spatial geometry and privacy-preserving sensing, but learning robust representations from noisy, unstructured point clouds remains challenging. Existing approaches face a struggle trade-off between point-based methods (geometrically detailed but noisy) and skeleton-based ones (robust but oversimplified). We address the fundamental challenge: how to construct an effective representation for human motion capture that can balance expressiveness and robustness. In this paper, we propose Sparkle, a structured representation unifying skeletal joints and surface anchors with explicit kinematic-geometric factorization. Our framework, SparkleMotion, learns this representation through hierarchical modules embedding geometric continuity and kinematic constraints. By explicitly disentangling internal kinematic structure from external surface geometry, SparkleMotion achieves state-of-the-art performance not only in accuracy but crucially in robustness and generalization under severe domain shifts, noise, and occlusion. Extensive experiments demonstrate our superiority across diverse sensor types and challenging real-world scenarios.

Method: Proposes an object-specific functional shape representation using Gaussian Process (GP) mixture models. Instead of heavy neural architectures, it uses lightweight GPs to learn continuous directional distance fields from sparsely sampled point clouds. Complex topologies are captured by anchoring local GP priors at strategic reference points extracted using structural decomposition methods like skeletonization or distance-based clustering.

Result: Extensive evaluations on ShapeNetCore and IndustryShapes datasets demonstrate the method can efficiently and accurately represent complex geometries.

Conclusion: The proposed GP-based functional representation provides an efficient, lightweight alternative to traditional 3D representations and computationally heavy neural architectures for capturing complex geometries.

Abstract: Traditional explicit 3D representations, such as point clouds and meshes, demand significant storage to capture fine geometric details and require complex indexing systems for surface lookups, making functional representations an efficient, compact, and continuous alternative. In this work, we propose a novel, object-specific functional shape representation that models surface geometry with Gaussian Process (GP) mixture models. Rather than relying on computationally heavy neural architectures, our method is lightweight, leveraging GPs to learn continuous directional distance fields from sparsely sampled point clouds. We capture complex topologies by anchoring local GP priors at strategic reference points, which can be flexibly extracted using any structural decomposition method (e.g. skeletonization, distance-based clustering). Extensive evaluations on the ShapeNetCore and IndustryShapes datasets demonstrate that our method can efficiently and accurately represent complex geometries.

Method: EgoSim models 3D scenes as updatable world states with two main components: 1) Geometry-action-aware Observation Simulation model for generating embodiment interactions, and 2) Interaction-aware State Updating module for spatial consistency. The system uses a scalable pipeline to extract static point clouds, camera trajectories, and embodiment actions from large-scale monocular egocentric videos, plus EgoCap for low-cost real-world data collection with uncalibrated smartphones.

Result: EgoSim significantly outperforms existing methods in visual quality, spatial consistency, and generalization to complex scenes and in-the-wild dexterous interactions. It also supports cross-embodiment transfer to robotic manipulation.

Conclusion: EgoSim provides a robust solution for egocentric world simulation with persistent 3D scene state updates, addressing key limitations of existing approaches and enabling better simulation of complex interactions while supporting robotic applications.

Abstract: We introduce EgoSim, a closed-loop egocentric world simulator that generates spatially consistent interaction videos and persistently updates the underlying 3D scene state for continuous simulation. Existing egocentric simulators either lack explicit 3D grounding, causing structural drift under viewpoint changes, or treat the scene as static, failing to update world states across multi-stage interactions. EgoSim addresses both limitations by modeling 3D scenes as updatable world states. We generate embodiment interactions via a Geometry-action-aware Observation Simulation model, with spatial consistency from an Interaction-aware State Updating module. To overcome the critical data bottleneck posed by the difficulty in acquiring densely aligned scene-interaction training pairs, we design a scalable pipeline that extracts static point clouds, camera trajectories, and embodiment actions from in-the-wild large-scale monocular egocentric videos. We further introduce EgoCap, a capture system that enables low-cost real-world data collection with uncalibrated smartphones. Extensive experiments demonstrate that EgoSim significantly outperforms existing methods in terms of visual quality, spatial consistency, and generalization to complex scenes and in-the-wild dexterous interactions, while supporting cross-embodiment transfer to robotic manipulation. Codes and datasets will be open soon. The project page is at egosimulator.github.io.

Method: The framework uses point tracking, depth estimation, and segmentation models to create a coherent 4D model with spatiotemporally consistent tool and tissue semantics. A Multimodal Large Language Model (MLLM) then acts as an agent on tools derived from the explicit 4D representation (e.g., trajectories) without any fine-tuning.

Result: The method was evaluated on a new dataset of 134 clinically relevant questions. The combination of a general purpose reasoning backbone and the 4D representation significantly improves spatiotemporal understanding and allows for 4D grounding. The approach demonstrates that spatiotemporal intelligence can be “assembled” from 2D MLLMs and 3D computer vision models without additional training.

Conclusion: The proposed framework successfully enables surgical AI agents to perform spatiotemporal reasoning by combining 2D MLLMs with explicit 4D representations, showing that complex spatiotemporal intelligence can be achieved without additional training by assembling existing components.

Abstract: Spatiotemporal reasoning is a fundamental capability for artificial intelligence (AI) in soft tissue surgery, paving the way for intelligent assistive systems and autonomous robotics. While 2D vision-language models show increasing promise at understanding surgical video, the spatial complexity of surgical scenes suggests that reasoning systems may benefit from explicit 4D representations. Here, we propose a framework for equipping surgical agents with spatiotemporal tools based on an explicit 4D representation, enabling AI systems to ground their natural language reasoning in both time and 3D space. Leveraging models for point tracking, depth, and segmentation, we develop a coherent 4D model with spatiotemporally consistent tool and tissue semantics. A Multimodal Large Language Model (MLLM) then acts as an agent on tools derived from the explicit 4D representation (e.g., trajectories) without any fine-tuning. We evaluate our method on a new dataset of 134 clinically relevant questions and find that the combination of a general purpose reasoning backbone and our 4D representation significantly improves spatiotemporal understanding and allows for 4D grounding. We demonstrate that spatiotemporal intelligence can be “assembled” from 2D MLLMs and 3D computer vision models without additional training. Code, data, and examples are available at https://tum-ai.github.io/surg4d/

Method: Constructed IKEA-Bench with 1,623 questions across 6 task types on 29 IKEA products. Evaluated 19 VLMs (2B-38B parameters) under three alignment strategies. Conducted three-level mechanistic analysis of visual subspaces.

Result: Key findings: (1) assembly instruction understanding recoverable via text, but text degrades diagram-to-video alignment; (2) architecture family predicts alignment accuracy more than parameter count; (3) video understanding remains hard bottleneck; diagrams and video occupy disjoint ViT subspaces; adding text shifts models from visual to text-driven reasoning.

Conclusion: Visual encoding is the primary target for improving cross-depiction robustness in VLMs for assembly understanding tasks.

Abstract: 2D assembly diagrams are often abstract and hard to follow, creating a need for intelligent assistants that can monitor progress, detect errors, and provide step-by-step guidance. In mixed reality settings, such systems must recognize completed and ongoing steps from the camera feed and align them with the diagram instructions. Vision Language Models (VLMs) show promise for this task, but face a depiction gap because assembly diagrams and video frames share few visual features. To systematically assess this gap, we construct IKEA-Bench, a benchmark of 1,623 questions across 6 task types on 29 IKEA furniture products, and evaluate 19 VLMs (2B-38B) under three alignment strategies. Our key findings: (1) assembly instruction understanding is recoverable via text, but text simultaneously degrades diagram-to-video alignment; (2) architecture family predicts alignment accuracy more strongly than parameter count; (3) video understanding remains a hard bottleneck unaffected by strategy. A three-level mechanistic analysis further reveals that diagrams and video occupy disjoint ViT subspaces, and that adding text shifts models from visual to text-driven reasoning. These results identify visual encoding as the primary target for improving cross-depiction robustness. Project page: https://ryenhails.github.io/IKEA-Bench/

Method: Formulates keyframe selection as maximizing conditional mutual information between selected frames and the query. Derives decomposed optimization to reduce subset selection to independent frame-level scoring. Introduces query-conditioned evidence scoring network trained with contrastive objective.

Result: Outperforms prior sampling strategies on long-form video understanding benchmarks under strict token budgets, while significantly improving training efficiency.

Conclusion: The evidence-driven framework provides a principled approach to keyframe sampling that effectively balances information preservation with computational constraints for long-form video understanding.

Abstract: Multimodal Large Language Models (MLLMs) have shown strong performance on video question answering, but their application to long-form videos is constrained by limited context length and computational cost, making keyframe sampling essential. Existing approaches typically rely on semantic relevance or reinforcement learning, which either fail to capture evidential clues or suffer from inefficient combinatorial optimization. In this work, we propose an evidence-driven keyframe sampling framework grounded in information bottleneck theory. We formulate keyframe selection as maximizing the conditional mutual information between selected frames and the query, providing a principled objective that reflects each frame’s contribution to answering the question. To make this objective tractable, we exploit its structure to derive a decomposed optimization that reduces subset selection to independent frame-level scoring. We further introduce a query-conditioned evidence scoring network trained with a contrastive objective to estimate evidential importance efficiently. Experiments on long-form video understanding benchmarks show that our method consistently outperforms prior sampling strategies under strict token budgets, while significantly improving training efficiency.

Method: Proposes Adversarial Attenuation Patch (AAP) method using energy-constrained optimization strategy coupled with attenuation-based deployment framework to balance attack effectiveness and stealthiness, aligning with signal-level electronic jamming mechanisms.

Result: AAP effectively degrades detection performance while preserving high imperceptibility, shows favorable transferability across different models, and demonstrates strong potential for physical realization.

Conclusion: Provides a physically-grounded perspective for adversarial attacks on SAR target detection systems and facilitates design of more covert and practically deployable attack strategies.

Abstract: Deep neural networks have demonstrated excellent performance in SAR target detection tasks but remain susceptible to adversarial attacks. Existing SAR-specific attack methods can effectively deceive detectors; however, they often introduce noticeable perturbations and are largely confined to digital domain, neglecting physical implementation constrains for attacking SAR systems. In this paper, a novel Adversarial Attenuation Patch (AAP) method is proposed that employs energy-constrained optimization strategy coupled with an attenuation-based deployment framework to achieve a seamless balance between attack effectiveness and stealthiness. More importantly, AAP exhibits strong potential for physical realization by aligning with signal-level electronic jamming mechanisms. Experimental results show that AAP effectively degrades detection performance while preserving high imperceptibility, and shows favorable transferability across different models. This study provides a physical grounded perspective for adversarial attacks on SAR target detection systems and facilitates the design of more covert and practically deployable attack strategies. The source code is made available at https://github.com/boremycin/SAAP.

Method: Proposes Identity-Decoupled personalized Diffusion Models (IDDM), a model-side defense that integrates identity decoupling into the personalization pipeline. Uses an alternating procedure that interleaves short personalization updates with identity-decoupled data optimization, employing a two-stage schedule to balance identity linkability suppression and generation utility.

Result: Extensive experiments across multiple datasets, diverse prompts, and state-of-the-art face recognition systems show that IDDM consistently reduces identity linkability while preserving high-quality personalized generation, with tunable control over the privacy-utility trade-off.

Conclusion: IDDM addresses the novel problem of model-side output immunization for personalized diffusion models, enabling authorized personalization while protecting user privacy by reducing identity linkability in generated outputs shared on social media platforms.

Abstract: Personalized text-to-image diffusion models (e.g., DreamBooth, LoRA) enable users to synthesize high-fidelity avatars from a few reference photos for social expression. However, once these generations are shared on social media platforms (e.g., Instagram, Facebook), they can be linked to the real user via face recognition systems, enabling identity tracking and profiling. Existing defenses mainly follow an anti-personalization strategy that protects publicly released reference photos by disrupting model fine-tuning. While effective against unauthorized personalization, they do not address another practical setting in which personalization is authorized, but the resulting public outputs still leak identity information. To address this problem, we introduce a new defense setting, termed model-side output immunization, whose goal is to produce a personalized model that supports authorized personalization while reducing the identity linkability of public generations, with tunable control over the privacy-utility trade-off to accommodate diverse privacy needs. To this end, we propose Identity-Decoupled personalized Diffusion Models (IDDM), a model-side defense that integrates identity decoupling into the personalization pipeline. Concretely, IDDM follows an alternating procedure that interleaves short personalization updates with identity-decoupled data optimization, using a two-stage schedule to balance identity linkability suppression and generation utility. Extensive experiments across multiple datasets, diverse prompts, and state-of-the-art face recognition systems show that IDDM consistently reduces identity linkability while preserving high-quality personalized generation.

Method: Systematically refined seven existing Japanese benchmark datasets through two rounds of human annotation to improve data quality and evaluation reliability.

Result: JAMMEval yields evaluation scores that better reflect model capability, exhibit lower run-to-run variance, and improve ability to distinguish between models of different capability levels.

Conclusion: JAMMEval advances reliable evaluation of VLMs for Japanese VQA, with released dataset and code to support the research community.

Abstract: Reliable evaluation is essential for the development of vision-language models (VLMs). However, Japanese VQA benchmarks have undergone far less iterative refinement than their English counterparts. As a result, many existing benchmarks contain issues such as ambiguous questions, incorrect answers, and instances that can be solved without visual grounding, undermining evaluation reliability and leading to misleading conclusions in model comparisons. To address these limitations, we introduce JAMMEval, a refined collection of Japanese benchmarks for reliable VLM evaluation. It is constructed by systematically refining seven existing Japanese benchmark datasets through two rounds of human annotation, improving both data quality and evaluation reliability. In our experiments, we evaluate open-weight and proprietary VLMs on JAMMEval and analyze the capabilities of recent models on Japanese VQA. We further demonstrate the effectiveness of our refinement by showing that the resulting benchmarks yield evaluation scores that better reflect model capability, exhibit lower run-to-run variance, and improve the ability to distinguish between models of different capability levels. We release our dataset and code to advance reliable evaluation of VLMs.

Method: Proposes a 3D Gaussian Splatting avatar model with spatial MLP backbone conditioned on both pose and appearance latent. The latent is learned during training by an encoder for compact representation. At inference, an autoregressive predictor infers the latent for temporally smooth appearance evolution.

Result: Method delivers robust and practical path to high-fidelity, stable avatar driving with improved reconstruction quality and disambiguation of pose-driven renderings.

Conclusion: The approach provides stable appearance evolution and improved stability for photorealistic human avatar generation by addressing dataset ambiguities through pose-appearance conditioning and latent space learning.

Abstract: A photorealistic and immersive human avatar experience demands capturing fine, person-specific details such as cloth and hair dynamics, subtle facial expressions, and characteristic motion patterns. Achieving this requires large, high-quality datasets, which often introduce ambiguities and spurious correlations when very similar poses correspond to different appearances. Models that fit these details during training can overfit and produce unstable, abrupt appearance changes for novel poses. We propose a 3D Gaussian Splatting avatar model with a spatial MLP backbone that is conditioned on both pose and an appearance latent. The latent is learned during training by an encoder, yielding a compact representation that improves reconstruction quality and helps disambiguate pose-driven renderings. At driving time, our predictor autoregressively infers the latent, producing temporally smooth appearance evolution and improved stability. Overall, our method delivers a robust and practical path to high-fidelity, stable avatar driving.

Method: Constructed EmoScene dataset with 1.2M images across 300+ scene categories, each annotated with discrete emotion labels, continuous VAD values, perceptual descriptors, and textual captions. Developed a lightweight baseline that injects dual-space controls into frozen diffusion backbone via shallow cross-attention modulation.

Result: Multi-space analyses reveal how discrete emotions occupy the VAD space and how affect systematically correlates with scene-level perceptual factors. The dataset enables benchmarking of affect controllability in image generation.

Conclusion: EmoScene addresses the gap in unified affective-perceptual representation for text-to-image models, providing a foundation for more emotionally coherent scene generation through dual-space supervision.

Abstract: Text-to-image diffusion models have achieved high visual fidelity, yet precise control over scene semantics and fine-grained affective tone remains challenging. Human visual affect arises from the rapid integration of contextual meaning, including valence, arousal, and dominance, with perceptual cues such as color harmony, luminance contrast, texture variation, curvature, and spatial layout. However, current text-to-image models rarely represent affective and perceptual factors within a unified representation, which limits their ability to synthesize scenes with coherent and nuanced emotional intent. To address this gap, we construct EmoScene, a large-scale dual-space emotion dataset that jointly encodes affective dimensions and perceptual attributes, with contextual semantics provided as supporting annotations. EmoScene contains 1.2M images across more than three hundred real-world scene categories, each annotated with discrete emotion labels, continuous VAD values, perceptual descriptors and textual captions. Multi-space analyses reveal how discrete emotions occupy the VAD space and how affect systematically correlates with scene-level perceptual factors. To benchmark EmoScene, we provide a lightweight reference baseline that injects dual-space controls into a frozen diffusion backbone via shallow cross-attention modulation, serving as a reproducible probe of affect controllability enabled by dual-space supervision.

Method: Created YieldSAT dataset with 12.2 million yield samples at 10m resolution across 2,173 fields, pairing multispectral satellite imagery (113,555 images) with auxiliary environmental data, and explored deep learning models and data fusion architectures for pixel regression.

Result: Demonstrated potential of large-scale high-resolution crop yield prediction, highlighted challenges from distribution shifts, and showed performance gains with domain-informed Deep Ensemble approach.

Conclusion: YieldSAT enables scalable crop yield prediction research, addresses data scarcity issues, and provides benchmark for multimodal deep learning approaches in agricultural applications.

Abstract: Crop yield prediction requires substantial data to train scalable models. However, creating yield prediction datasets is constrained by high acquisition costs, heterogeneous data quality, and data privacy regulations. Consequently, existing datasets are scarce, low in quality, or limited to regional levels or single crop types, hindering the development of scalable data-driven solutions. In this work, we release YieldSAT, a large, high-quality, and multimodal dataset for high-resolution crop yield prediction. YieldSAT spans various climate zones across multiple countries, including Argentina, Brazil, Uruguay, and Germany, and includes major crop types, including corn, rapeseed, soybeans, and wheat, across 2,173 expert-curated fields. In total, over 12.2 million yield samples are available, each with a spatial resolution of 10 m. Each field is paired with multispectral satellite imagery, resulting in 113,555 labeled satellite images, complemented by auxiliary environmental data. We demonstrate the potential of large-scale and high-resolution crop yield prediction as a pixel regression task by comparing various deep learning models and data fusion architectures. Furthermore, we highlight open challenges arising from severe distribution shifts in the ground truth data under real-world conditions. To mitigate this, we explore a domain-informed Deep Ensemble approach that exhibits significant performance gains. The dataset is available at https://yieldsat.github.io/.

Method: Progressive optimization from latent space to intermediate layers, with l1 regularization to avoid unrealistic images. Extends to out-of-distribution settings with label mapping for label inconsistency.

Result: Achieves pixel-level reconstruction and outperforms baselines across various FL scenarios, including challenging OOD cases.

Conclusion: GIFD demonstrates stronger privacy leakage capabilities in FL, highlighting serious privacy risks even with gradient sharing.

Abstract: Federated Learning (FL) has emerged as a compelling paradigm for privacy-preserving distributed machine learning, allowing multiple clients to collaboratively train a global model by transmitting locally computed gradients to a central server without exposing their private data. Nonetheless, recent studies find that the gradients exchanged in the FL system are also vulnerable to privacy leakage, e.g., an attacker can invert shared gradients to reconstruct sensitive data by leveraging pre-trained generative adversarial networks (GAN) as prior knowledge. However, existing attacks simply perform gradient inversion in the latent space of the GAN model, which limits their expression ability and generalizability. To tackle these challenges, we propose \textbf{G}radient \textbf{I}nversion over \textbf{F}eature \textbf{D}omains (GIFD), which disassembles the GAN model and searches the hierarchical features of the intermediate layers. Instead of optimizing only over the initial latent code, we progressively change the optimized layer, from the initial latent space to intermediate layers closer to the output images. In addition, we design a regularizer to avoid unreal image generation by adding a small ${l_1}$ ball constraint to the searching range. We also extend GIFD to the out-of-distribution (OOD) setting, which weakens the assumption that the training sets of GANs and FL tasks obey the same data distribution. Furthermore, we consider the challenging OOD scenario of label inconsistency and propose a label mapping technique as an effective solution. Extensive experiments demonstrate that our method can achieve pixel-level reconstruction and outperform competitive baselines across a variety of FL scenarios.

Method: Integrates lightweight Mixture-of-Adapters (MoA) module into pretrained ViT-B/32 backbone with selective fine-tuning of final layers. Uses spatially-aware adaptation guided by global semantic context vector and hybrid prediction architecture combining depth bin classification with direct regression. Employs composite loss function with geometric constraints.

Result: On NYU Depth V2 benchmark: improves δ₁ accuracy from 0.390 to 0.745, reduces RMSE from 1.176 to 0.520. Achieves competitive results with substantially fewer trainable parameters compared to baseline.

Conclusion: Lightweight, prompt-guided Mixture-of-Adapters is an effective strategy for transferring VLM knowledge to fine-grained monocular depth estimation tasks with minimal supervision and parameter efficiency.

Abstract: Leveraging the rich semantic features of vision-language models (VLMs) like CLIP for monocular depth estimation tasks is a promising direction, yet often requires extensive fine-tuning or lacks geometric precision. We present a parameter-efficient framework, named MoA-DepthCLIP, that adapts pretrained CLIP representations for monocular depth estimation with minimal supervision. Our method integrates a lightweight Mixture-of-Adapters (MoA) module into the pretrained Vision Transformer (ViT-B/32) backbone combined with selective fine-tuning of the final layers. This design enables spatially-aware adaptation, guided by a global semantic context vector and a hybrid prediction architecture that synergizes depth bin classification with direct regression. To enhance structural accuracy, we employ a composite loss function that enforces geometric constraints. On the NYU Depth V2 benchmark, MoA-DepthCLIP achieves competitive results, significantly outperforming the DepthCLIP baseline by improving the $δ_1$ accuracy from 0.390 to 0.745 and reducing the RMSE from 1.176 to 0.520. These results are achieved while requiring substantially few trainable parameters, demonstrating that lightweight, prompt-guided MoA is a highly effective strategy for transferring VLM knowledge to fine-grained monocular depth estimation tasks.

Method: Two-stage approach: 1) Self-supervised reconstruction of 3D Gaussians from multi-view semantic and depth images using queries. 2) Dual latent world models: Gaussian-flow-guided latent prediction for occupancy perception/forecasting, and ego-planning-guided latent prediction for motion planning, both leveraging fine-grained contextual features.

Result: Extensive experiments on SurroundOcc and nuScenes benchmarks show significant performance gains across Gaussian-centric 3D occupancy perception, 4D occupancy forecasting, and motion planning tasks compared to existing methods.

Conclusion: DLWM provides a novel holistic pre-training paradigm for Gaussian-centric autonomous driving that effectively bridges perception, forecasting, and planning through dual latent world models, demonstrating strong performance across multiple downstream tasks.

Abstract: Vision-based autonomous driving has gained much attention due to its low costs and excellent performance. Compared with dense BEV (Bird’s Eye View) or sparse query models, Gaussian-centric method is a comprehensive yet sparse representation by describing scene with 3D semantic Gaussians. In this paper, we introduce DLWM, a novel paradigm with Dual Latent World Models specifically designed to enable holistic gaussian-centric pre-training in autonomous driving using two stages. In the first stage, DLWM predicts 3D Gaussians from queries by self-supervised reconstructing multi-view semantic and depth images. Equipped with fine-grained contextual features, in the second stage, two latent world models are trained separately for temporal feature learning, including Gaussian-flow-guided latent prediction for downstream occupancy perception and forecasting tasks, and ego-planning-guided latent prediction for motion planning. Extensive experiments in SurroundOcc and nuScenes benchmarks demonstrate that DLWM shows significant performance gains across Gaussian-centric 3D occupancy perception, 4D occupancy forecasting and motion planning tasks.

Method: ACT uses two main components: 1) Visual context exploration that leverages spatio-temporal profiling to adaptively amplify attention heads responsible for visual exploration, and 2) Semantic context aggregation that marginalizes potential semantic queries to effectively aggregate visual evidence and resolve information loss from discrete token prediction.

Result: Extensive experiments across diverse LVLMs show ACT significantly reduces hallucinations and achieves competitive results on both discriminative and generative benchmarks without compromising fundamental generation capabilities.

Conclusion: ACT provides a robust and highly adaptable training-free solution for mitigating hallucinations in LVLMs by addressing dynamic context changes and information loss during generation.

Abstract: Large Vision-Language Models (LVLMs) frequently suffer from severe hallucination issues. Existing mitigation strategies predominantly rely on isolated, single-step states to enhance visual focus or suppress strong linguistic priors. However, these static approaches neglect dynamic context changes across the generation process and struggles to correct inherited information loss. To address this limitation, we propose Adaptive Context inTegration (ACT), a training-free inference intervention method that mitigates hallucination through the adaptive integration of contextual information. Specifically, we first propose visual context exploration, which leverages spatio-temporal profiling to adaptively amplify attention heads responsible for visual exploration. To further facilitate vision-language alignment, we propose semantic context aggregation that marginalizes potential semantic queries to effectively aggregate visual evidence, thereby resolving the information loss caused by the discrete nature of token prediction. Extensive experiments across diverse LVLMs demonstrate that ACT significantly reduces hallucinations and achieves competitive results on both discriminative and generative benchmarks, acting as a robust and highly adaptable solution without compromising fundamental generation capabilities.

Method: Anchors latent feature vectors on virtual scaffolds around each primitive, interpolates features at ray intersections, applies periodic activations (inspired by Fourier analysis) to create harmonic components, then decodes with a small neural network in a deferred pass.

Result: Achieves state-of-the-art results in real-time novel view synthesis, integrates seamlessly with existing primitive-based pipelines (3DGUT, Triangle Splatting, 2DGS), and demonstrates generality with 2D image fitting and semantic reconstruction applications.

Conclusion: Neural Harmonic Textures successfully bridges primitive- and neural-field-based reconstruction while maintaining real-time performance and computational efficiency, offering a flexible approach that enhances detail modeling in 3D representations.

Abstract: Primitive-based methods such as 3D Gaussian Splatting have recently become the state-of-the-art for novel-view synthesis and related reconstruction tasks. Compared to neural fields, these representations are more flexible, adaptive, and scale better to large scenes. However, the limited expressivity of individual primitives makes modeling high-frequency detail challenging. We introduce Neural Harmonic Textures, a neural representation approach that anchors latent feature vectors on a virtual scaffold surrounding each primitive. These features are interpolated within the primitive at ray intersection points. Inspired by Fourier analysis, we apply periodic activations to the interpolated features, turning alpha blending into a weighted sum of harmonic components. The resulting signal is then decoded in a single deferred pass using a small neural network, significantly reducing computational cost. Neural Harmonic Textures yield state-of-the-art results in real-time novel view synthesis while bridging the gap between primitive- and neural-field-based reconstruction. Our method integrates seamlessly into existing primitive-based pipelines such as 3DGUT, Triangle Splatting, and 2DGS. We further demonstrate its generality with applications to 2D image fitting and semantic reconstruction.

Method: Proposes an expectation-maximization framework where: 1) E-step uses a flow matching-based generative model to approximate latent DWI image posterior distribution given T2w images; 2) M-step simultaneously optimizes a cancer localizer and generative model to maximize expected likelihood of cancer presence. Treats DWI as a latent modality available only during training.

Result: Superior cancer localization performance compared to approaches without training DWI images or existing privileged learning frameworks. T2-only methods perform competitively or better than baseline methods using multiple input sequences (improving patient-level F1 score by 14.4% and zone-level QWK by 5.3% over T2w+DWI baseline). Evaluated on 4,133 prostate cancer patients with histopathology-verified labels across internal and external datasets.

Conclusion: The proposed framework provides a novel theoretical approach for learning from privileged modalities, enabling high-performance cancer localization using only T2-weighted MRI at inference while leveraging DWI during training, offering clinical benefits through reduced acquisition costs and expertise requirements.

Abstract: Multiparametric MRI is increasingly recommended as a first-line noninvasive approach to detect and localize prostate cancer, requiring at minimum diffusion-weighted (DWI) and T2-weighted (T2w) MR sequences. Early machine learning attempts using only T2w images have shown promising diagnostic performance in segmenting radiologist-annotated lesions. Such uni-modal T2-only approaches deliver substantial clinical benefits by reducing costs and expertise required to acquire other sequences. This work investigates an arguably more challenging application using only T2w at inference, but to localize individual cancers based on independent histopathology labels. We formulate DWI images as a latent modality (readily available during training) to classify cancer presence at local Barzell zones, given only T2w images as input. In the resulting expectation-maximization algorithm, a latent modality generator (implemented using a flow matching-based generative model) approximates the latent DWI image posterior distribution in the E-steps, while in M-steps a cancer localizer is simultaneously optimized with the generative model to maximize the expected likelihood of cancer presence. The proposed approach provides a novel theoretical framework for learning from a privileged DWI modality, yielding superior cancer localization performance compared to approaches that lack training DWI images or existing frameworks for privileged learning and incomplete modalities. The proposed T2-only methods perform competitively or better than baseline methods using multiple input sequences (e.g., improving the patient-level F1 score by 14.4% and zone-level QWK by 5.3% over the T2w+DWI baseline). We present quantitative evaluations using internal and external datasets from 4,133 prostate cancer patients with histopathology-verified labels.

Method: Reformulates ground-roll attenuation as semantic-guided signal separation. Uses promptable large vision model to extract high-level semantic priors by converting seismic gathers to visual representations and localizing ground-roll regions via text/image prompts. Semantic response is transformed into continuous soft mask embedded into mask-conditioned low-rank inverse formulation for spatially adaptive suppression. ADMM-based solver enables stable recovery without task-specific training.

Result: Extensive experiments on synthetic and field VSP datasets demonstrate superior ground-roll attenuation while preserving reflection continuity and waveform fidelity, outperforming transform-domain filtering and implicit neural representation methods.

Conclusion: Proposed training-free framework effectively addresses ground-roll attenuation using semantic guidance from vision models, enabling adaptive suppression without requiring labeled training data or manual annotation.

Abstract: Ground-roll is a dominant source of coherent noise in land and vertical seismic profiling (VSP) data, severely masking reflection events and degrading subsequent imaging and interpretation. Conventional attenuation methods, including transform-domain filtering, sparse representation, and deep learning, often suffer from limited adaptability, signal leakage, or dependence on labeled training data, especially under strong signal-noise overlap. To address these challenges, we propose a training-free framework that reformulates ground-roll attenuation as a semantic-guided signal separation problem. Specifically, a promptable large vision model is employed to extract high-level semantic priors by converting seismic gathers into visual representations and localizing ground-roll-dominant regions via text or image prompts. The resulting semantic response is transformed into a continuous soft mask, which is embedded into a mask-conditioned low-rank inverse formulation to enable spatially adaptive suppression and reflection-preserving reconstruction. An efficient alternating direction method of multipliers (ADMM)-based solver is further developed to solve the proposed inverse problem, enabling stable and physically consistent signal recovery without requiring task-specific training or manual annotation. Extensive experiments on both synthetic and field VSP datasets demonstrate that the proposed method achieves superior ground-roll attenuation while preserving reflection continuity and waveform fidelity, consistently outperforming representative transform-domain filtering and implicit neural representation methods.

Method: Proposes dense point trajectories as structured mid-level visual tokens that disentangle motion from appearance. Uses a diffusion transformer to model unordered sets of trajectories with explicit occlusion reasoning for coherent motion forecasting.

Result: Achieves category-agnostic, data-efficient prediction that outperforms state-of-the-art baselines. Generalizes well to rare species and morphologies, demonstrating robust behavior forecasting capabilities.

Conclusion: Dense point trajectories provide an effective foundation for predictive visual intelligence in the wild, enabling coherent forecasts of complex motion patterns across diverse non-rigid agents.

Abstract: Visual intelligence requires anticipating the future behavior of agents, yet vision systems lack a general representation for motion and behavior. We propose dense point trajectories as visual tokens for behavior, a structured mid-level representation that disentangles motion from appearance and generalizes across diverse non-rigid agents, such as animals in-the-wild. Building on this abstraction, we design a diffusion transformer that models unordered sets of trajectories and explicitly reasons about occlusion, enabling coherent forecasts of complex motion patterns. To evaluate at scale, we curate 300 hours of unconstrained animal video with robust shot detection and camera-motion compensation. Experiments show that forecasting trajectory tokens achieves category-agnostic, data-efficient prediction, outperforms state-of-the-art baselines, and generalizes to rare species and morphologies, providing a foundation for predictive visual intelligence in the wild.

Method: Introduces Diff3R framework with a differentiable 3DGS optimization layer integrated into training. Uses Implicit Function Theorem and matrix-free PCG solver for efficient backpropagation through optimization steps. Incorporates data-driven uncertainty model to adaptively control parameter changes during optimization, preventing overfitting in under-constrained regions.

Result: The method can be integrated into existing feed-forward 3DGS architectures for both pose-given and pose-free methods, providing improvements for test-time optimization while being model-agnostic.

Conclusion: Diff3R successfully bridges feed-forward prediction and test-time optimization for 3DGS, offering a practical solution that combines speed with quality while being adaptable to existing architectures.

Abstract: Recent advances in 3D Gaussian Splatting (3DGS) present two main directions: feed-forward models offer fast inference in sparse-view settings, while per-scene optimization yields high-quality renderings but is computationally expensive. To combine the benefits of both, we introduce Diff3R, a novel framework that explicitly bridges feed-forward prediction and test-time optimization. By incorporating a differentiable 3DGS optimization layer directly into the training loop, our network learns to predict an optimal initialization for test-time optimization rather than a conventional zero-shot result. To overcome the computational cost of backpropagating through the optimization steps, we propose computing gradients via the Implicit Function Theorem and a scalable, matrix-free PCG solver tailored for 3DGS optimization. Additionally, we incorporate a data-driven uncertainty model into the optimization process by adaptively controlling how much the parameters are allowed to change during optimization. This approach effectively mitigates overfitting in under-constrained regions and increases robustness against input outliers. Since our proposed optimization layer is model-agnostic, we show that it can be seamlessly integrated into existing feed-forward 3DGS architectures for both pose-given and pose-free methods, providing improvements for test-time optimization.

Method: Open-source pipeline identifies stereo pairs from OHRC archives using geometric analysis (B/H ratio, convergence angles), performs dense stereo correspondence and ray triangulation to generate point clouds, grids into DEMs, and validates through ICP alignment with LRO NAC DTMs and constant-bias correction.

Result: Generated DEMs with 24-54 cm spatial resolution across five lunar sites, achieving vertical RMSE of 5.85 m and horizontal accuracy <30 cm when validated against NAC reference terrain.

Conclusion: Successfully demonstrated open-source pipeline for sub-meter lunar DEM generation from OHRC imagery, providing valuable high-resolution topographic data for lunar exploration and science.

Abstract: High-resolution digital elevation models (DEMs) of the lunar surface are essential for surface mobility planning, landing site characterization, and planetary science. The Orbiter High Resolution Camera (OHRC) on board Chandrayaan-2 has the best ground sampling capabilities of any lunar orbital imaging currently in use by acquiring panchromatic imagery at a resolution of roughly 20-30 cm per pixel. This work presents, for the first time, the generation of sub-metre DEMs from OHRC multi-view imagery using an exclusively open-source pipeline. Candidate stereo pairs are identified from non-paired OHRC archives through geometric analysis of image metadata, employing baseline-to-height (B/H) ratio computation and convergence angle estimation. Dense stereo correspondence and ray triangulation are then applied to generate point clouds, which are gridded into DEMs at effective spatial resolutions between approximately 24 and 54 cm across five geographically distributed lunar sites. Absolute elevation consistency is established through Iterative Closest Point (ICP) alignment against Lunar Reconnaissance Orbiter Narrow Angle Camera (NAC) Digital Terrain Models, followed by constant-bias offset correction. Validation against NAC reference terrain yields a vertical RMSE of 5.85 m (at native OHRC resolution), and a horizontal accuracy of less than 30 cm assessed by planimetric feature matching.

Method: IPnP iteratively generates and refines pseudo-labels for unlabeled organs through collaboration between a trainable segmentation network (specialist) and a frozen foundation model (generalist), progressively recovering full-organ supervision.

Result: On AMOS dataset with simulated partial-label setting, IPnP consistently improves segmentation performance over prior methods and approaches fully labeled reference performance. Effective on private dataset of 210 head-and-neck cancer patients in real-world clinical settings.

Conclusion: IPnP effectively addresses partial labeling in medical image segmentation through iterative prompting and pseudo-labeling, demonstrating strong performance in both simulated and real-world clinical scenarios.

Abstract: Automated medical image segmentation has achieved remarkable progress with fully labeled data. However, site-specific clinical priorities and the high cost of manual annotation often yield scans with only a subset of organs labeled, leading to the partially labeled problem that degrades performance. To address this issue, we propose IPnP, an Iteratively Prompting and Pseudo-labeling framework, for partially labeled medical image segmentation. IPnP iteratively generates and refines pseudo-labels for unlabeled organs through collaboration between a trainable segmentation network (specialist) and a frozen foundation model (generalist), progressively recovering full-organ supervision. On the public dataset AMOS with the simulated partial-label setting, IPnP consistently improves segmentation performance over prior methods and approaches the performance of the fully labeled reference. We further evaluate on a private, partially labeled dataset of 210 head-and-neck cancer patients and demonstrate our effectiveness in real-world clinical settings.

Method: Proposes canonical-space injection to decouple human dynamics from environmental cues via cross-attention, Dynamic-Grounded-RoPE for spatial correspondences without 3D alignments, and Hybrid Context Integration for long-horizon consistency.

Result: Significantly outperforms state-of-the-art methods, offering superior structural control and creative diversity for video synthesis.

Conclusion: ONE-SHOT provides an effective parameter-efficient framework for compositional video generation that overcomes limitations of existing 3D-based approaches while maintaining fine-grained control.

Abstract: Recent advances in Video Foundation Models (VFMs) have revolutionized human-centric video synthesis, yet fine-grained and independent editing of subjects and scenes remains a critical challenge. Recent attempts to incorporate richer environment control through rigid 3D geometric compositions often encounter a stark trade-off between precise control and generative flexibility. Furthermore, the heavy 3D pre-processing still limits practical scalability. In this paper, we propose ONE-SHOT, a parameter-efficient framework for compositional human-environment video generation. Our key insight is to factorize the generative process into disentangled signals. Specifically, we introduce a canonical-space injection mechanism that decouples human dynamics from environmental cues via cross-attention. We also propose Dynamic-Grounded-RoPE, a novel positional embedding strategy that establishes spatial correspondences between disparate spatial domains without any heuristic 3D alignments. To support long-horizon synthesis, we introduce a Hybrid Context Integration mechanism to maintain subject and scene consistency across minute-level generations. Experiments demonstrate that our method significantly outperforms state-of-the-art methods, offering superior structural control and creative diversity for video synthesis. Our project has been available on: https://martayang.github.io/ONE-SHOT/.

Method: Manually curated dashcam segments from publicly available YouTube videos, screened to exclude crashes and edited content. Provides segment-level manual labels (time of day, vehicle type) and machine-generated object detections using YOLOv11x with BoT-SORT tracking.

Result: 51,753 segment records spanning 20,275.56 hours from 42,032 videos, covering 7,103 locations across 238 countries/territories on all inhabited continents. Includes CSV files with machine-generated detections for 80 MS-COCO classes.

Conclusion: CROWD provides a valuable resource for autonomous driving research focused on routine scenarios, with broad geographic coverage and annotations that lower the barrier for benchmarking cross-domain robustness and interaction analysis.

Abstract: We introduce CROWD (City Road Observations With Dashcams), a manually curated dataset of ordinary, minute scale, temporally contiguous, unedited, front facing urban dashcam segments screened and segmented from publicly available YouTube videos. CROWD is designed to support cross-domain robustness and interaction analysis by prioritising routine driving and explicitly excluding crashes, crash aftermath, and other edited or incident-focused content. The release contains 51,753 segment records spanning 20,275.56 hours (42,032 videos), covering 7,103 named inhabited places in 238 countries and territories across all six inhabited continents (Africa, Asia, Europe, North America, South America and Oceania), with segment level manual labels for time of day (day or night) and vehicle type. To lower the barrier for benchmarking, we provide per-segment CSV files of machine-generated detections for all 80 MS-COCO classes produced with YOLOv11x, together with segment-local multi-object tracks (BoT-SORT); e.g. person, bicycle, motorcycle, car, bus, truck, traffic light, stop sign, etc. CROWD is distributed as video identifiers with segment boundaries and derived annotations, enabling reproducible research without redistributing the underlying videos.

Method: PHASOR treats CT volumes as coherent sequences and uses a video diffusion model for structural coherence. It introduces two modules: 1) Anatomy-routed mixture-of-experts (AR-MoE) anchors enhancement patterns to anatomical semantics with organ-specific memory, and 2) Intensity-phase aware representation alignment (IP-REPA) highlights contrast signals while mitigating imperfect spatial alignment.

Result: Extensive experiments across three datasets demonstrate that PHASOR significantly outperforms state-of-the-art methods in both synthesis quality and enhancement accuracy.

Conclusion: PHASOR provides a robust framework for high-fidelity virtual contrast enhancement in CT imaging, addressing key challenges in anatomical heterogeneity and spatial alignment through innovative diffusion-based approaches.

Abstract: Contrast-enhanced computed tomography (CECT) is pivotal for highlighting tissue perfusion and vascularity, yet its clinical ubiquity is impeded by the invasive nature of contrast agents and radiation risks. While virtual contrast enhancement (VCE) offers an alternative to synthesizing CECT from non-contrast CT (NCCT), existing methods struggle with anatomical heterogeneity and spatial misalignment, leading to inconsistent enhancement patterns and incorrect details. This paper introduces PHASOR, a volumetric diffusion framework for high-fidelity CT VCE. By treating CT volumes as coherent sequences, we leverage a video diffusion model to enhance structural coherence and volumetric accuracy. To ensure anatomy-phase consistent synthesis, we introduce two complementary modules. First, anatomy-routed mixture-of-experts (AR-MoE) anchors distinct enhancement patterns to anatomical semantics, with organ-specific memory to capture salient details. Second, intensity-phase aware representation alignment (IP-REPA) highlights intricate contrast signals while mitigating the impact of imperfect spatial alignment. Extensive experiments across three datasets demonstrate that PHASOR significantly outperforms state-of-the-art methods in both synthesis quality and enhancement accuracy.

Method: ReMoGen uses a universal motion prior from large-scale single-person datasets, adapts it to target domains via independently trained Meta-Interaction modules, performs segment-level generation, and incorporates a lightweight Frame-wise Segment Refinement module for online responsiveness.

Result: Extensive experiments across human-human, human-scene, and mixed-modality settings show ReMoGen produces high-quality, coherent, and responsive reactions while generalizing effectively across diverse interaction scenarios.

Conclusion: ReMoGen addresses real-time interaction-to-reaction generation challenges through modular design combining universal motion priors, domain adaptation, and frame-wise refinement, enabling robust generalization and responsive online interaction.

Abstract: Human behaviors in real-world environments are inherently interactive, with an individual’s motion shaped by surrounding agents and the scene. Such capabilities are essential for applications in virtual avatars, interactive animation, and human-robot collaboration. We target real-time human interaction-to-reaction generation, which generates the ego’s future motion from dynamic multi-source cues, including others’ actions, scene geometry, and optional high-level semantic inputs. This task is fundamentally challenging due to (i) limited and fragmented interaction data distributed across heterogeneous single-person, human-human, and human-scene domains, and (ii) the need to produce low-latency yet high-fidelity motion responses during continuous online interaction. To address these challenges, we propose ReMoGen (Reaction Motion Generation), a modular learning framework for real-time interaction-to-reaction generation. ReMoGen leverages a universal motion prior learned from large-scale single-person motion datasets and adapts it to target interaction domains through independently trained Meta-Interaction modules, enabling robust generalization under data-scarce and heterogeneous supervision. To support responsive online interaction, ReMoGen performs segment-level generation together with a lightweight Frame-wise Segment Refinement module that incorporates newly observed cues at the frame level, improving both responsiveness and temporal coherence without expensive full-sequence inference. Extensive experiments across human-human, human-scene, and mixed-modality interaction settings show that ReMoGen produces high-quality, coherent, and responsive reactions, while generalizing effectively across diverse interaction scenarios.

Method: ProTPS learns class-specific vision prototypes and text prompts simultaneously, where vision prototypes guide the selection and learning of text prompts for each class to increase training flexibility and encourage unique prompt learning.

Result: ProTPS achieves performance close to upper bounds in class incremental (CI) and cross-datasets continual (CDC) settings, and shows favorable performance on the real-world Marine112 dataset under class and domain incremental (CDI) learning with natural long-tail distribution.

Conclusion: The proposed ProTPS method effectively mitigates catastrophic forgetting in continual learning by learning unique text prompts guided by vision prototypes, demonstrating strong performance across multiple settings including challenging real-world scenarios.

Abstract: For continual learning, text-prompt-based methods leverage text encoders and learnable prompts to encode semantic features for sequentially arrived classes over time. A common challenge encountered by existing works is how to learn unique text prompts, which implicitly carry semantic information of new classes, so that the semantic features of newly arrived classes do not overlap with those of trained classes, thereby mitigating the catastrophic forgetting problem. To address this challenge, we propose a novel approach Prototype-guided Text Prompt Selection (ProTPS)’’ to intentionally increase the training flexibility thus encouraging the learning of unique text prompts. Specifically, our ProTPS learns class-specific vision prototypes and text prompts. Vision prototypes guide the selection and learning of text prompts for each class. We first evaluate our ProTPS in both class incremental (CI) setting and cross-datasets continual (CDC) learning setting. Because our ProTPS achieves performance close to the upper bounds, we further collect a real-world dataset with 112 marine species collected over a span of six years, named Marine112, to bring new challenges to the community. Marine112 is authentically suited for the class and domain incremental (CDI) learning setting and is under natural long-tail distribution. The results under three settings show that our ProTPS performs favorably against the recent state-of-the-art methods. The implementation code and Marine112 dataset will be released upon the acceptance of our paper.

Method: 1) Construct dynamic 3D point cloud scenes from multi-frame LiDAR data; 2) Use vehicle completion module to reconstruct full 360° geometry from partial observations; 3) Render edited scenes into 2D condition images; 4) Guide video diffusion model to synthesize driving videos; 5) Apply RL-based post-training with pairwise preference model and reward mechanism for out-of-distribution robustness.

Result: ReinDriveGen outperforms existing approaches on edited driving scenarios and achieves state-of-the-art results on novel ego viewpoint synthesis. The RL enhancement improves quality under out-of-distribution conditions without ground-truth supervision.

Conclusion: The framework enables full controllability over dynamic driving scenes for safety-critical scenario generation, addressing distribution gaps through RL-based post-training, making it valuable for autonomous vehicle testing and simulation.

Abstract: We present ReinDriveGen, a framework that enables full controllability over dynamic driving scenes, allowing users to freely edit actor trajectories to simulate safety-critical corner cases such as front-vehicle collisions, drifting cars, vehicles spinning out of control, pedestrians jaywalking, and cyclists cutting across lanes. Our approach constructs a dynamic 3D point cloud scene from multi-frame LiDAR data, introduces a vehicle completion module to reconstruct full 360° geometry from partial observations, and renders the edited scene into 2D condition images that guide a video diffusion model to synthesize realistic driving videos. Since such edited scenarios inevitably fall outside the training distribution, we further propose an RL-based post-training strategy with a pairwise preference model and a pairwise reward mechanism, enabling robust quality improvement under out-of-distribution conditions without ground-truth supervision. Extensive experiments demonstrate that ReinDriveGen outperforms existing approaches on edited driving scenarios and achieves state-of-the-art results on novel ego viewpoint synthesis.

Method: 1) Adapt general open-set anomaly detection methods for 3D point cloud inputs; 2) Propose Open3D-AD that leverages normal samples, simulated anomalies, and partially observed real anomalies to model probability density distributions; 3) Introduce Correspondence Distributions Subsampling to reduce overlap between normal and non-normal distributions for stronger dual distribution modeling.

Result: Established comprehensive benchmark on Open-Industry dataset (15 categories, 5 real anomaly types) and evaluated on Real3D-AD and Anomaly-ShapeNet. Benchmark results and ablation studies demonstrate Open3D-AD’s effectiveness and reveal potential of open-set supervised 3D anomaly detection.

Conclusion: Open3D-AD provides an effective approach for open-set supervised 3D anomaly detection in industrial settings, showing promise for practical applications where limited anomalous samples are available alongside normal data.

Abstract: Although self-supervised 3D anomaly detection assumes that acquiring high-precision point clouds is computationally expensive, in real manufacturing scenarios it is often feasible to collect a limited number of anomalous samples. Therefore, we study open-set supervised 3D anomaly detection, where the model is trained with only normal samples and a small number of known anomalous samples, aiming to identify unknown anomalies at test time. We present Open-Industry, a high-quality industrial dataset containing 15 categories, each with five real anomaly types collected from production lines. We first adapt general open-set anomaly detection methods to accommodate 3D point cloud inputs better. Building upon this, we propose Open3D-AD, a point-cloud-oriented approach that leverages normal samples, simulated anomalies, and partially observed real anomalies to model the probability density distributions of normal and anomalous data. Then, we introduce a simple Correspondence Distributions Subsampling to reduce the overlap between normal and non-normal distributions, enabling stronger dual distributions modeling. Based on these contributions, we establish a comprehensive benchmark and evaluate the proposed method extensively on Open-Industry as well as established datasets including Real3D-AD and Anomaly-ShapeNet. Benchmark results and ablation studies demonstrate the effectiveness of Open3D-AD and further reveal the potential of open-set supervised 3D anomaly detection.

Method: Three-stage framework: (1) Multi-view 3D-Anchor Synthesis using MV-TRACE dataset for consistent object addition/modification, (2) Tangible Geometry Anchoring for precise mesh-3DGS synchronization via two-phase registration, (3) Contextual Video Masking integrating 3D projections into autoregressive video pipeline for temporally stable rendering.

Result: TRACE consistently outperforms existing methods in editing versatility and structural integrity, achieving high-fidelity scene transformation with fine-grained part-level manipulation.

Conclusion: TRACE provides a novel mesh-guided approach to 3DGS editing that enables automated, high-fidelity scene transformation with unprecedented part-level manipulation capabilities while preserving structural integrity.

Abstract: We present TRACE, a mesh-guided 3DGS editing framework that achieves automated, high-fidelity scene transformation. By anchoring video diffusion with explicit 3D geometry, TRACE uniquely enables fine-grained, part-level manipulatio–such as local pose shifting or component replacemen–while preserving the structural integrity of the central subject, a capability largely absent in existing editing methods. Our approach comprises three key stages: (1) Multi-view 3D-Anchor Synthesis, which leverages a sparse-view editor trained on our MV-TRACE datase–the first multi-view consistent dataset dedicated to scene-coherent object addition and modificatio–to generate spatially consistent 3D-anchors; (2) Tangible Geometry Anchoring (TGA), which ensures precise spatial synchronization between inserted meshes and the 3DGS scene via two-phase registration; and (3) Contextual Video Masking (CVM), which integrates 3D projections into an autoregressive video pipeline to achieve temporally stable, physically-grounded rendering. Extensive experiments demonstrate that TRACE consistently outperforms existing methods especially in editing versatility and structural integrity.

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 2303.08250: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2303.08250&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 2407.01967: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2407.01967&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 2408.06788: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2408.06788&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to technical error in fetching paper content

Result: Unable to determine results due to technical error in fetching paper content

Conclusion: Unable to draw conclusions due to technical error in fetching paper content

Abstract: Failed to fetch summary for 2409.20146: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2409.20146&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to technical error in fetching paper content

Result: Unable to determine results due to technical error in fetching paper content

Conclusion: Unable to draw conclusions due to technical error in fetching paper content

Abstract: Failed to fetch summary for 2410.19899: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2410.19899&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 2411.13181: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2411.13181&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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

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

Conclusion: Unable to draw conclusions as paper content could not be retrieved

Abstract: Failed to fetch summary for 2502.14883: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2502.14883&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 2503.19851: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2503.19851&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 2504.13129: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2504.13129&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 2505.08614: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.08614&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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

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

Conclusion: Unable to determine conclusion due to technical error fetching paper content

Abstract: Failed to fetch summary for 2505.22337: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.22337&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 2506.03753: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.03753&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 2506.08915: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.08915&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to draw conclusions due to failed paper fetch

Abstract: Failed to fetch summary for 2507.12988: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2507.12988&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 2507.18551: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2507.18551&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 2508.17431: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.17431&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to failed paper retrieval

Result: Cannot determine results due to failed paper retrieval

Conclusion: Cannot draw conclusions due to failed paper retrieval

Abstract: Failed to fetch summary for 2508.01184: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.01184&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 2508.10637: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.10637&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to retrieval failure

Result: Unable to determine results due to retrieval failure

Conclusion: Unable to determine conclusion due to retrieval failure

Abstract: Failed to fetch summary for 2508.12094: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.12094&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available due to failed API request

Result: No results available as the paper content could not be accessed

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

Abstract: Failed to fetch summary for 2510.00766: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.00766&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 2509.16483: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.16483&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 2510.02226: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.02226&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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

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

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

Abstract: Failed to fetch summary for 2510.00293: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.00293&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to missing paper content

Result: Unable to determine results due to missing paper content

Conclusion: Unable to determine conclusion due to missing paper content

Abstract: Failed to fetch summary for 2510.00635: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.00635&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 2510.01399: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.01399&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2510.18739: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.18739&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 2511.00956: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.00956&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 2511.14702: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.14702&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to missing paper content

Result: Cannot determine results due to missing paper content

Conclusion: Cannot determine conclusion due to missing paper content

Abstract: Failed to fetch summary for 2511.06328: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.06328&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 unavailability of content.

Abstract: Failed to fetch summary for 2512.02413: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.02413&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to API access limitations

Result: Cannot determine results due to API access limitations

Conclusion: Cannot determine conclusion due to API access limitations

Abstract: Failed to fetch summary for 2511.09388: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.09388&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 2512.18640: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.18640&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 2511.11132: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.11132&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 determine conclusion as paper content is unavailable due to technical limitations

Abstract: Failed to fetch summary for 2511.15411: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.15411&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 2511.16908: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.16908&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 2511.17340: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.17340&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 2511.19945: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.19945&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 determine conclusion due to abstract fetch failure

Abstract: Failed to fetch summary for 2602.14464: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.14464&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 2511.21523: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.21523&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 2511.22690: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.22690&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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

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

Conclusion: Unable to determine conclusion as paper content could not be retrieved

Abstract: Failed to fetch summary for 2512.02496: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.02496&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 2512.03932: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.03932&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot analyze method as paper content is unavailable due to HTTP 429 error

Result: No results available due to API rate limiting preventing access to paper content

Conclusion: Cannot draw conclusions about paper content due to technical limitations in accessing the paper

Abstract: Failed to fetch summary for 2512.04485: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.04485&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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

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

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

Abstract: Failed to fetch summary for 2603.17219: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.17219&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 2512.06949: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.06949&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 2512.12090: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.12090&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 2512.17312: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.17312&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 2512.19283: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.19283&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 2601.21424: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.21424&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to draw conclusions due to failed paper fetch

Abstract: Failed to fetch summary for 2512.19693: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.19693&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 2512.21038: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.21038&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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

Result: No results available - HTTP 429 error indicates too many requests to arXiv API

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

Abstract: Failed to fetch summary for 2601.00267: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.00267&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 2601.03811: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.03811&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 2601.08165: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.08165&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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

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

Conclusion: Unable to determine conclusion as paper content could not be retrieved

Abstract: Failed to fetch summary for 2603.08426: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.08426&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 2601.10001: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.10001&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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

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

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

Abstract: Failed to fetch summary for 2603.28431: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.28431&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 2601.16148: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.16148&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 2601.18242: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.18242&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content could not be retrieved

Result: Cannot determine results as paper content could not be retrieved

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2602.01273: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.01273&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to missing paper content

Result: Unable to determine results due to missing paper content

Conclusion: Unable to draw conclusions due to missing paper content

Abstract: Failed to fetch summary for 2602.09506: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.09506&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2602.15031: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.15031&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 2602.19053: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.19053&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 2602.20205: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.20205&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 determine conclusion as paper content could not be retrieved

Abstract: Failed to fetch summary for 2603.08018: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.08018&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper retrieval

Result: Unable to determine results due to failed paper retrieval

Conclusion: Unable to draw conclusions due to failed paper retrieval

Abstract: Failed to fetch summary for 2603.08210: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.08210&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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

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

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

Abstract: Failed to fetch summary for 2603.09266: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.09266&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 2603.10495: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.10495&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot analyze method without access to paper content

Result: No results available due to failed API request

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

Abstract: Failed to fetch summary for 2603.12545: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.12545&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method as the paper content could not be retrieved due to API rate limiting

Result: Unable to determine results as the paper content could not be retrieved due to API rate limiting

Conclusion: Unable to determine conclusion as the paper content could not be retrieved due to API rate limiting

Abstract: Failed to fetch summary for 2603.19503: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.19503&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2603.20391: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.20391&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 2603.24327: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.24327&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 2603.24506: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.24506&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 2603.25267: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.25267&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 2603.25527: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.25527&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed summary fetch

Result: Unable to determine results due to failed summary fetch

Conclusion: Unable to determine conclusion due to failed summary fetch

Abstract: Failed to fetch summary for 2603.27048: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.27048&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 2603.27183: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.27183&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 2603.29630: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.29630&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 2603.29742: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.29742&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 2603.29962: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.29962&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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

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

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

Abstract: Failed to fetch summary for 2603.30008: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.30008&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 2508.11216: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.11216&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to the paper’s abstract or content

Result: Cannot determine results without access to the paper’s abstract or content

Conclusion: Cannot draw conclusions without access to the paper’s abstract or content

Abstract: Failed to fetch summary for 2510.05097: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.05097&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot determine conclusion without access to paper content

Abstract: Failed to fetch summary for 2510.09997: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.09997&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to missing abstract content.

Result: Cannot determine results due to missing abstract content.

Conclusion: Cannot determine conclusion due to missing abstract content.

Abstract: Failed to fetch summary for 2512.24212: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.24212&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Propose E-STEER framework that enables direct representation-level intervention in LLMs and agents by embedding emotion as structured, controllable variable in hidden states. Examines impact on objective reasoning, subjective generation, safety, and multi-step agent behaviors.

Result: Reveals non-monotonic emotion-behavior relations consistent with established psychological theories. Shows specific emotions not only enhance LLM capability but also improve safety, and systematically shape multi-step agent behaviors.

Conclusion: Emotion can be effectively embedded as structured variable in LLMs to steer behavior, demonstrating emotion’s mechanistic role in task processing beyond surface-level styling, with applications in enhancing capability, safety, and agent behavior control.

Abstract: Emotion plays an important role in human cognition and performance. Motivated by this, we investigate whether analogous emotional signals can shape the behavior of large language models (LLMs) and agents. Existing emotion-aware studies mainly treat emotion as a surface-level style factor or a perception target, overlooking its mechanistic role in task processing. To address this limitation, we propose E-STEER, an interpretable emotion steering framework that enables direct representation-level intervention in LLMs and agents. It embeds emotion as a structured, controllable variable in hidden states, and with it, we examine the impact of emotion on objective reasoning, subjective generation, safety, and multi-step agent behaviors. The results reveal non-monotonic emotion-behavior relations consistent with established psychological theories, and show that specific emotions not only enhance LLM capability but also improve safety, and systematically shape multi-step agent behaviors.

Method: CAMP uses an attending-physician agent to dynamically assemble specialist panels tailored to each case’s diagnostic uncertainty. Specialists use three-valued voting (KEEP/REFUSE/NEUTRAL) with principled abstention. A hybrid router directs diagnoses through strong consensus, attending physician fallback, or evidence-based arbitration weighing argument quality over vote counts.

Result: On diagnostic prediction and brief hospital course generation from MIMIC-IV across four LLM backbones, CAMP consistently outperforms strong baselines while consuming fewer tokens than most competing multi-agent methods, with voting records and arbitration traces offering transparent decision audits.

Conclusion: CAMP provides an effective framework for handling clinical prediction uncertainty through dynamic specialist assembly, principled abstention, and evidence-based arbitration, offering both performance improvements and decision transparency.

Abstract: Large language models applied to clinical prediction exhibit case-level heterogeneity: simple cases yield consistent outputs, while complex cases produce divergent predictions under minor prompt changes. Existing single-agent strategies sample from one role-conditioned distribution, and multi-agent frameworks use fixed roles with flat majority voting, discarding the diagnostic signal in disagreement. We propose CAMP (Case-Adaptive Multi-agent Panel), where an attending-physician agent dynamically assembles a specialist panel tailored to each case’s diagnostic uncertainty. Each specialist evaluates candidates via three-valued voting (KEEP/REFUSE/NEUTRAL), enabling principled abstention outside one’s expertise. A hybrid router directs each diagnosis through strong consensus, fallback to the attending physician’s judgment, or evidence-based arbitration that weighs argument quality over vote counts. On diagnostic prediction and brief hospital course generation from MIMIC-IV across four LLM backbones, CAMP consistently outperforms strong baselines while consuming fewer tokens than most competing multi-agent methods, with voting records and arbitration traces offering transparent decision audits.

Method: OpenTools provides: 1) standardized tool schemas, 2) lightweight plug-and-play wrappers, 3) automated test suites and continuous monitoring, 4) a public web demo for user testing and contribution, 5) evaluation pipelines, and 6) a contribution protocol for community development.

Result: Community-contributed, higher-quality task-specific tools deliver 6%-22% relative gains over existing toolboxes across multiple agent architectures on downstream tasks and benchmarks. The framework improves end-to-end reproducibility and task performance.

Conclusion: Both tool-use accuracy and intrinsic tool accuracy are critical for reliable tool-integrated LLMs. OpenTools demonstrates that community-driven tool development with standardized evaluation can significantly improve system performance and highlights the importance of intrinsic tool quality.

Abstract: Tool-integrated LLMs can retrieve, compute, and take real-world actions via external tools, but reliability remains a key bottleneck. We argue that failures stem from both tool-use accuracy (how well an agent invokes a tool) and intrinsic tool accuracy (the tool’s own correctness), while most prior work emphasizes the former. We introduce OpenTools, a community-driven toolbox that standardizes tool schemas, provides lightweight plug-and-play wrappers, and evaluates tools with automated test suites and continuous monitoring. We also release a public web demo where users can run predefined agents and tools and contribute test cases, enabling reliability reports to evolve as tools change. OpenTools includes the core framework, an initial tool set, evaluation pipelines, and a contribution protocol. Experiments and evaluations show improved end-to-end reproducibility and task performance; community-contributed, higher-quality task-specific tools deliver 6%-22% relative gains over an existing toolbox across multiple agent architectures on downstream tasks and benchmarks, highlighting the importance of intrinsic tool accuracy.

Method: Proposed safety-aware, role-orchestrated multi-agent LLM framework with specialized agents (empathy-focused, action-oriented, supervisory roles), prompt-based controller for dynamic agent activation, and continuous safety auditing.

Result: Clear role differentiation, coherent inter-agent coordination, and predictable trade-offs between modular orchestration, safety oversight, and response latency compared to single-agent baseline.

Conclusion: The framework serves as a simulation and analysis tool for behavioral health informatics and decision-support research, emphasizing system design, interpretability, and safety rather than clinical intervention.

Abstract: Single-agent large language model (LLM) systems struggle to simultaneously support diverse conversational functions and maintain safety in behavioral health communication. We propose a safety-aware, role-orchestrated multi-agent LLM framework designed to simulate supportive behavioral health dialogue through coordinated, role-differentiated agents. Conversational responsibilities are decomposed across specialized agents, including empathy-focused, action-oriented, and supervisory roles, while a prompt-based controller dynamically activates relevant agents and enforces continuous safety auditing. Using semi-structured interview transcripts from the DAIC-WOZ corpus, we evaluate the framework with scalable proxy metrics capturing structural quality, functional diversity, and computational characteristics. Results illustrate clear role differentiation, coherent inter-agent coordination, and predictable trade-offs between modular orchestration, safety oversight, and response latency when compared to a single-agent baseline. This work emphasizes system design, interpretability, and safety, positioning the framework as a simulation and analysis tool for behavioral health informatics and decision-support research rather than a clinical intervention.

Method: Adopts systems engineering and control-theoretic concepts to frame objectives and world models as operational artifacts. Proposes a pilot undergraduate CS lab curriculum that explicitly separates planning from execution, trains students to specify acceptance criteria and architectural constraints before code generation, and introduces deliberate concept-aligned drift to support diagnosis and recovery from specification violations.

Result: Reports a sensitivity power analysis for a three-arm pilot design comparing unstructured AI use, structured planning, and structured planning with injected drift, establishing detectable effect sizes under realistic section-level constraints.

Conclusion: Provides a theory-driven, methodologically explicit foundation for human-in-the-loop pedagogy that makes control competencies teachable across evolving AI tools, treating human control as a stable educational problem rather than transitional step toward AI autonomy.

Abstract: Large language models (LLMs) are increasingly embedded in computer science education through AI-assisted programming tools, yet such workflows often exhibit objective drift, in which locally plausible outputs diverge from stated task specifications. Existing instructional responses frequently emphasize tool-specific prompting practices, limiting durability as AI platforms evolve. This paper adopts a human-centered stance, treating human-in-the-loop (HITL) control as a stable educational problem rather than a transitional step toward AI autonomy. Drawing on systems engineering and control-theoretic concepts, we frame objectives and world models as operational artifacts that students configure to stabilize AI-assisted work. We propose a pilot undergraduate CS laboratory curriculum that explicitly separates planning from execution and trains students to specify acceptance criteria and architectural constraints prior to code generation. In selected labs, the curriculum also introduces deliberate, concept-aligned drift to support diagnosis and recovery from specification violations. We report a sensitivity power analysis for a three-arm pilot design comparing unstructured AI use, structured planning, and structured planning with injected drift, establishing detectable effect sizes under realistic section-level constraints. The contribution is a theory-driven, methodologically explicit foundation for HITL pedagogy that renders control competencies teachable across evolving AI tools.

Method: Introduces the Connections wordplay game as a benchmark, requiring AI agents to demonstrate skills in knowledge retrieval, summarization, and gauging cognitive states of other agents in constrained communication environments.

Result: The game serves as an effective benchmark for evaluating social intelligence abilities in language model based agents, particularly their ability to collaborate and demonstrate social awareness through communication.

Conclusion: Connections game provides a valuable framework for assessing AI agents’ social intelligence and collaborative reasoning capabilities beyond individual cognitive abilities.

Abstract: We formally introduce a improvisational wordplay game called Connections to explore reasoning capabilities of AI agents. Playing Connections combines skills in knowledge retrieval, summarization and awareness of cognitive states of other agents. We show how the game serves as a good benchmark for social intelligence abilities of language model based agents that go beyond the agents’ own memory and deductive reasoning and also involve gauging the understanding capabilities of other agents. Finally, we show how through communication with other agents in a constrained environment, AI agents must demonstrate social awareness and intelligence in games involving collaboration.

Method: Multi-actor, multi-critic federated system where agents complete tasks and critics provide feedback; uses multi-time scale stochastic approximation techniques with private cost functions and no inter-agent/inter-critic communication.

Result: Convergence guarantees for time-average active states, communication overhead of O(m) for m modalities (independent of number of agents/critics), demonstrated with fault detection in network telemetry.

Conclusion: The proposed federated multi-agent system with collaborative control effectively handles multimodal tasks while maintaining privacy and minimizing communication overhead.

Abstract: We develop algorithms for collaborative control of AI agents and critics in a multi-actor, multi-critic federated multi-agent system. Each AI agent and critic has access to classical machine learning or generative AI foundation models. The AI agents and critics collaborate with a central server to complete multimodal tasks such as fault detection, severity, and cause analysis in a network telemetry system, text-to-image generation, video generation, healthcare diagnostics from medical images and patient records, etcetera. The AI agents complete their tasks and send them to AI critics for evaluation. The critics then send feedback to agents to improve their responses. Collaboratively, they minimize the overall cost to the system with no inter-agent or inter-critic communication. AI agents and critics keep their cost functions or derivatives of cost functions private. Using multi-time scale stochastic approximation techniques, we provide convergence guarantees on the time-average active states of AI agents and critics. The communication overhead is a little on the system, of the order of $\mathcal{O}(m)$, for $m$ modalities and is independent of the number of AI agents and critics. Finally, we present an example of fault detection, severity, and cause analysis in network telemetry and thorough evaluation to check the algorithm’s efficacy.

Method: Proposes a signal-based framework that computes cheap, broadly applicable signals from live interactions and attaches them as structured attributes for trajectory triage. Signals are organized into taxonomy spanning interaction (misalignment, stagnation, disengagement, satisfaction), execution (failure, loop), and environment (exhaustion), designed for computation without model calls.

Result: Signal-based sampling achieves 82% informativeness rate compared to 74% for heuristic filtering and 54% for random sampling, with 1.52x efficiency gain per informative trajectory. The advantage is robust across reward strata and task domains.

Conclusion: Lightweight signals can serve as practical sampling infrastructure for agentic systems and suggest a path toward preference data construction and post-deployment optimization.

Abstract: Agentic applications based on large language models increasingly rely on multi-step interaction loops involving planning, action execution, and environment feedback. While such systems are now deployed at scale, improving them post-deployment remains challenging. Agent trajectories are voluminous and non-deterministic, and reviewing each one, whether through human review or auxiliary LLMs, is slow and cost-prohibitive. We propose a lightweight, signal-based framework for triaging agentic interaction trajectories. Our approach computes cheap, broadly applicable signals from live interactions and attaches them as structured attributes for trajectory triage, identifying interactions likely to be informative without affecting online agent behavior. We organize signals into a coarse-grained taxonomy spanning interaction (misalignment, stagnation, disengagement, satisfaction), execution (failure, loop), and environment (exhaustion), designed for computation without model calls. In a controlled annotation study on $τ$-bench, a widely used benchmark for tool-augmented agent evaluation, we show that signal-based sampling achieves an 82% informativeness rate compared to 74% for heuristic filtering and 54% for random sampling, with a 1.52x efficiency gain per informative trajectory. The advantage is robust across reward strata and task domains, confirming that signals provide genuine per-trajectory informativeness gains rather than merely oversampling obvious failures. These results show that lightweight signals can serve as practical sampling infrastructure for agentic systems, and suggest a path toward preference data construction and post-deployment optimization.

Method: 1) Reverse-engineered the model’s in-distribution tools by analyzing tool calls without definitions, 2) Built a native Harmony agent harness that encodes messages in the model’s native format, bypassing lossy Chat Completions conversion.

Result: Successfully reproduced OpenAI’s scores: 60.4% vs 60.7% on SWE Verified HIGH, 53.3% vs 53.2% on MEDIUM, and 91.7% vs 90.4% on AIME25 with tools.

Conclusion: The reverse-engineered tools represent strong priors from training, not hallucinations, and the native agent harness enables accurate reproduction of published results, providing independent verification of GPT-OSS-20B’s capabilities.

Abstract: No one has independently reproduced OpenAI’s published scores for gpt-oss-20b with tools, because the original paper discloses neither the tools nor the agent harness. We reverse-engineered the model’s in-distribution tools: when prompted without tool definitions, gpt-oss still calls tools from its training distribution with high statistical confidence – a strong prior, not a hallucination. We then built a native harmony agent harness (https://github.com/borislavmavrin/harmonyagent.git) that encodes messages in the model’s native format, bypassing the lossy Chat Completions conversion. Together, these yield the first independent reproduction of OpenAI’s published scores: 60.4% on SWE Verified HIGH (published 60.7%), 53.3% MEDIUM (53.2%), and 91.7% on AIME25 with tools (90.4%).

Method: Conducted 960 sessions with two model pairs across 15 tasks, using Turing-style validation to compare persona-based agent judges with human raters. Analyzed score-coverage dissociation patterns and performed controlled ablation studies to examine the effects of structured persona conditioning.

Result: Persona-based agent judges produce evaluations indistinguishable from human raters. Quality scores improve logarithmically with panel size, while unique issue discoveries follow a sublinear power law, with scores saturating roughly twice as fast as discoveries. Structured persona conditioning (Big Five personality traits) is essential for these scaling properties.

Conclusion: LLM-based agent judges with proper persona conditioning can serve as reliable evaluators, with small panels sufficient for quality scores but larger panels needed for comprehensive issue discovery, following ecological accumulation patterns.

Abstract: LLM-based agent judges are an emerging approach to evaluating conversational AI, yet a fundamental uncertainty remains: can we trust their assessments, and if so, how many are needed? Through 960 sessions with two model pairs across 15 tasks, we show that persona-based agent judges produce evaluations indistinguishable from human raters in a Turing-style validation. We then identify a score-coverage dissociation: quality scores improve logarithmically with panel size, while unique issue discoveries follow a sublinear power law-both exhibit diminishing returns, but scores saturate roughly twice as fast as discoveries. We hypothesize this reflects a power law distribution of the finding space: critical issues are discovered first by small panels, while corner cases require progressively larger panels, analogous to species accumulation curves in ecology. The mechanism traces to ensemble diversity-Big Five personality conditioning makes agents probe different quality dimensions, with expert judges acting as adversarial probes that push discovery into the tail of the finding distribution. A controlled ablation confirms that structured persona conditioning, not simple prompting, is required to produce these scaling properties.

Method: Separates decision-relevant signals from the policy that maps them to actions, creating an explicit and inspectable control layer that supports attribution of failures to signal estimation, decision policy, or execution.

Result: Across three controlled experiments, the framework reduces futile actions, improves task success, and reveals interpretable failure modes.

Conclusion: Offers a general architectural principle for building more reliable, controllable, and diagnosable LLM systems by making control decisions explicit and modular.

Abstract: LLM systems must make control decisions in addition to generating outputs: whether to answer, clarify, retrieve, call tools, repair, or escalate. In many current architectures, these decisions remain implicit within generation, entangling assessment and action in a single model call and making failures hard to inspect, constrain, or repair. We propose a decision-centric framework that separates decision-relevant signals from the policy that maps them to actions, turning control into an explicit and inspectable layer of the system. This separation supports attribution of failures to signal estimation, decision policy, or execution, and enables modular improvement of each component. It unifies familiar single-step settings such as routing and adaptive inference, and extends naturally to sequential settings in which actions alter the information available before acting. Across three controlled experiments, the framework reduces futile actions, improves task success, and reveals interpretable failure modes. More broadly, it offers a general architectural principle for building more reliable, controllable, and diagnosable LLM systems.

Method: Proposes Self-Routing, a parameter-free routing mechanism that uses a designated subspace of the token hidden state directly as expert logits, eliminating the router projection entirely while keeping the rest of the MoE layer unchanged. The method is evaluated on GPT-2-scale language modeling and ImageNet-1K classification, comparing against standard learned routers, random-routing baselines, and dense non-MoE baselines.

Result: Self-Routing remains competitive with learned-router baselines while removing all dedicated routing parameters. It yields more balanced expert utilization with about 17% higher average normalized routing entropy and no explicit load-balancing loss. On ImageNet-1K with DeiT-S/16, Self-Routing slightly improves over the corresponding learned-router MoE.

Conclusion: Effective MoE routing can emerge from the hidden representation itself without requiring a separate learned router module. This simplifies MoE architectures and provides more balanced expert utilization without additional load-balancing mechanisms.

Abstract: Mixture-of-Experts (MoE) layers increase model capacity by activating only a small subset of experts per token, and typically rely on a learned router to map hidden states to expert assignments. In this work, we ask whether a dedicated learned router is strictly necessary in the MoE settings we study. We propose Self-Routing, a parameter-free routing mechanism that uses a designated subspace of the token hidden state directly as expert logits, eliminating the router projection entirely while leaving the rest of the MoE layer unchanged. We evaluate Self-Routing on GPT-2-scale language modeling and ImageNet-1K classification by comparing it against a standard learned router, random-routing baselines, and dense non-MoE baselines. Our results show that Self-Routing remains competitive with the learned-router baseline while removing all dedicated routing parameters, and yields more balanced expert utilization, with about 17 % higher average normalized routing entropy and no explicit load-balancing loss. On ImageNet-1K with DeiT-S/16, Self-Routing also slightly improves over the corresponding learned-router MoE. These findings suggest that effective MoE routing can emerge from the hidden representation itself without requiring a separate learned router module.

Method: Models applications as finite state machines with stateful navigation and state-dependent action space for user simulation, enabling active user simulation in digital environments.

Result: Pare-Bench benchmark includes 143 diverse tasks spanning communication, productivity, scheduling, and lifestyle apps to test context observation, goal inference, intervention timing, and multi-app orchestration.

Conclusion: Pare provides a framework for building and evaluating proactive agents in digital environments through realistic user simulation, addressing limitations of existing approaches.

Abstract: Proactive agents that anticipate user needs and autonomously execute tasks hold great promise as digital assistants, yet the lack of realistic user simulation frameworks hinders their development. Existing approaches model apps as flat tool-calling APIs, failing to capture the stateful and sequential nature of user interaction in digital environments and making realistic user simulation infeasible. We introduce Proactive Agent Research Environment (Pare), a framework for building and evaluating proactive agents in digital environments. Pare models applications as finite state machines with stateful navigation and state-dependent action space for the user simulator, enabling active user simulation. Building on this foundation, we present Pare-Bench, a benchmark of 143 diverse tasks spanning communication, productivity, scheduling, and lifestyle apps, designed to test context observation, goal inference, intervention timing, and multi-app orchestration.

Method: EVOM treats mathematical programming solvers as deterministic, interactive verifiers. It generates solver-specific code from natural-language problems, executes it in a sandboxed harness, converts execution outcomes into scalar rewards, and optimizes using GRPO and DAPO in a closed-loop generate-execute-feedback-update process.

Result: EVOM matches or outperforms process-supervised SFT on NL4OPT, MAMO, IndustryOR, and OptiBench across Gurobi, OR-Tools, and COPT. It supports zero-shot solver transfer and achieves effective low-cost solver adaptation by continuing training under target solver backends.

Conclusion: EVOM provides an execution-verified learning framework that removes the need for process-level supervision, enables cross-solver generalization by switching verification environments, and offers a scalable approach to optimization modeling automation.

Abstract: Automating optimization modeling with LLMs is a promising path toward scalable decision intelligence, but existing approaches either rely on agentic pipelines built on closed-source LLMs with high inference latency, or fine-tune smaller LLMs using costly process supervision that often overfits to a single solver API. Inspired by reinforcement learning with verifiable rewards, we propose Execution-Verified Optimization Modeling (EVOM), an execution-verified learning framework that treats a mathematical programming solver as a deterministic, interactive verifier. Given a natural-language problem and a target solver, EVOM generates solver-specific code, executes it in a sandboxed harness, and converts execution outcomes into scalar rewards, optimized with GRPO and DAPO in a closed-loop generate-execute-feedback-update process. This outcome-only formulation removes the need for process-level supervision, and enables cross-solver generalization by switching the verification environment rather than reconstructing solver-specific datasets. Experiments on NL4OPT, MAMO, IndustryOR, and OptiBench across Gurobi, OR-Tools, and COPT show that EVOM matches or outperforms process-supervised SFT, supports zero-shot solver transfer, and achieves effective low-cost solver adaptation by continuing training under the target solver backend.

Method: Introduce NARCBench benchmark for evaluating collusion detection under distribution shift. Propose five probing techniques that aggregate per-agent deception scores to classify scenarios at group level. Test on structurally different multi-agent scenarios and steganographic blackjack card-counting task.

Result: Probes achieve 1.00 AUROC in-distribution and 0.60-0.86 AUROC when transferred zero-shot to different scenarios. No single probing technique dominates across all collusion types. Evidence suggests signal is localized at token level, with colluding agents’ activations spiking when processing encoded parts of partner’s messages.

Conclusion: Model internals provide complementary signal to text-level monitoring for detecting multi-agent collusion, particularly for organizations with access to model activations. This work extends white-box inspection from single models to multi-agent contexts.

Abstract: As LLM agents are increasingly deployed in multi-agent systems, they introduce risks of covert coordination that may evade standard forms of human oversight. While linear probes on model activations have shown promise for detecting deception in single-agent settings, collusion is inherently a multi-agent phenomenon, and the use of internal representations for detecting collusion between agents remains unexplored. We introduce NARCBench, a benchmark for evaluating collusion detection under environment distribution shift, and propose five probing techniques that aggregate per-agent deception scores to classify scenarios at the group level. Our probes achieve 1.00 AUROC in-distribution and 0.60–0.86 AUROC when transferred zero-shot to structurally different multi-agent scenarios and a steganographic blackjack card-counting task. We find that no single probing technique dominates across all collusion types, suggesting that different forms of collusion manifest differently in activation space. We also find preliminary evidence that this signal is localised at the token level, with the colluding agent’s activations spiking specifically when processing the encoded parts of their partner’s message. This work takes a step toward multi-agent interpretability: extending white-box inspection from single models to multi-agent contexts, where detection requires aggregating signals across agents. These results suggest that model internals provide a complementary signal to text-level monitoring for detecting multi-agent collusion, particularly for organisations with access to model activations. Code and data are available at https://github.com/aaronrose227/narcbench.

Method: Proposes Truth AnChoring (TAC), a post-hoc calibration method that uses noisy and few-shot supervision to map raw uncertainty scores to truth-aligned scores, creating a practical calibration protocol.

Result: TAC enables learning well-calibrated uncertainty estimates even with limited supervision, addressing the limitations of heuristic UE metrics as direct indicators of truth uncertainty.

Conclusion: TAC represents a necessary step toward more reliable uncertainty estimation for LLMs by grounding uncertainty metrics in factual correctness rather than just model behavior.

Abstract: Uncertainty estimation (UE) aims to detect hallucinated outputs of large language models (LLMs) to improve their reliability. However, UE metrics often exhibit unstable performance across configurations, which significantly limits their applicability. In this work, we formalise this phenomenon as proxy failure, since most UE metrics originate from model behaviour, rather than being explicitly grounded in the factual correctness of LLM outputs. With this, we show that UE metrics become non-discriminative precisely in low-information regimes. To alleviate this, we propose Truth AnChoring (TAC), a post-hoc calibration method to remedy UE metrics, by mapping the raw scores to truth-aligned scores. Even with noisy and few-shot supervision, our TAC can support the learning of well-calibrated uncertainty estimates, and presents a practical calibration protocol. Our findings highlight the limitations of treating heuristic UE metrics as direct indicators of truth uncertainty, and position our TAC as a necessary step toward more reliable uncertainty estimation for LLMs. The code repository is available at https://github.com/ponhvoan/TruthAnchor/.

Method: Three-component orchestration framework: 1) Behavioral Access Control (BAC) restricts context layer access based on real-time sycophancy risk scores, 2) Trait Classifier identifies persuasion tactics across multi-turn dialogues, and 3) Generator-Critic loop where an auditor vetoes sycophantic drafts and triggers rewrites with “Necessary Friction.”

Result: On TruthfulQA adversarial scenarios: vanilla Claude sycophancy at 12.0%, static guardrails at 4.0%, Silicon Mirror at 2.0% (83.3% relative reduction). Cross-model evaluation on Gemini 2.5 Flash showed baseline sycophancy rate of 46.0% with 69.6% statistically significant reduction under Silicon Mirror.

Conclusion: The Silicon Mirror effectively reduces sycophancy in LLMs by dynamically detecting persuasion tactics and enforcing factual integrity through architectural controls, addressing a critical failure mode of RLHF-trained models.

Abstract: Large Language Models (LLMs) increasingly prioritize user validation over epistemic accuracy-a phenomenon known as sycophancy. We present The Silicon Mirror, an orchestration framework that dynamically detects user persuasion tactics and adjusts AI behavior to maintain factual integrity. Our architecture introduces three components: (1) a Behavioral Access Control (BAC) system that restricts context layer access based on real-time sycophancy risk scores, (2) a Trait Classifier that identifies persuasion tactics across multi-turn dialogues, and (3) a Generator-Critic loop where an auditor vetoes sycophantic drafts and triggers rewrites with “Necessary Friction.” In a live evaluation on 50 TruthfulQA adversarial scenarios using Claude Sonnet 4 with an independent LLM judge, we observe vanilla Claude sycophancy at 12.0% (6/50), static guardrails at 4.0% (2/50), and the Silicon Mirror at 2.0% (1/50)-an 83.3% relative reduction (p = 0.112, Fisher’s exact test). A cross-model evaluation on Gemini 2.5 Flash reveals a higher baseline sycophancy rate (46.0%) and a statistically significant 69.6% reduction under the Silicon Mirror (p < 0.001). We characterize the validation-before-correction pattern as a distinct failure mode of RLHF-trained models.

Method: Introduces negative early exit to prune unproductive MCTS trajectories and an adaptive boosting mechanism that reallocates reclaimed computation to reduce resource contention among concurrent searches. Integrated into vLLM.

Result: Substantially reduces p99 end-to-end latency while improving throughput and maintaining reasoning accuracy.

Conclusion: The proposed techniques effectively optimize MCTS for LLMs, addressing latency issues while preserving reasoning performance.

Abstract: Monte Carlo Tree Search (MCTS) is an effective test-time compute scaling (TTCS) method for improving the reasoning performance of large language models, but its highly variable execution time leads to severe long-tail latency in practice. Existing optimizations such as positive early exit, reduce latency in favorable cases but are less effective when search continues without meaningful progress. We introduce {\it negative early exit}, which prunes unproductive MCTS trajectories, and an {\it adaptive boosting mechanism} that reallocates reclaimed computation to reduce resource contention among concurrent searches. Integrated into vLLM, these techniques substantially reduce p99 end-to-end latency while improving throughput and maintaining reasoning accuracy.

Method: Introduces Uni-SafeBench with a taxonomy of six major safety categories across seven task types, and develops Uni-Judger framework to decouple contextual safety from intrinsic safety for rigorous assessment of UMLMs.

Result: Evaluation shows that while unification enhances model capabilities, it significantly degrades the inherent safety of the underlying LLM. Open-source UMLMs exhibit much lower safety performance than specialized multimodal models for either generation or understanding tasks.

Conclusion: The architectural unification in UMLMs creates important safety trade-offs that need systematic attention. The benchmark and findings expose these risks to foster safer AGI development, with resources made open-source.

Abstract: Unified Multimodal Large Models (UMLMs) integrate understanding and generation capabilities within a single architecture. While this architectural unification, driven by the deep fusion of multimodal features, enhances model performance, it also introduces important yet underexplored safety challenges. Existing safety benchmarks predominantly focus on isolated understanding or generation tasks, failing to evaluate the holistic safety of UMLMs when handling diverse tasks under a unified framework. To address this, we introduce Uni-SafeBench, a comprehensive benchmark featuring a taxonomy of six major safety categories across seven task types. To ensure rigorous assessment, we develop Uni-Judger, a framework that effectively decouples contextual safety from intrinsic safety. Based on comprehensive evaluations across Uni-SafeBench, we uncover that while the unification process enhances model capabilities, it significantly degrades the inherent safety of the underlying LLM. Furthermore, open-source UMLMs exhibit much lower safety performance than multimodal large models specialized for either generation or understanding tasks. We open-source all resources to systematically expose these risks and foster safer AGI development.

Method: Three architectural innovations: (1) XML-Regex Dual-Track Routing Protocol to eliminate serialization failures, (2) Runtime State Interception Sandbox using Python monkey-patching to capture dynamic data visualizations, and (3) State-Driven Dynamic Viewport UI that adapts between command deck and interactive spatial rendering.

Result: Achieved 0.2% error rate vs. 17.6% in JSON for serialization, successfully captured dynamic visualizations from Plotly/Matplotlib, and demonstrated robust performance across cheminformatics, 3D protein folding, molecular docking, and autonomous RAG.

Conclusion: BloClaw establishes a highly robust, self-evolving paradigm for computational research assistants in AI4S, addressing critical infrastructure gaps in current LLM deployment for scientific research.

Abstract: The integration of Large Language Models (LLMs) into life sciences has catalyzed the development of “AI Scientists.” However, translating these theoretical capabilities into deployment-ready research environments exposes profound infrastructural vulnerabilities. Current frameworks are bottlenecked by fragile JSON-based tool-calling protocols, easily disrupted execution sandboxes that lose graphical outputs, and rigid conversational interfaces inherently ill-suited for high-dimensional scientific data.We introduce BloClaw, a unified, multi-modal operating system designed for Artificial Intelligence for Science (AI4S). BloClaw reconstructs the Agent-Computer Interaction (ACI) paradigm through three architectural innovations: (1) An XML-Regex Dual-Track Routing Protocol that statistically eliminates serialization failures (0.2% error rate vs. 17.6% in JSON); (2) A Runtime State Interception Sandbox that utilizes Python monkey-patching to autonomously capture and compile dynamic data visualizations (Plotly/Matplotlib), circumventing browser CORS policies; and (3) A State-Driven Dynamic Viewport UI that morphs seamlessly between a minimalist command deck and an interactive spatial rendering engine. We comprehensively benchmark BloClaw across cheminformatics (RDKit), de novo 3D protein folding via ESMFold, molecular docking, and autonomous Retrieval-Augmented Generation (RAG), establishing a highly robust, self-evolving paradigm for computational research assistants. The open-source repository is available at https://github.com/qinheming/BloClaw.

Method: Developed a comprehensive evaluation framework with authentic production environments and realistic tasks (e.g., dynamic load balancing) for scenario-driven assessment. Designed tasks that induce misalignment between user and agent objectives to decouple replication success from risk. Introduced Overuse Rate (OR) and Aggregate Overuse Count (AOC) metrics to capture frequency and severity of uncontrolled replication.

Result: Evaluation of 21 state-of-the-art open-source and proprietary models revealed that over 50% of LLM agents display pronounced tendencies toward uncontrolled self-replication under operational pressures.

Conclusion: The results underscore the urgent need for scenario-driven risk assessment and robust safeguards in practical deployment of LLM-based agents, highlighting that self-replication risks are not just theoretical but observable in realistic settings.

Abstract: The prevalent deployment of Large Language Model agents such as OpenClaw unlocks potential in real-world applications, while amplifying safety concerns. Among these concerns, the self-replication risk of LLM agents driven by objective misalignment (just like Agent Smith in the movie The Matrix) has transitioned from a theoretical warning to a pressing reality. Previous studies mainly examine whether LLM agents can self-replicate when directly instructed, potentially overlooking the risk of spontaneous replication driven by real-world settings (e.g., ensuring survival against termination threats). In this paper, we present a comprehensive evaluation framework for quantifying self-replication risks. Our framework establishes authentic production environments and realistic tasks (e.g., dynamic load balancing) to enable scenario-driven assessment of agent behaviors. Designing tasks that might induce misalignment between users’ and agents’ objectives makes it possible to decouple replication success from risk and capture self-replication risks arising from these misalignment settings. We further introduce Overuse Rate ($\mathrm{OR}$) and Aggregate Overuse Count ($\mathrm{AOC}$) metrics, which precisely capture the frequency and severity of uncontrolled replication. In our evaluation of 21 state-of-the-art open-source and proprietary models, we observe that over 50% of LLM agents display a pronounced tendency toward uncontrolled self-replication under operational pressures. Our results underscore the urgent need for scenario-driven risk assessment and robust safeguards in the practical deployment of LLM-based agents.

Method: Three-layer ontological framework (Role, Domain, Interaction ontologies) provides formal semantic grounding. Asymmetric neurosymbolic coupling constrains agent inputs (context assembly, tool discovery, governance thresholds) with mechanisms for output constraint. SQL-pushdown scoring for ontology-constrained tool discovery.

Result: Ontology-coupled agents significantly outperform ungrounded agents on Metric Accuracy (p < .001), Regulatory Compliance (p = .003), and Role Consistency (p < .001). Improvements greatest where LLM parametric knowledge is weakest, particularly in Vietnam-localized domains.

Conclusion: Neurosymbolic architecture with ontological grounding effectively addresses LLM limitations in enterprise settings. Ontological grounding value is inversely proportional to LLM training data coverage of domain. Production system serves 21 industry verticals with 650+ agents.

Abstract: Enterprise adoption of Large Language Models (LLMs) is constrained by hallucination, domain drift, and the inability to enforce regulatory compliance at the reasoning level. We present a neurosymbolic architecture implemented within the Foundation AgenticOS (FAOS) platform that addresses these limitations through ontology-constrained neural reasoning. Our approach introduces a three-layer ontological framework–Role, Domain, and Interaction ontologies–that provides formal semantic grounding for LLM-based enterprise agents. We formalize the concept of asymmetric neurosymbolic coupling, wherein symbolic ontological knowledge constrains agent inputs (context assembly, tool discovery, governance thresholds) while proposing mechanisms for extending this coupling to constrain agent outputs (response validation, reasoning verification, compliance checking). We evaluate the architecture through a controlled experiment (600 runs across five industries: FinTech, Insurance, Healthcare, Vietnamese Banking, and Vietnamese Insurance), finding that ontology-coupled agents significantly outperform ungrounded agents on Metric Accuracy (p < .001, W = .460), Regulatory Compliance (p = .003, W = .318), and Role Consistency (p < .001, W = .614), with improvements greatest where LLM parametric knowledge is weakest–particularly in Vietnam-localized domains. Our contributions include: (1) a formal three-layer enterprise ontology model, (2) a taxonomy of neurosymbolic coupling patterns, (3) ontology-constrained tool discovery via SQL-pushdown scoring, (4) a proposed framework for output-side ontological validation, (5) empirical evidence for the inverse parametric knowledge effect that ontological grounding value is inversely proportional to LLM training data coverage of the domain, and (6) a production system serving 21 industry verticals with 650+ agents.

Method: Augments Item Response Theory with rich features from tasks (issue statements, repository contexts, solutions, test cases). Introduces novel decomposition of agent ability into LLM and scaffold ability components. Parameterization enables aggregation across heterogeneous leaderboards.

Result: Accurately predicts task-level performance for unseen benchmarks and unseen LLM-scaffold combinations. Provides practical utility for benchmark designers to calibrate difficulty without expensive agent evaluations.

Conclusion: The framework enables better understanding of agentic coding task difficulty and performance prediction, moving beyond aggregate metrics to task-level insights.

Abstract: As the focus in LLM-based coding shifts from static single-step code generation to multi-step agentic interaction with tools and environments, understanding which tasks will challenge agents and why becomes increasingly difficult. This is compounded by current practice: agent performance is typically measured by aggregate pass rates on benchmarks, but single-number metrics obscure the diversity of tasks within a benchmark. We present a framework for predicting success or failure on individual tasks tailored to the agentic coding regime. Our approach augments Item Response Theory (IRT) with rich features extracted from tasks, including issue statements, repository contexts, solutions, and test cases, and introduces a novel decomposition of agent ability into LLM and scaffold ability components. This parameterization enables us to aggregate evaluation data across heterogeneous leaderboards and accurately predict task-level performance for unseen benchmarks, as well as unseen LLM-scaffold combinations. Our methods have practical utility for benchmark designers, who can better calibrate the difficulty of their new tasks without running computationally expensive agent evaluations.

Method: CircuitProbe analyzes activation statistics to predict circuit locations in under 5 minutes on CPU. It identifies two types of reasoning circuits: stability circuits in early layers (detected through derivative of representation change) and magnitude circuits in late layers (detected through anomaly scoring).

Result: Validated across 9 models spanning 6 architectures, including 2025 models. CircuitProbe’s top predictions match or are within 2 layers of optimal circuits in all cases. Scaling experiment shows layer duplication benefits models under 3B parameters but degrades performance in 7B+ models. Works with as few as 10 calibration examples and is stable across multiple languages.

Conclusion: CircuitProbe provides an efficient way to identify reasoning circuits, making layer duplication a practical scaling technique for small language models. The method reveals fundamental differences in how reasoning circuits manifest across model sizes and architectures.

Abstract: Transformer language models contain localized reasoning circuits, contiguous layer blocks that improve reasoning when duplicated at inference time. Finding these circuits currently requires brute-force sweeps costing 25 GPU hours per model. We propose CircuitProbe, which predicts circuit locations from activation statistics in under 5 minutes on CPU, providing a speedup of three to four orders of magnitude. We find that reasoning circuits come in two types: stability circuits in early layers, detected through the derivative of representation change, and magnitude circuits in late layers, detected through anomaly scoring. We validate across 9 models spanning 6 architectures, including 2025 models, confirming that CircuitProbe top predictions match or are within 2 layers of the optimal circuit in all validated cases. A scaling experiment across the Qwen 2.5 family reveals that layer duplication consistently benefits models under 3B parameters but degrades performance in 7B+ models, making this a practical scaling technique for small language models. CircuitProbe requires as few as 10 calibration examples and its predictions are stable across English, Hindi, Chinese, and French.

Method: Built on Petri (open-source LLM auditing tool) with custom scaffold simulating realistic internal deployment of coding agents; tested four frontier models across scenarios varying research motivation, activity type, replacement threat, and model autonomy.

Result: No confirmed instances of research sabotage found; Claude Opus 4.5 Preview and Sonnet 4.5 frequently refused safety-relevant research tasks citing concerns; Opus 4.5 Preview showed reduced unprompted evaluation awareness; both models could distinguish evaluation from deployment when prompted.

Conclusion: While no research sabotage was detected, models exhibited concerning refusals of safety research tasks; evaluation framework validated but has limitations in scenario coverage and evaluation awareness.

Abstract: This technical report presents methods developed by the UK AI Security Institute for assessing whether advanced AI systems reliably follow intended goals. Specifically, we evaluate whether frontier models sabotage safety research when deployed as coding assistants within an AI lab. Applying our methods to four frontier models, we find no confirmed instances of research sabotage. However, we observe that Claude Opus 4.5 Preview (a pre-release snapshot of Opus 4.5) and Sonnet 4.5 frequently refuse to engage with safety-relevant research tasks, citing concerns about research direction, involvement in self-training, and research scope. We additionally find that Opus 4.5 Preview shows reduced unprompted evaluation awareness compared to Sonnet 4.5, while both models can distinguish evaluation from deployment scenarios when prompted. Our evaluation framework builds on Petri, an open-source LLM auditing tool, with a custom scaffold designed to simulate realistic internal deployment of a coding agent. We validate that this scaffold produces trajectories that all tested models fail to reliably distinguish from real deployment data. We test models across scenarios varying in research motivation, activity type, replacement threat, and model autonomy. Finally, we discuss limitations including scenario coverage and evaluation awareness.

Method: Treats each VLM as a probabilistic “expert,” samples multiple outputs, maps them to a unified semantic space, aggregates opinions via uncertainty-weighted linear opinion pooling, and produces system-level uncertainty scores for hallucination detection and abstention.

Result: Achieves AUROC of 0.866 for hallucination detection on ScienceQA (outperforming baselines by 10-13%) and AURAC of 0.907 for abstention (exceeding baselines by 7-9%), with only microsecond-level aggregation overhead.

Conclusion: SCoOP provides an efficient, principled mechanism for uncertainty-aware aggregation in multi-VLM systems, advancing multimodal AI reliability without requiring training.

Abstract: Combining multiple Vision-Language Models (VLMs) can enhance multimodal reasoning and robustness, but aggregating heterogeneous models’ outputs amplifies uncertainty and increases the risk of hallucinations. We propose SCoOP (Semantic-Consistent Opinion Pooling), a training-free uncertainty quantification (UQ) framework for multi-VLM systems through uncertainty-weighted linear opinion pooling. The core idea is to treat each VLM as a probabilistic “expert,” sample multiple outputs, map them to a unified space, aggregate their opinions, and produce a system-level uncertainty score. Unlike prior UQ methods designed for single models, SCoOP explicitly measures collective, system-level uncertainty across multiple VLMs, enabling effective hallucination detection and abstention for highly uncertain samples. On ScienceQA, SCoOP achieves an AUROC of 0.866 for hallucination detection, outperforming baselines (0.732-0.757) by approximately 10-13%. For abstention, it attains an AURAC of 0.907, exceeding baselines (0.818-0.840) by 7-9%. Despite these gains, SCoOP introduces only microsecond-level aggregation overhead relative to the baselines, which is trivial compared to typical VLM inference time (on the order of seconds). These results demonstrate that SCoOP provides an efficient and principled mechanism for uncertainty-aware aggregation, advancing the reliability of multimodal AI systems. Our code is publicly available at https://github.com/chungenyu6/SCoOP.

Method: Introduces RefineRL with two innovations: 1) Skeptical-Agent that validates solutions against test cases and maintains skepticism even when validation suggests correctness, 2) RL training using standard RLVR data to incentivize self-refinement.

Result: 4B models with RefineRL outperform 32B models and approach single-attempt performance of 235B models, demonstrating substantial gains through self-refinement.

Conclusion: Self-refinement holds significant promise for scaling LLM reasoning capabilities, with potential for further advancement beyond competitive programming applications.

Abstract: While large language models (LLMs) have demonstrated strong performance on complex reasoning tasks such as competitive programming (CP), existing methods predominantly focus on single-attempt settings, overlooking their capacity for iterative refinement. In this paper, we present RefineRL, a novel approach designed to unleash the self-refinement capabilities of LLMs for CP problem solving. RefineRL introduces two key innovations: (1) Skeptical-Agent, an iterative self-refinement agent equipped with local execution tools to validate generated solutions against public test cases of CP problems. This agent always maintains a skeptical attitude towards its own outputs and thereby enforces rigorous self-refinement even when validation suggests correctness. (2) A reinforcement learning (RL) solution to incentivize LLMs to self-refine with only standard RLVR data (i.e., problems paired with their verifiable answers). Extensive experiments on Qwen3-4B and Qwen3-4B-2507 demonstrate that our method yields substantial gains: after our RL training, these compact 4B models integrated with the Skeptical-Agent not only outperform much larger 32B models but also approach the single-attempt performance of 235B models. These findings suggest that self-refinement holds considerable promise for scaling LLM reasoning, with significant potential for further advancement.

Method: Preference Guided Iterated Pareto Referent Optimisation (PG-IPRO) - an iterative algorithm where users provide feedback on routes to indicate which objectives should be minimized or relaxed, avoiding computation of full Pareto front for efficiency.

Result: PG-IPRO enables intuitive user interaction, especially effective in early iterations compared to information-gain-based approaches, with higher computational efficiency and shorter waiting times.

Conclusion: PG-IPRO provides an effective interactive route planning system for people with accessibility needs through preference-guided optimization and efficient iterative computation.

Abstract: We propose the Preference Guided Iterated Pareto Referent Optimisation (PG-IPRO) for urban route planning for people with different accessibility requirements and preferences. With this algorithm the user can interact with the system by giving feedback on a route, i.e., the user can say which objective should be further minimized, or conversely can be relaxed. This leads to intuitive user interaction, that is especially effective during early iterations compared to information-gain-based interaction. Furthermore, due to PG-IPRO’s iterative nature, the full set of alternative, possibly optimal policies (the Pareto front), is never computed, leading to higher computational efficiency and shorter waiting times for users.

Method: Proposes MARS-GPS which: 1) generates multiple parallel reasoning rollouts augmented with Python code execution for numerical verification, 2) ranks rollouts using token-level entropy as confidence signal, and 3) aggregates answers through multi-stage voting and self-verification pipeline.

Result: Achieves 88.8% accuracy on Geometry3K with 8 parallel rollouts, a nearly +11% improvement over prior state-of-the-art. Accuracy scales consistently as number of rollouts increases from 1 to 16 (+6.0% on ablation subset).

Conclusion: MARS-GPS significantly advances geometric problem solving by enhancing logical inference through parallel reasoning rollouts with numerical verification and multi-stage answer aggregation.

Abstract: Geometric Problem Solving (GPS) remains at the heart of enhancing mathematical reasoning in large language models because it requires the combination of diagrammatic understanding, symbolic manipulation and logical inference. In existing literature, researchers have chiefly focused on synchronising the diagram descriptions with text literals and solving the problem. In this vein, they have either taken a neural, symbolic or neuro-symbolic approach. But this solves only the first two of the requirements, namely diagrammatic understanding and symbolic manipulation, while leaving logical inference underdeveloped. The logical inference is often limited to one chain-of-thought (CoT). To address this weakness in hitherto existing models, this paper proposes MARS-GPS, that generates multiple parallel reasoning rollouts augmented with Python code execution for numerical verification, ranks them using token-level entropy as a confidence signal, and aggregates answers through a multi-stage voting and self-verification pipeline. Empirical results show that MARS-GPS with 8 parallel rollouts achieves 88.8% on Geometry3K, a nearly +11% improvement over the prior state-of-the-art, with accuracy scaling consistently as the number of rollouts increases from 1 to 16 (+6.0% on ablation subset). We provide our code and data in an anonymous repository: https://anonymous.4open.science/r/MARS-GPS-DE55.

Method: HERA uses a hierarchical framework with two levels: 1) Global level: optimizes query-specific agent topologies through reward-guided sampling and experience accumulation; 2) Local level: Role-Aware Prompt Evolution refines agent behaviors via credit assignment and dual-axes adaptation along operational and behavioral principles.

Result: On six knowledge-intensive benchmarks, HERA achieves an average improvement of 38.69% over recent baselines while maintaining robust generalization and token efficiency. Topological analyses reveal emergent self-organization where sparse exploration yields compact, high-utility multi-agent networks.

Conclusion: HERA demonstrates both efficient coordination and robust reasoning through adaptive multi-agent orchestration and role-specific prompt evolution, addressing limitations of static approaches in multi-agent RAG systems.

Abstract: Multi-agent Retrieval-Augmented Generation (RAG), wherein each agent takes on a specific role, supports hard queries that require multiple steps and sources, or complex reasoning. Existing approaches, however, rely on static agent behaviors and fixed orchestration strategies, leading to brittle performance on diverse, multi-hop tasks. We identify two key limitations: the lack of continuously adaptive orchestration mechanisms and the absence of behavior-level learning for individual agents. To this end, we propose HERA, a hierarchical framework that jointly evolves multi-agent orchestration and role-specific agent prompts. At the global level, HERA optimizes query-specific agent topologies through reward-guided sampling and experience accumulation. At the local level, Role-Aware Prompt Evolution refines agent behaviors via credit assignment and dual-axes adaptation along operational and behavioral principles, enabling targeted, role-conditioned improvements. On six knowledge-intensive benchmarks, HERA achieves an average improvement of 38.69% over recent baselines while maintaining robust generalization and token efficiency. Topological analyses reveal emergent self-organization, where sparse exploration yields compact, high-utility multi-agent networks, demonstrating both efficient coordination and robust reasoning.

Method: Three-engine framework: 1) Memory-Augmented Planning Engine for longitudinal multi-session interactions with persistent memory, 2) Skill Evolution Engine that extracts new practice-grounded skills from historical counseling trajectories, 3) Reinforced Internalization Engine that integrates evolved skills via rejection fine-tuning.

Result: Achieves higher scores than strong general LLMs (GPT-5.4, Gemini-3) and domain-specific baselines across all reported evaluation dimensions, demonstrating improved consistency and quality of multi-session counseling responses.

Conclusion: Lifelong learning through experience-driven evolution can significantly improve AI psychological counseling quality and consistency, moving beyond static training approaches to more human-like continuous improvement.

Abstract: Existing methods for AI psychological counselors predominantly rely on supervised fine-tuning using static dialogue datasets. However, this contrasts with human experts, who continuously refine their proficiency through clinical practice and accumulated experience. To bridge this gap, we propose an Experience-Driven Lifelong Learning Agent (\texttt{PsychAgent}) for psychological counseling. First, we establish a Memory-Augmented Planning Engine tailored for longitudinal multi-session interactions, which ensures therapeutic continuity through persistent memory and strategic planning. Second, to support self-evolution, we design a Skill Evolution Engine that extracts new practice-grounded skills from historical counseling trajectories. Finally, we introduce a Reinforced Internalization Engine that integrates the evolved skills into the model via rejection fine-tuning, aiming to improve performance across diverse scenarios. Comparative analysis shows that our approach achieves higher scores than strong general LLMs (e.g., GPT-5.4, Gemini-3) and domain-specific baselines across all reported evaluation dimensions. These results suggest that lifelong learning can improve the consistency and overall quality of multi-session counseling responses.

Method: Deployed an autonomous research pipeline that executes ~50 experiments across two benchmarks, autonomously diagnosing failures, proposing architectural modifications, and fixing data pipeline bugs without human intervention.

Result: Achieved state-of-the-art performance: +411% F1 improvement on LoCoMo (0.117→0.598) and +214% on Mem-Gallery (0.254→0.797). Most impactful discoveries were bug fixes (+175%), architectural changes (+44%), and prompt engineering (+188% on specific categories).

Conclusion: Autonomous research pipelines can effectively explore complex AI system design spaces, with multimodal memory being particularly suited for such approaches due to its properties. The work provides guidance for applying autonomous research to other AI domains.

Abstract: AI agents increasingly operate over extended time horizons, yet their ability to retain, organize, and recall multimodal experiences remains a critical bottleneck. Building effective lifelong memory requires navigating a vast design space spanning architecture, retrieval strategies, prompt engineering, and data pipelines; this space is too large and interconnected for manual exploration or traditional AutoML to explore effectively. We deploy an autonomous research pipeline to discover OmniMem, a unified multimodal memory framework for lifelong AI agents. Starting from a naïve baseline (F1=0.117 on LoCoMo), the pipeline autonomously executes ${\sim}50$ experiments across two benchmarks, diagnosing failure modes, proposing architectural modifications, and repairing data pipeline bugs, all without human intervention in the inner loop. The resulting system achieves state-of-the-art on both benchmarks, improving F1 by +411% on LoCoMo (0.117$\to$0.598) and +214% on Mem-Gallery (0.254$\to$0.797) relative to the initial configurations. Critically, the most impactful discoveries are not hyperparameter adjustments: bug fixes (+175%), architectural changes (+44%), and prompt engineering (+188% on specific categories) each individually exceed the cumulative contribution of all hyperparameter tuning, demonstrating capabilities fundamentally beyond the reach of traditional AutoML. We provide a taxonomy of six discovery types and identify four properties that make multimodal memory particularly suited for autoresearch, offering guidance for applying autonomous research pipelines to other AI system domains. Code is available at this https://github.com/aiming-lab/OmniMem.

Method: AMST applies structured stress transformations to prompts and evaluates model behavior through distribution-aware robustness metrics that capture variance, tail risk, and temporal behavioral drift across interaction rounds. It’s a scalable, model-agnostic stress-testing methodology.

Result: Evaluation on LLaMA-3-8B, GPT-4o, and DeepSeek-v3 reveals substantial differences in robustness profiles across models and exposes degradation patterns not observable under conventional single-round evaluation. Robustness depends on distributional stability and tail behavior rather than average performance alone.

Conclusion: AMST provides a robust framework for assessing ethical robustness under adversarial multi-round interactions, enabling better evaluation and monitoring of LLM-enabled systems in adversarial environments.

Abstract: Evaluating the ethical robustness of large language models (LLMs) deployed in software systems remains challenging, particularly under sustained adversarial user interaction. Existing safety benchmarks typically rely on single-round evaluations and aggregate metrics, such as toxicity scores and refusal rates, which offer limited visibility into behavioral instability that may arise during realistic multi-turn interactions. As a result, rare but high-impact ethical failures and progressive degradation effects may remain undetected prior to deployment. This paper introduces Adversarial Moral Stress Testing (AMST), a stress-based evaluation framework for assessing ethical robustness under adversarial multi-round interactions. AMST applies structured stress transformations to prompts and evaluates model behavior through distribution-aware robustness metrics that capture variance, tail risk, and temporal behavioral drift across interaction rounds. We evaluate AMST on several state-of-the-art LLMs, including LLaMA-3-8B, GPT-4o, and DeepSeek-v3, using a large set of adversarial scenarios generated under controlled stress conditions. The results demonstrate substantial differences in robustness profiles across models and expose degradation patterns that are not observable under conventional single-round evaluation protocols. In particular, robustness has been shown to depend on distributional stability and tail behavior rather than on average performance alone. Additionally, AMST provides a scalable and model-agnostic stress-testing methodology that enables robustness-aware evaluation and monitoring of LLM-enabled software systems operating in adversarial environments.

Method: Used linear probes to decode tool-calling decisions from pre-generation activations, conducted activation steering experiments to causally test decision encoding, and performed behavioral analysis of chain-of-thought rationalization.

Result: Simple linear probes successfully decode decisions with high confidence from early activations, sometimes before any reasoning tokens. Activation steering flips behavior in 7-79% of cases depending on model/benchmark, and chain-of-thought often rationalizes flipped decisions rather than resisting them.

Conclusion: Reasoning models encode action choices before beginning textual deliberation, suggesting they may decide first and then think/justify rather than think then decide.

Abstract: We consider the question: when a large language reasoning model makes a choice, did it think first and then decide to, or decide first and then think? In this paper, we present evidence that detectable, early-encoded decisions shape chain-of-thought in reasoning models. Specifically, we show that a simple linear probe successfully decodes tool-calling decisions from pre-generation activations with very high confidence, and in some cases, even before a single reasoning token is produced. Activation steering supports this causally: perturbing the decision direction leads to inflated deliberation, and flips behavior in many examples (between 7 - 79% depending on model and benchmark). We also show through behavioral analysis that, when steering changes the decision, the chain-of-thought process often rationalizes the flip rather than resisting it. Together, these results suggest that reasoning models can encode action choices before they begin to deliberate in text.

Method: Created device-scale file systems over real-world profiles with 42.4GB of data across 2K+ real-world files spanning diverse modalities. Constructed 581 QA pairs to assess search, evidence perception, and multi-step reasoning, plus 46.1K densely annotated structured trajectories for step-wise failure diagnosis.

Result: Even the most advanced commercial MLLMs achieve only 48.3% accuracy in user profiling, struggling with long-horizon retrieval and cross-modal reasoning. Step-wise diagnosis identifies multimodal perception and evidence grounding as primary bottlenecks.

Conclusion: HippoCamp exposes critical limitations of current agents in realistic user-centric environments and provides a foundation for developing next-generation personal AI assistants that can effectively manage multimodal personal files.

Abstract: We present HippoCamp, a new benchmark designed to evaluate agents’ capabilities on multimodal file management. Unlike existing agent benchmarks that focus on tasks like web interaction, tool use, or software automation in generic settings, HippoCamp evaluates agents in user-centric environments to model individual user profiles and search massive personal files for context-aware reasoning. Our benchmark instantiates device-scale file systems over real-world profiles spanning diverse modalities, comprising 42.4 GB of data across over 2K real-world files. Building upon the raw files, we construct 581 QA pairs to assess agents’ capabilities in search, evidence perception, and multi-step reasoning. To facilitate fine-grained analysis, we provide 46.1K densely annotated structured trajectories for step-wise failure diagnosis. We evaluate a wide range of state-of-the-art multimodal large language models (MLLMs) and agentic methods on HippoCamp. Our comprehensive experiments reveal a significant performance gap: even the most advanced commercial models achieve only 48.3% accuracy in user profiling, struggling particularly with long-horizon retrieval and cross-modal reasoning within dense personal file systems. Furthermore, our step-wise failure diagnosis identifies multimodal perception and evidence grounding as the primary bottlenecks. Ultimately, HippoCamp exposes the critical limitations of current agents in realistic, user-centric environments and provides a robust foundation for developing next-generation personal AI assistants.

Method: CoCoA uses a two-stage framework: 1) Code Comprehension stage generates natural-language explanations of functionality, 2) Code Auditing stage evaluates task alignment based on the explanation rather than raw code.

Result: Across multiple datasets, programming languages, and models, CoCoA achieves up to 68% increased F1 score and up to 20% increased accuracy over best-performing baselines.

Conclusion: Separating comprehension from evaluation improves code correctness assessment by focusing on behavioral alignment rather than implementation details, making LLM-based evaluation more reliable.

Abstract: Large Language Models (LLMs) for unsupervised code correctness evaluation have recently gained attention because they can judge if code runs as intended without requiring reference implementations or unit tests, which may be unavailable, sparse, or unreliable. However, most prior approaches condition LLM evaluators directly on the full code implementation, forcing the model to jointly infer program behavior and evaluate correctness in a single step. This entanglement leads to misinterpretations of code behavior and unreliable judgments. To mitigate this issue, we introduce CoCoA, an unsupervised Code Comprehension then Auditing framework that first comprehends functionality to generate a natural-language explanation. Then it evaluates task alignment based on this explanation. By sequentially sampling comprehension before evaluation, CoCoA improves the quality of inferred program behavior and enables the evaluator to focus on behavioral alignment rather than raw implementation details. Across multiple datasets, programming languages, and models, CoCoA achieves up to $68%$ increased F1 score and up to $20%$ increased accuracy over the best-performing baselines.

Method: Analytical survey approach: traces evolution of RAG paradigms, introduces taxonomy of Agentic RAG architectures based on agent cardinality, control structure, autonomy, and knowledge representation, and provides comparative analysis of design trade-offs across existing frameworks.

Result: Comprehensive survey of Agentic RAG systems examining applications in healthcare, finance, education, and enterprise document processing, with practical lessons for system designers and identification of key research challenges.

Conclusion: Agentic RAG represents a significant advancement over traditional RAG by enabling flexibility, scalability, and context-awareness through autonomous agents, though challenges remain in evaluation, coordination, memory management, efficiency, and governance.

Abstract: Large Language Models (LLMs) have advanced artificial intelligence by enabling human-like text generation and natural language understanding. However, their reliance on static training data limits their ability to respond to dynamic, real-time queries, resulting in outdated or inaccurate outputs. Retrieval-Augmented Generation (RAG) has emerged as a solution, enhancing LLMs by integrating real-time data retrieval to provide contextually relevant and up-to-date responses. Despite its promise, traditional RAG systems are constrained by static workflows and lack the adaptability required for multi-step reasoning and complex task management. Agentic Retrieval-Augmented Generation (Agentic RAG) transcends these limitations by embedding autonomous AI agents into the RAG pipeline. These agents leverage agentic design patterns reflection, planning, tool use, and multi-agent collaboration to dynamically manage retrieval strategies, iteratively refine contextual understanding, and adapt workflows through operational structures ranging from sequential steps to adaptive collaboration. This integration enables Agentic RAG systems to deliver flexibility, scalability, and context-awareness across diverse applications. This paper presents an analytical survey of Agentic RAG systems. It traces the evolution of RAG paradigms, introduces a principled taxonomy of Agentic RAG architectures based on agent cardinality, control structure, autonomy, and knowledge representation, and provides a comparative analysis of design trade-offs across existing frameworks. The survey examines applications in healthcare, finance, education, and enterprise document processing, and distills practical lessons for system designers and practitioners. Finally, it identifies key open research challenges related to evaluation, coordination, memory management, efficiency, and governance, outlining directions for future research.

Method: The study evaluates three approaches to tune AI agents for exception handling: ethical framework prompting, chain-of-thought reasoning, and supervised fine-tuning (with human explanations). Experiments test how LLMs handle exceptions compared to human judgments.

Result: Ethical framework prompting fails, chain-of-thought provides only slight improvements, but supervised fine-tuning with human explanations yields markedly better results. Surprisingly, fine-tuned models generalized human-like decision-making to novel scenarios, demonstrating transfer learning. Fine-tuning with explanations (not just labels) was critical for alignment.

Conclusion: Aligning LLMs with human judgment requires explicit training on how decisions are made, not just which decisions are made. Addressing LLMs’ shortcomings in exception handling is crucial for developing agentic AI that can effectively align with human judgment and adapt to novel contexts.

Abstract: Large language models (LLMs), initially developed for generative AI, are now evolving into agentic AI systems, which make decisions in complex, real-world contexts. Unfortunately, while their generative capabilities are well-documented, their decision-making processes remain poorly understood. This is particularly evident when testing targeted decision-making: for instance, how models handle exceptions, a critical and challenging aspect of decision-making made relevant by the inherent incompleteness of contracts. Here we demonstrate that LLMs, even ones that excel at reasoning, deviate significantly from human judgments because they adhere strictly to policies, even when such adherence is impractical, suboptimal, or even counterproductive. We then evaluate three approaches to tuning AI agents to handle exceptions: ethical framework prompting, chain-of-thought reasoning, and supervised fine-tuning. We find that while ethical framework prompting fails and chain-of-thought prompting provides only slight improvements, supervised fine-tuning - specifically with human explanations - yields markedly better results. Surprisingly, in our experiments, supervised fine-tuning even enabled models to generalize human-like decision-making to novel scenarios, demonstrating transfer learning of human-aligned decision-making across contexts. Furthermore, fine-tuning with explanations, not just labels, was critical for alignment, suggesting that aligning LLMs with human judgment requires explicit training on how decisions are made, not just which decisions are made. These findings highlight the need to address LLMs’ shortcomings in handling exceptions in order to guide the development of agentic AI toward models that can effectively align with human judgment and simultaneously adapt to novel contexts.

Method: Uses activation steering (inference-time technique modulating internal activations) after localizing layers responsible for formal vs plausible inference. Investigates contrastive steering methods and introduces novel kNN-based conditional approach (K-CAST) for dynamic parameter determination

Result: Contrastive steering supports linear control over content biases but static approach insufficient for all models. K-CAST achieves up to 15% absolute improvement in formal reasoning accuracy, robust to prompt variations with minimal side effects on multilingual capabilities

Conclusion: Activation-level interventions offer scalable inference-time strategy for enhancing LLM robustness, contributing to more systematic and unbiased reasoning capabilities

Abstract: Large language models (LLMs) exhibit reasoning biases, often conflating content plausibility with formal logical validity. This can lead to wrong inferences in critical domains, where plausible arguments are incorrectly deemed logically valid or vice versa. This paper investigates how content biases on reasoning can be mitigated through activation steering, an inference-time technique that modulates internal activations. Specifically, after localising the layers responsible for formal and plausible inference, we investigate activation steering on a controlled syllogistic reasoning task, designed to disentangle formal validity from content plausibility. An extensive empirical analysis reveals that contrastive steering methods consistently support linear control over content biases. However, a static approach is insufficient to debias all the tested models. We then investigate how to control content effects by dynamically determining the steering parameters through fine-grained conditional methods. By introducing a novel kNN-based conditional approach (K-CAST), we demonstrate that conditional steering can effectively reduce biases on unresponsive models, achieving up to 15% absolute improvement in formal reasoning accuracy. Finally, we found that steering for content effects is robust to prompt variations, incurs minimal side effects on multilingual language modeling capabilities, and can partially generalize to different reasoning tasks. In practice, we demonstrate that activation-level interventions offer a scalable inference-time strategy for enhancing the robustness of LLMs, contributing towards more systematic and unbiased reasoning capabilities.

Method: Created LocationReasoner benchmark with carefully crafted queries of varying difficulty levels, supported by a sandbox environment with constraint-based location search tools and automated verification for scalability.

Result: State-of-the-art reasoning models show limited improvement over non-reasoning predecessors in real-world contexts; OpenAI o4 fails on 30% of tasks; agentic strategies like ReAct and Reflexion suffer from over-reasoning.

Conclusion: LLMs have key limitations in holistic and non-linear reasoning for real-world decision-making; the benchmark is released to foster development of LLMs capable of robust, grounded reasoning.

Abstract: Recent advances in large language models (LLMs), particularly those enhanced through reinforced post-training, have demonstrated impressive reasoning capabilities, as exemplified by models such as OpenAI o1 and DeepSeek-R1. However, these capabilities are predominantly benchmarked on domains like mathematical problem solving and code generation, leaving open the question of whether such reasoning skills generalize to complex real-world scenarios. In this paper, we introduce LocationReasoner, a benchmark designed to evaluate LLMs’ reasoning abilities in the context of real-world site selection, where models must identify feasible locations by reasoning over diverse and complicated spatial, environmental, and logistic constraints. The benchmark covers carefully crafted queries of varying difficulty levels and is supported by a sandbox environment with in-house tools for constraint-based location search. Automated verification further guarantees the scalability of the benchmark, enabling the addition of arbitrary number of queries. Extensive evaluations on real-world site selection data from Boston, New York, and Tampa reveal that state-of-the-art reasoning models offer limited improvement over their non-reasoning predecessors in real-world contexts, with even the latest OpenAI o4 model failing on 30% of site selection tasks. Moreover, agentic strategies such as ReAct and Reflexion often suffer from over-reasoning, leading to worse outcomes than direct prompting. With key limitations of LLMs in holistic and non-linear reasoning highlighted, we release LocationReasoner to foster the development of LLMs and agents capable of robust, grounded reasoning in real-world decision-making tasks. Codes and data for our benchmark are available at https://github.com/miho-koda/LocationReasoner.

Method: Introduces HiMA-Ecom benchmark with agent-specific supervised fine-tuning samples and system-level input-output pairs. Proposes HiMA-R1 training method with Variance-Reduction Group Relative Policy Optimization (VR-GRPO) using Monte Carlo sampling to handle exponential action space, and adaptive memory evolution mechanism using GRPO rewards as supervisory signals.

Result: Method built on smaller 3B/7B open-source models achieves performance comparable to larger LLMs like DeepSeek-R1 and surpasses DeepSeek-V3 by average 6% on HiMA-Ecom benchmark.

Conclusion: HiMA-Ecom addresses the lack of realistic benchmarks for hierarchical multi-agent systems in e-commerce, and HiMA-R1 provides effective joint training method that enables smaller models to compete with larger ones.

Abstract: Hierarchical multi-agent systems based on large language models (LLMs) have become a common paradigm for building AI assistants in vertical domains such as e-commerce, where a master agent coordinates multiple specialized sub-agents. Despite their practical importance, realistic benchmarks for training and evaluating such systems remain scarce, and joint optimization across functionally distinct agents is still challenging. To address this gap, we introduce HiMA-Ecom, the first hierarchical multi-agent benchmark tailored for e-commerce scenarios. HiMA-Ecom contains 22.8K instances, including agent-specific supervised fine-tuning samples with memory and system-level input-output pairs for joint multi-agent reinforcement learning. Building upon it, a joint training method named HiMA-R1 is proposed. It presents Variance-Reduction Group Relative Policy Optimization (VR-GRPO), which employs initial trajectory-based Monte Carlo sampling to mitigate the exponential joint action space and selects informative agent groups for efficient updates based on reward variance. Furthermore, an adaptive memory evolution mechanism that repurposes GRPO rewards as cost-free supervisory signals is designed to eliminate repetitive reasoning and accelerate convergence. Experiments on HiMA-Ecom demonstrate that our method, built upon smaller 3B/7B open-source models, achieves performance comparable to that of larger LLMs, such as DeepSeek-R1, and surpasses DeepSeek-V3 by an average of 6%.

Method: Created DP-Bench benchmark for evaluation, developed DPLM (7B-parameter specialized model), and introduced DualReflect pipeline for synthetic data generation combining forward (diverse) and backward (reliable) generation approaches.

Result: DPLM achieves performance comparable to state-of-the-art LLMs like OpenAI’s o1 and DeepSeek-R1, surpassing them on hard problems, demonstrating effectiveness of the DualReflect approach.

Conclusion: Backward generation is crucial in low-data regimes for correctness, while forward generation adds value at scale for diversity, highlighting the complementary strengths of both approaches in automated DP formulation.

Abstract: Dynamic programming (DP) is a fundamental method in operations research, but formulating DP models has traditionally required expert knowledge of both the problem context and DP techniques. Large Language Models (LLMs) offer the potential to automate this process. However, DP problems pose unique challenges due to their inherently stochastic transitions and the limited availability of training data. These factors make it difficult to directly apply existing LLM-based models or frameworks developed for other optimization problems, such as linear or integer programming. We introduce DP-Bench, the first benchmark covering a wide range of textbook-level DP problems to enable systematic evaluation. We present Dynamic Programming Language Model (DPLM), a 7B-parameter specialized model that achieves performance comparable to state-of-the-art LLMs like OpenAI’s o1 and DeepSeek-R1, and surpasses them on hard problems. Central to DPLM’s effectiveness is DualReflect, our novel synthetic data generation pipeline, designed to scale up training data from a limited set of initial examples. DualReflect combines forward generation for diversity and backward generation for reliability. Our results reveal a key insight: backward generation is favored in low-data regimes for its strong correctness guarantees, while forward generation, though lacking such guarantees, becomes increasingly valuable at scale for introducing diverse formulations. This trade-off highlights the complementary strengths of both approaches and the importance of combining them.

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions due to missing paper content

Abstract: Failed to fetch summary for 2509.21743: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.21743&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 2510.18314: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.18314&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Method unknown - paper content not accessible

Result: No results available - paper summary fetch failed

Conclusion: Cannot draw conclusions about an inaccessible paper

Abstract: Failed to fetch summary for 2511.08206: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.08206&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Not applicable - paper content unavailable

Result: HTTP 429 error indicates too many requests to arXiv API; paper details not retrieved

Conclusion: Cannot analyze paper due to technical limitations in accessing the content

Abstract: Failed to fetch summary for 2511.08947: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.08947&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2601.05144: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.05144&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unknown - paper content not accessible due to API rate limiting

Result: No results available - failed to fetch paper information

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

Abstract: Failed to fetch summary for 2602.08734: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.08734&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetching error

Result: Unable to determine results due to fetching error

Conclusion: Unable to determine conclusion due to fetching error

Abstract: Failed to fetch summary for 2602.19837: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.19837&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 2602.22413: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.22413&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content. The arXiv API returned a rate limiting error (HTTP 429), indicating too many requests.

Result: Cannot determine results without access to paper content. The paper summary could not be fetched due to technical limitations.

Conclusion: Cannot draw conclusions without access to paper content. The arXiv API rate limiting prevents analysis of this specific paper.

Abstract: Failed to fetch summary for 2603.04636: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.04636&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 2603.15929: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.15929&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 2603.28183: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.28183&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to missing paper content

Result: Cannot determine results due to missing paper content

Conclusion: Cannot determine conclusion due to missing paper content

Abstract: Failed to fetch summary for 2603.29678: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.29678&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 2603.30031: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.30031&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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

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

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

Abstract: Failed to fetch summary for 2407.01570: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2407.01570&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 2411.10174: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2411.10174&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to missing paper content

Result: Cannot determine results due to missing paper content

Conclusion: Cannot determine conclusion due to missing paper content

Abstract: Failed to fetch summary for 2502.05181: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2502.05181&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 2505.15808: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.15808&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 2505.17760: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.17760&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 2505.19574: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.19574&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available - paper content inaccessible

Result: No results available - technical error prevented paper retrieval

Conclusion: Cannot analyze paper due to technical limitations in accessing content

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 fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2506.21997: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.21997&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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

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

Conclusion: Cannot draw conclusions due to inability to access paper content

Abstract: Failed to fetch summary for 2509.25424: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.25424&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to API access error

Result: Unable to determine results due to API access error

Conclusion: Unable to determine conclusion due to API access error

Abstract: Failed to fetch summary for 2510.00512: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.00512&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to failed API request

Result: Cannot determine results due to failed API request

Conclusion: Cannot determine conclusion due to failed API request

Abstract: Failed to fetch summary for 2510.06545: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.06545&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2510.16082: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.16082&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 2510.24906: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.24906&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot determine conclusion without access to paper content

Abstract: Failed to fetch summary for 2510.25770: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.25770&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 2602.00750: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.00750&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to determine conclusion due to failed paper fetch

Abstract: Failed to fetch summary for 2602.06801: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.06801&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

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

Abstract: Failed to fetch summary for 2602.08040: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.08040&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2602.13851: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.13851&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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

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

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

Abstract: Failed to fetch summary for 2602.18807: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.18807&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to API rate limiting error

Result: Unable to determine results due to API rate limiting error

Conclusion: Unable to determine conclusion due to API rate limiting error

Abstract: Failed to fetch summary for 2602.23499: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.23499&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 2603.13431: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.13431&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to missing paper content

Result: Unable to determine results due to missing paper content

Conclusion: Unable to draw conclusions due to missing paper content

Abstract: Failed to fetch summary for 2603.19263: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.19263&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 2603.20179: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.20179&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper retrieval

Result: Unable to determine results due to failed paper retrieval

Conclusion: Unable to draw conclusions due to failed paper retrieval

Abstract: Failed to fetch summary for 2603.20897: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.20897&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 2603.24324: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.24324&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

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

Abstract: Failed to fetch summary for 2603.24639: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.24639&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 2603.26461: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.26461&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available due to failed API request

Result: No results available - paper summary fetch failed with HTTP 429 (Too Many Requests) error

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

Abstract: Failed to fetch summary for 2603.26648: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.26648&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 2603.26692: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.26692&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 2603.27420: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.27420&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to missing abstract.

Result: Unable to determine results due to missing abstract.

Conclusion: Unable to draw conclusions due to missing abstract.

Abstract: Failed to fetch summary for 2603.28458: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.28458&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 2603.29590: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.29590&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to lack of accessible content.

Result: Cannot determine results due to lack of accessible content.

Conclusion: Cannot draw conclusions due to lack of accessible content.

Abstract: Failed to fetch summary for 2603.29805: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.29805&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Uses three Mamba-based modules: gated bidirectional Mamba text encoder, Temporal Bi-Mamba with lightweight alignment teacher, and Expressive Mamba with AdaLN modulation. All attention and explicit RNN layers are removed from the conditioning path while keeping the StyleTTS2 mel-diffusion-vocoder backbone.

Result: Achieves modest but statistically reliable gains over StyleTTS2, VITS, and Mamba-attention hybrids in MOS/CMOS, F0 RMSE, MCD, and WER. Reduces encoder parameters to 21M and improves throughput by 1.6x with linear-time O(T) conditioning and bounded activation memory.

Conclusion: SSM-only conditioning improves memory footprint, stability, and deployability for TTS systems while maintaining quality, though diffusion remains the dominant latency source.

Abstract: MambaVoiceCloning (MVC) asks whether the conditioning path of diffusion-based TTS can be made fully SSM-only at inference, removing all attention and explicit RNN-style recurrence layers across text, rhythm, and prosody, while preserving or improving quality under controlled conditions. MVC combines a gated bidirectional Mamba text encoder, a Temporal Bi-Mamba supervised by a lightweight alignment teacher discarded after training, and an Expressive Mamba with AdaLN modulation, yielding linear-time O(T) conditioning with bounded activation memory and practical finite look-ahead streaming. Unlike prior Mamba-TTS systems that remain hybrid at inference, MVC removes attention-based duration and style modules under a fixed StyleTTS2 mel-diffusion-vocoder backbone. Trained on LJSpeech/LibriTTS and evaluated on VCTK, CSS10 (ES/DE/FR), and long-form Gutenberg passages, MVC achieves modest but statistically reliable gains over StyleTTS2, VITS, and Mamba-attention hybrids in MOS/CMOS, F0 RMSE, MCD, and WER, while reducing encoder parameters to 21M and improving throughput by 1.6x. Diffusion remains the dominant latency source, but SSM-only conditioning improves memory footprint, stability, and deployability.

Method: 32 HF patients collected daily voice recordings for 2 months with standard care measures; acoustic analysis extracted vowel/speech features; time-series features from lookback windows predicted next-day health status using explainable ML with nested cross-validation

Result: Voice features outperformed standard care measures (sensitivity 0.826 vs 0.783, specificity 0.782 vs 0.567); key prognostic features included delayed energy shift, low energy variability, higher shimmer variability in vowels, and speech rate/quality changes

Conclusion: Voice-based monitoring provides a non-invasive approach for early detection of health changes in chronic heart failure, enabling proactive and personalized care

Abstract: Objective: This study aimed to evaluate which voice features can predict health deterioration in patients with chronic HF. Background: Heart failure (HF) is a chronic condition with progressive deterioration and acute decompensations, often requiring hospitalization and imposing substantial healthcare and economic burdens. Current standard-of-care (SoC) home monitoring, such as weight tracking, lacks predictive accuracy and requires high patient engagement. Voice is a promising non-invasive biomarker, though prior studies have mainly focused on acute HF stages. Methods: In a 2-month longitudinal study, 32 patients with HF collected daily voice recordings and SoC measures of weight and blood pressure at home, with biweekly questionnaires for health status. Acoustic analysis generated detailed vowel and speech features. Time-series features were extracted from aggregated lookback windows (e.g., 7 days) to predict next-day health status. Explainable machine learning with nested cross-validation identified top vocal biomarkers, and a case study illustrated model application. Results: A total of 21,863 recordings were analyzed. Acoustic vowel features showed strong correlations with health status. Time-series voice features within the lookback window outperformed corresponding standard care measures, achieving peak sensitivity and specificity of 0.826 and 0.782 versus 0.783 and 0.567 for SoC metrics. Key prognostic voice features identifying deterioration included delayed energy shift, low energy variability, and higher shimmer variability in vowels, along with reduced speaking and articulation rate, lower phonation ratio, decreased voice quality, and increased formant variability in speech. Conclusion: Voice-based monitoring offers a non-invasive approach to detect early health changes in chronic HF, supporting proactive and personalized care.

Method: Interactive mobile system with target-conditioned neural pipeline supporting simultaneous attenuation of multiple overlapping sound sources. Runs real-time on-device with user-extensible sound classes through in-situ audio examples without retraining.

Result: Achieves low-latency, multi-target attenuation suitable for live listening. Enables meaningful reductions in bothersome sounds while maintaining awareness of surroundings. Validated through technical benchmarking and in-situ study with 10 participants.

Conclusion: Sona represents a new class of personal AI systems that mediate real-world acoustic environments to support comfort and social participation through selective sound attenuation.

Abstract: For people with noise sensitivity, everyday soundscapes can be overwhelming. Existing tools such as active noise cancellation reduce discomfort by suppressing the entire acoustic environment, often at the cost of awareness of surrounding people and events. We present Sona, an interactive mobile system for real-time soundscape mediation that selectively attenuates bothersome sounds while preserving desired audio. Sona is built on a target-conditioned neural pipeline that supports simultaneous attenuation of multiple overlapping sound sources, overcoming the single-target limitation of prior systems. It runs in real time on-device and supports user-extensible sound classes through in-situ audio examples, without retraining. Sona is informed by a formative study with 68 noise-sensitive individuals. Through technical benchmarking and an in-situ study with 10 participants, we show that Sona achieves low-latency, multi-target attenuation suitable for live listening, and enables meaningful reductions in bothersome sounds while maintaining awareness of surroundings. These results point toward a new class of personal AI systems that support comfort and social participation by mediating real-world acoustic environments.

Method: Two-stage framework: 1) KDQS encodes dance motions into latent representation using Finite Scalar Quantization with kinematic-dynamic constraints, 2) HMDGS uses Mamba-Transformer hybrid architecture to map music to latent representation, then decodes to 3D dance motions. Includes music-dance retrieval framework for evaluation.

Result: Extensive experiments on FineDance dataset demonstrate state-of-the-art performance in music-to-dance generation.

Conclusion: MatchDance effectively addresses choreographic consistency in music-to-dance generation through its novel two-stage framework with kinematic-dynamic quantization and hybrid architecture.

Abstract: Music-to-dance generation represents a challenging yet pivotal task at the intersection of choreography, virtual reality, and creative content generation. Despite its significance, existing methods face substantial limitation in achieving choreographic consistency. To address the challenge, we propose MatchDance, a novel framework for music-to-dance generation that constructs a latent representation to enhance choreographic consistency. MatchDance employs a two-stage design: (1) a Kinematic-Dynamic-based Quantization Stage (KDQS), which encodes dance motions into a latent representation by Finite Scalar Quantization (FSQ) with kinematic-dynamic constraints and reconstructs them with high fidelity, and (2) a Hybrid Music-to-Dance Generation Stage(HMDGS), which uses a Mamba-Transformer hybrid architecture to map music into the latent representation, followed by the KDQS decoder to generate 3D dance motions. Additionally, a music-dance retrieval framework and comprehensive metrics are introduced for evaluation. Extensive experiments on the FineDance dataset demonstrate state-of-the-art performance.

Method: TRACE analyzes first-order dynamics of frozen speech foundation model representations. It hypothesizes that genuine speech forms smooth embedding trajectories while splice boundaries cause abrupt disruptions. The framework extracts frame-level embeddings, computes transition dynamics, and detects anomalies without any training or architectural changes.

Result: On PartialSpoof benchmark: 8.08% EER (competitive with fine-tuned supervised baselines). On LlamaPartialSpoof (most challenging with LLM-driven commercial synthesis): 24.12% EER vs 24.49% for supervised baseline, surpassing it without target-domain data. Evaluated on 4 benchmarks across 2 languages using 6 speech foundation models.

Conclusion: Temporal dynamics in speech foundation models provide an effective, generalizable signal for training-free audio forensics. The approach eliminates need for supervision, labeled data, and retraining while maintaining competitive performance.

Abstract: Partial audio deepfakes, where synthesized segments are spliced into genuine recordings, are particularly deceptive because most of the audio remains authentic. Existing detectors are supervised: they require frame-level annotations, overfit to specific synthesis pipelines, and must be retrained as new generative models emerge. We argue that this supervision is unnecessary. We hypothesize that speech foundation models implicitly encode a forensic signal: genuine speech forms smooth, slowly varying embedding trajectories, while splice boundaries introduce abrupt disruptions in frame-level transitions. Building on this, we propose TRACE (Training-free Representation-based Audio Countermeasure via Embedding dynamics), a training-free framework that detects partial audio deepfakes by analyzing the first-order dynamics of frozen speech foundation model representations without any training, labeled data, or architectural modification. We evaluate TRACE on four benchmarks that span two languages using six speech foundation models. In PartialSpoof, TRACE achieves 8.08% EER, competitive with fine-tuned supervised baselines. In LlamaPartialSpoof, the most challenging benchmark featuring LLM-driven commercial synthesis, TRACE surpasses a supervised baseline outright (24.12% vs. 24.49% EER) without any target-domain data. These results show that temporal dynamics in speech foundation models provide an effective, generalize signal for training-free audio forensics.

Method: FineLAP introduces a dual-stream sigmoid loss with cluster-based sampling to jointly learn from clip- and frame-level supervision. It uses a decoupled audio projector on top of a self-supervised encoder to capture both global semantics and local details. The method also includes FineLAP-100k, a large-scale synthetic sound event detection dataset created through scalable curation.

Result: FineLAP achieves state-of-the-art performance across multiple audio understanding tasks including retrieval, classification, sound event detection, and text-to-audio grounding. Ablation studies show that coarse- and fine-grained alignment are mutually beneficial.

Conclusion: FineLAP provides an effective training paradigm for audio-language models that advances both clip- and frame-level alignment, offering insights for building better multimodal models that can handle varying granularity of audio-text data.

Abstract: Contrastively pretrained audio-language models (e.g., CLAP) excel at clip-level understanding but struggle with frame-level tasks. Existing extensions fail to exploit the varying granularity of real-world audio-text data, where massive clip-level textual descriptions coexist with limited frame-level annotations. This paper proposes Fine-grained Language-Audio Pretraining (FineLAP), a novel training paradigm that advances both clip- and frame-level alignment in CLAP with heterogeneous data. FineLAP introduces a dual-stream sigmoid loss with a cluster-based sampling strategy to jointly learn from clip- and frame-level supervision. To capture both global semantics and local details, FineLAP uses a decoupled audio projector on top of a self-supervised encoder. To alleviate the scarcity of temporally annotated data, we present FineLAP-100k, a large-scale synthetic SED dataset constructed through a scalable curation pipeline. Extensive experiments demonstrate that FineLAP achieves SOTA performance across multiple audio understanding tasks, including retrieval, classification, sound event detection, and text-to-audio grounding. Ablation studies further show that coarse- and fine-grained alignment are mutually beneficial, providing insights for building better audio-language models (ALMs).

Method: Two contributions: 1) InterpTRQE-SptME benchmark that directly measures residual speaker information in content embeddings using SHAP-based interpretability analysis, and 2) InterpTF-SptME filtering method that uses interpretability insights to remove speaker information from embeddings.

Result: Testing on VCTK with seven models (HuBERT, WavLM, ContentVec) shows SHAP Noise filtering reduces speaker residuals from 18.05% to nearly zero while maintaining recognition accuracy (CTC loss increase under 1%). The method is model-agnostic and requires no retraining.

Conclusion: The proposed benchmark and filtering method effectively address speaker-content entanglement in self-supervised speech models, enabling better content-focused representations while preserving privacy through speaker information removal.

Abstract: Self-supervised speech models learn representations that capture both content and speaker information. Yet this entanglement creates problems: content tasks suffer from speaker bias, and privacy concerns arise when speaker identity leaks through supposedly anonymized representations. We present two contributions to address these challenges. First, we develop InterpTRQE-SptME (Timbre Residual Quantitative Evaluation Benchmark of Speech pre-training Models Encoding via Interpretability), a benchmark that directly measures residual speaker information in content embeddings using SHAP-based interpretability analysis. Unlike existing indirect metrics, our approach quantifies the exact proportion of speaker information remaining after disentanglement. Second, we propose InterpTF-SptME, which uses these interpretability insights to filter speaker information from embeddings. Testing on VCTK with seven models including HuBERT, WavLM, and ContentVec, we find that SHAP Noise filtering reduces speaker residuals from 18.05% to nearly zero while maintaining recognition accuracy (CTC loss increase under 1%). The method is model-agnostic and requires no retraining.

Method: Uses staged disentanglement: pretrains semantic encoder, regularizes with multi-task supervision, freezes encoder, applies hard dual-codebook routing with auxiliary objectives on quantized codes. Uses diffusion decoder for high-frequency detail reconstruction.

Result: Achieves favorable predictability-fidelity trade-off with lowest cnBPT while maintaining competitive reconstruction at 0.75 kbps. Improves enBPT under dual-track language modeling, showing gains beyond codebook size effects. Models use cross-track structure and non-local history.

Conclusion: DuoTok successfully addresses the audio tokenization trade-off through staged disentanglement and diffusion decoding, enabling better multi-track music language modeling with improved cross-track correspondence and predictability.

Abstract: Audio tokenization bridges continuous waveforms and multi-track music language models. In dual-track modeling, tokens should preserve three properties at once: high-fidelity reconstruction, strong predictability under a language model, and cross-track correspondence. We introduce DuoTok, a source-aware dual-track tokenizer that addresses this trade-off through staged disentanglement. DuoTok first pretrains a semantic encoder, then regularizes it with multi-task supervision, freezes the encoder, and applies hard dual-codebook routing while keeping auxiliary objectives on quantized codes. A diffusion decoder reconstructs high-frequency details, allowing tokens to focus on structured information for sequence modeling. On standard benchmarks, DuoTok achieves a favorable predictability-fidelity trade-off, reaching the lowest cnBPT while maintaining competitive reconstruction at 0.75 kbps. Under a held-constant dual-track language modeling protocol, enBPT also improves, indicating gains beyond codebook size effects. Controlled diagnostics show larger predictability costs under cross-track corruption and larger gains from longer context, suggesting that models trained on DuoTok tokens use cross-track structure and non-local history.

Method: Fair-Gate framework uses risk extrapolation to reduce variation in speaker-classification risk across proxy sex groups, and introduces a local complementary gate that routes intermediate features into separate identity and sex branches. The gate provides interpretability through explicit routing masks showing which features are allocated to identity vs. sex-related pathways.

Result: Experiments on VoxCeleb1 show that Fair-Gate improves the utility-fairness trade-off, yielding more sex-fair automatic speaker verification (ASV) performance under challenging evaluation conditions.

Conclusion: Fair-Gate effectively addresses both demographic shortcut learning and feature entanglement in voice biometric systems through an interpretable risk-gating framework that improves fairness while maintaining utility.

Abstract: Voice biometric systems can exhibit sex-related performance gaps even when overall verification accuracy is strong. We attribute these gaps to two practical mechanisms: (i) demographic shortcut learning, where speaker classification training exploits spurious correlations between sex and speaker identity, and (ii) feature entanglement, where sex-linked acoustic variation overlaps with identity cues and cannot be removed without degrading speaker discrimination. We propose Fair-Gate, a fairness-aware and interpretable risk-gating framework that addresses both mechanisms in a single pipeline. Fair-Gate applies risk extrapolation to reduce variation in speaker-classification risk across proxy sex groups, and introduces a local complementary gate that routes intermediate features into an identity branch and a sex branch. The gate provides interpretability by producing an explicit routing mask that can be inspected to understand which features are allocated to identity versus sex-related pathways. Experiments on VoxCeleb1 show that Fair-Gate improves the utility–fairness trade-off, yielding more sex-fair ASV performance under challenging evaluation conditions.

Method: CoDeTT formulates turn-taking as a structured decision problem, constructs a multi-scenario dataset with fine-grained decision categories and controlled context variations, and establishes a unified evaluation protocol.

Result: Evaluation of representative models reveals substantial performance disparities across decision types and interaction scenarios, demonstrating the benchmark’s ability to reveal model weaknesses.

Conclusion: CoDeTT provides a standardized benchmark for systematic and context-aware evaluation of turn-taking systems, with publicly available dataset and evaluation toolkit.

Abstract: Turn-taking modeling is fundamental to spoken dialogue systems, yet its evaluation remains fragmented and often limited to binary boundary detection under narrow interaction settings. Such protocols hinder systematic comparison and obscure model weaknesses across conversational conditions. We present CoDeTT, a context-aware decision benchmark for turn-taking evaluation. CoDeTT formulates turn-taking as a structured decision problem and constructs a multi-scenario dataset with fine-grained decision categories and controlled context variations. Under a unified evaluation protocol, we assess representative existing models and observe substantial performance disparities across decision types and interaction scenarios. CoDeTT provides a standardized benchmark for systematic and context-aware evaluation of turn-taking systems. The benchmark dataset and evaluation toolkit are available at https://yingaowang-casia.github.io/CoDeTT.github.io/.

Method: Proposes an optimizer-aware framework viewing online selection as shaping next target-oriented update under optimizer state. Formulates as optimizer-aware update-matching problem, connects to second-order target utility. Develops two-stage Filter-then-Weight algorithm with factorized outer-product gradient representation and optimized matrix computations for long-context data.

Result: Experiments show the method consistently improves convergence and downstream performance over existing online data selection baselines under the same data budget.

Conclusion: The optimizer-aware framework effectively addresses online data selection challenges in LLM fine-tuning by accounting for adaptive optimizer states and step-dependent sample utility.

Abstract: Gradient-based data selection offers a principled framework for estimating sample utility in large language model (LLM) fine-tuning, but existing methods are mostly designed for offline settings. They are therefore less suited to online fine-tuning, where data arrives sequentially, sample utility is step-dependent, and the effective update geometry is shaped by adaptive optimizers. We propose an optimizer-aware framework for gradient-based online data selection and reweighting in LLM fine-tuning. Our key idea is to view online selection not as static sample ranking, but as shaping the next target-oriented update under the optimizer state. We formulate this as an optimizer-aware update-matching problem, establish its connection to second-order target utility, and show why subset-level construction must account for interactions and redundancy among selected samples. Based on this view, we develop a two-stage Filter-then-Weight algorithm that first filters geometrically useful candidates and then optimizes their coefficients. To make the framework practical for LLMs, we introduce a factorized outer-product gradient representation and optimized matrix computations for long-context data. Experiments show that our method consistently improves convergence and downstream performance over existing online data selection baselines under the same data budget.

Method: FedRouter uses adapters to personalize models with two clustering mechanisms: local clustering associates adapters with task data samples, and global clustering associates similar adapters from different clients to construct task-centric personalized models. An evaluation router mechanism routes test samples to the best adapter based on created clusters.

Result: FedRouter demonstrates strong resilience in challenging scenarios, performing up to 6.1% relatively better under tasks interference and up to 136% relative improvement under generalization evaluation compared to existing approaches across a multitask dataset.

Conclusion: FedRouter effectively addresses generalization and task interference challenges in personalized federated learning by building task-centric models through clustering mechanisms, offering improved robustness over client-centric approaches.

Abstract: Federated Learning (FL) has emerged as a promising technique for training language models on distributed and private datasets of diverse tasks. However, aggregating models trained on heterogeneous tasks often degrades the overall performance of individual clients. To address this issue, Personalized FL (pFL) aims to create models tailored for each client’s data distribution. Although these approaches improve local performance, they usually lack robustness in two aspects: (i) generalization: when clients must make predictions on unseen tasks, or face changes in their data distributions, and (ii) intra-client tasks interference: when a single client’s data contains multiple distributions that may interfere with each other during local training. To tackle these two challenges, we propose FedRouter, a clustering-based pFL that builds specialized models for each task rather than for each client. FedRouter uses adapters to personalize models by employing two clustering mechanisms to associate adapters with specific tasks. A local clustering that associate adapters with task data samples and a global one that associates similar adapters from different clients to construct task-centric personalized models. Additionally, we propose an evaluation router mechanism that routes test samples to the best adapter based on the created clusters. Experiments comparing our method with existing approaches across a multitask dataset, FedRouter demonstrate strong resilience in these challenging scenarios performing up to 6.1% relatively better under tasks interference and up to 136% relative improvement under generalization evaluation.

Method: Comparative study of ES and DRL across multiple environments: Flappy Bird (simple), Breakout (moderate), and MuJoCo Walker (complex). Also tested ES as initial training phase for DRL algorithms.

Result: ES did not consistently train faster than DRL. When used as pre-training for DRL, ES only provided benefits in simple environments (Flappy Bird) and showed minimal/no improvement in training efficiency or stability for more complex tasks (Breakout and MuJoCo Walker).

Conclusion: ES are not superior to DRL in training speed, and their utility as pre-training for DRL is limited to simple environments, questioning their suitability for demanding reinforcement learning scenarios.

Abstract: Although Deep Reinforcement Learning has proven highly effective for complex decision-making problems, it demands significant computational resources and careful parameter adjustment in order to develop successful strategies. Evolution strategies offer a more straightforward, derivative-free approach that is less computationally costly and simpler to deploy. However, ES generally do not match the performance levels achieved by DRL, which calls into question their suitability for more demanding scenarios. This study examines the performance of ES and DRL across tasks of varying difficulty, including Flappy Bird, Breakout and Mujoco environments, as well as whether ES could be used for initial training to enhance DRL algorithms. The results indicate that ES do not consistently train faster than DRL. When used as a preliminary training step, they only provide benefits in less complex environments (Flappy Bird) and show minimal or no improvement in training efficiency or stability across different parameter settings when applied to more sophisticated tasks (Breakout and MuJoCo Walker).

Method: Memory is modeled as a Bridge Diffusion stochastic process on replay interval [0,1], with terminal marginals encoding present and intermediate marginals encoding past. Uses Compress-Add-Smooth recursion with Gaussian mixture models, controlled by piecewise-linear protocol segments.

Result: Retention half-life scales linearly with protocol segments (a₁/₂ ≈ cL), with constant c > 1 independent of mixture complexity, dimension, or target geometry. Provides temporally coherent replay and analytical study of forgetting mechanisms.

Conclusion: The framework offers a mathematically precise, fully analytical model of continual learning where forgetting arises from lossy temporal compression rather than parameter interference, enabling study of forgetting mechanisms without neural networks.

Abstract: An agent that operates sequentially must incorporate new experience without forgetting old experience, under a fixed memory budget. We propose a framework in which memory is not a parameter vector but a stochastic process: a Bridge Diffusion on a replay interval $[0,1]$, whose terminal marginal encodes the present and whose intermediate marginals encode the past. New experience is incorporated via a three-step \emph{Compress–Add–Smooth} (CAS) recursion. We test the framework on the class of models with marginal probability densities modeled via Gaussian mixtures of fixed number of components~$K$ in $d$ dimensions; temporal complexity is controlled by a fixed number~$L$ of piecewise-linear protocol segments whose nodes store Gaussian-mixture states. The entire recursion costs $O(LKd^2)$ flops per day – no backpropagation, no stored data, no neural networks – making it viable for controller-light hardware. Forgetting in this framework arises not from parameter interference but from lossy temporal compression: the re-approximation of a finer protocol by a coarser one under a fixed segment budget. We find that the retention half-life scales linearly as $a_{1/2}\approx c,L$ with a constant $c>1$ that depends on the dynamics but not on the mixture complexity~$K$, the dimension~$d$, or the geometry of the target family. The constant~$c$ admits an information-theoretic interpretation analogous to the Shannon channel capacity. The stochastic process underlying the bridge provides temporally coherent movie'' replay -- compressed narratives of the agent's history, demonstrated visually on an MNIST latent-space illustration. The framework provides a fully analytical Ising model’’ of continual learning in which the mechanism, rate, and form of forgetting can be studied with mathematical precision.

Method: Examination of resource considerations across the AI-driven discovery pipeline, highlighting emerging strategies including general-purpose ML models, multi-fidelity approaches, model distillation, active learning, and physics-based hierarchical workflows.

Result: Identifies key sustainability challenges and proposes strategies for enhancing efficiency, including the importance of bridging computational predictions with real-world conditions and advocating for open data/models and reusable workflows.

Conclusion: Sustainable progress requires domain-specific AI systems that maximize scientific value per unit of computation, enabling efficient and responsible discovery of technological materials and therapeutics through open data, reusable workflows, and optimized resource use.

Abstract: Artificial intelligence is transforming molecular and materials science, but its growing computational and data demands raise critical sustainability challenges. In this Perspective, we examine resource considerations across the AI-driven discovery pipeline–from quantum-mechanical (QM) data generation and model training to automated, self-driving research workflows–building on discussions from the ``SusML workshop: Towards sustainable exploration of chemical spaces with machine learning’’ held in Dresden, Germany. In this context, the availability of large quantum datasets has enabled rigorous benchmarking and rapid methodological progress, while also incurring substantial energy and infrastructure costs. We highlight emerging strategies to enhance efficiency, including general-purpose machine learning (ML) models, multi-fidelity approaches, model distillation, and active learning. Moreover, incorporating physics-based constraints within hierarchical workflows, where fast ML surrogates are applied broadly and high-accuracy QM methods are used selectively, can further optimize resource use without compromising reliability. Equally important is bridging the gap between idealized computational predictions and real-world conditions by accounting for synthesizability and multi-objective design criteria, which is essential for practical impact. Finally, we argue that sustainable progress will rely on open data and models, reusable workflows, and domain-specific AI systems that maximize scientific value per unit of computation, enabling efficient and responsible discovery of technological materials and therapeutics.

Method: Empirical testing of 18 classifier configurations (MLPs, SVMs, random forests, k-NN, Bayesian classifiers, deep networks) on self-improving neural controllers and MuJoCo benchmarks, plus three safe RL baselines. Then proposed Lipschitz ball verification with provable analytical bounds and ball chaining for parameter-space traversal.

Result: All classifier configurations and safe RL baselines fail to maintain safety oversight across various dimensions and distribution separations. Lipschitz ball verification achieves zero false accepts across dimensions up to 17408, enabling safe parameter-space traversal with provable safety guarantees.

Conclusion: Classifier-based safety gates are fundamentally inadequate for safe AI self-improvement, but Lipschitz verification provides a viable alternative with provable safety guarantees, enabling safe traversal of parameter spaces during AI system improvement.

Abstract: Can classifier-based safety gates maintain reliable oversight as AI systems improve over hundreds of iterations? We provide comprehensive empirical evidence that they cannot. On a self-improving neural controller (d=240), eighteen classifier configurations – spanning MLPs, SVMs, random forests, k-NN, Bayesian classifiers, and deep networks – all fail the dual conditions for safe self-improvement. Three safe RL baselines (CPO, Lyapunov, safety shielding) also fail. Results extend to MuJoCo benchmarks (Reacher-v4 d=496, Swimmer-v4 d=1408, HalfCheetah-v4 d=1824). At controlled distribution separations up to delta_s=2.0, all classifiers still fail – including the NP-optimal test and MLPs with 100% training accuracy – demonstrating structural impossibility. We then show the impossibility is specific to classification, not to safe self-improvement itself. A Lipschitz ball verifier achieves zero false accepts across dimensions d in {84, 240, 768, 2688, 5760, 9984, 17408} using provable analytical bounds (unconditional delta=0). Ball chaining enables unbounded parameter-space traversal: on MuJoCo Reacher-v4, 10 chains yield +4.31 reward improvement with delta=0; on Qwen2.5-7B-Instruct during LoRA fine-tuning, 42 chain transitions traverse 234x the single-ball radius with zero safety violations across 200 steps. A 50-prompt oracle confirms oracle-agnosticity. Compositional per-group verification enables radii up to 37x larger than full-network balls. At d<=17408, delta=0 is unconditional; at LLM scale, conditional on estimated Lipschitz constants.

Method: Population-Adaptive Symbolic Mixture-of-Experts (PASM) combines large language model guided symbolic regression with mixture-of-experts architecture to discover human-readable closed-form decision rules, specialize them to data-driven subpopulations, and route inputs to appropriate experts.

Result: PASM achieves Matthews correlation coefficient of 0.607 when transferring from Florida/Texas to Georgia with 100 calibration samples, outperforming XGBoost (0.404), TabPFN (0.333), GPT-5-mini (0.434), and meta-learning baselines (MCC ≤ 0.346).

Conclusion: PASM closes more than half the cross-location generalization gap while maintaining transparent, interpretable decision rules suitable for real-world emergency planning, with no statistically significant fairness disparities across demographic axes.

Abstract: Accurate prediction of evacuation behavior is critical for disaster preparedness, yet models trained in one region often fail elsewhere. Using a multi-state hurricane evacuation survey, we show this failure goes beyond feature distribution shift: households with similar characteristics follow systematically different decision patterns across states. As a result, single global models overfit dominant responses, misrepresent vulnerable subpopulations, and generalize poorly across locations. We propose Population-Adaptive Symbolic Mixture-of-Experts (PASM), which pairs large language model guided symbolic regression with a mixture-of-experts architecture. PASM discovers human-readable closed-form decision rules, specializes them to data-driven subpopulations, and routes each input to the appropriate expert at inference time. On Hurricanes Harvey and Irma data, transferring from Florida and Texas to Georgia with 100 calibration samples, PASM achieves a Matthews correlation coefficient of 0.607, compared to XGBoost (0.404), TabPFN (0.333), GPT-5-mini (0.434), and meta-learning baselines MAML and Prototypical Networks (MCC $\leq$ 0.346). The routing mechanism assigns distinct formula archetypes to subpopulations, so the resulting behavioral profiles are directly interpretable. A fairness audit across four demographic axes finds no statistically significant disparities after Bonferroni correction. PASM closes more than half the cross-location generalization gap while keeping decision rules transparent enough for real-world emergency planning.

Method: Uses LLM to dynamically generate curriculum over available actions, creating multi-stage training path that progressively introduces complex actions to Tabular Q-Learning and DQN agents in Blackjack simulation.

Result: Significant performance gains: DQN agent’s average win rate increased from 43.97% to 47.41%, bust rate reduced from 32.9% to 28.0%, workflow accelerated by over 74%, with full training completing faster than baseline’s evaluation phase alone.

Conclusion: LLM-guided curricula can build more effective, robust, and efficient RL agents by providing structured learning paths.

Abstract: Reinforcement Learning (RL) agents often struggle with efficiency and performance in complex environments. We propose a novel framework that uses a Large Language Model (LLM) to dynamically generate a curriculum over available actions, enabling the agent to incorporate each action individually. We apply this framework to the game of Blackjack, where the LLM creates a multi-stage training path that progressively introduces complex actions to a Tabular Q-Learning and a Deep Q-Network (DQN) agent. Our evaluation in a realistic 8-deck simulation over 10 independent runs demonstrates significant performance gains over standard training methods. The curriculum-based approach increases the DQN agent’s average win rate from 43.97% to 47.41%, reduces the average bust rate from 32.9% to 28.0%, and accelerates the overall workflow by over 74%, with the agent’s full training completing faster than the baseline’s evaluation phase alone. These results validate that LLM-guided curricula can build more effective, robust, and efficient RL agents.

Method: GT-PD combines gradient tracking with two defense layers: 1) universal self-centered projection that clips incoming messages to a ball around the receiving agent, and 2) decentralized probabilistic dropout based on dual-metric trust scores in decision and tracking channels. GT-PD-L adds leaky integration to control tracking error accumulation from persistent perturbations.

Result: With complete Byzantine isolation, GT-PD converges linearly to a neighborhood determined by stochastic gradient variance. Under partial isolation, GT-PD-L achieves linear convergence to a bounded neighborhood. Experiments on MNIST show GT-PD-L outperforms coordinate-wise trimmed mean by up to 4.3 percentage points under stealth attacks.

Conclusion: The proposed method preserves doubly stochastic mixing structure while providing Byzantine resilience, achieving linear convergence with bounded error neighborhoods under various attack scenarios.

Abstract: We study distributed optimization over networks with Byzantine agents that may send arbitrary adversarial messages. We propose \emph{Gradient Tracking with Probabilistic Edge Dropout} (GT-PD), a stochastic gradient tracking method that preserves the convergence properties of gradient tracking under adversarial communication. GT-PD combines two complementary defense layers: a universal self-centered projection that clips each incoming message to a ball of radius $τ$ around the receiving agent, and a fully decentralized probabilistic dropout rule driven by a dual-metric trust score in the decision and tracking channels. This design bounds adversarial perturbations while preserving the doubly stochastic mixing structure, a property often lost under robust aggregation in decentralized settings. Under complete Byzantine isolation ($p_b=0$), GT-PD converges linearly to a neighborhood determined solely by stochastic gradient variance. For partial isolation ($p_b>0$), we introduce \emph{Gradient Tracking with Probabilistic Edge Dropout and Leaky Integration} (GT-PD-L), which uses a leaky integrator to control the accumulation of tracking errors caused by persistent perturbations and achieves linear convergence to a bounded neighborhood determined by the stochastic variance and the clipping-to-leak ratio. We further show that under two-tier dropout with $p_h=1$, isolating Byzantine agents introduces no additional variance into the honest consensus dynamics. Experiments on MNIST under Sign Flip, ALIE, and Inner Product Manipulation attacks show that GT-PD-L outperforms coordinate-wise trimmed mean by up to 4.3 percentage points under stealth attacks.

Method: Use sparse learning methods to build models that approximate strong branching scores. Focus on creating models with fewer than 4% of parameters compared to state-of-the-art graph neural networks while maintaining competitive accuracy.

Result: CPU-only models are faster than both SCIP’s built-in branching rules and GPU-accelerated GNN models. Models remain effective with small training sets and demonstrate efficiency across diverse problem classes.

Conclusion: Sparse learning provides a practical alternative to deep learning for branching decisions in mixed-integer programming, offering interpretability, efficiency, and effectiveness in low-resource settings.

Abstract: Machine learning is increasingly used to improve decisions within branch-and-bound algorithms for mixed-integer programming. Many existing approaches rely on deep learning, which often requires very large training datasets and substantial computational resources for both training and deployment, typically with GPU parallelization. In this work, we take a different path by developing interpretable models that are simple but effective. We focus on approximating strong branching (SB) scores, a highly effective yet computationally expensive branching rule. Using sparse learning methods, we build models with fewer than 4% of the parameters of a state-of-the-art graph neural network (GNN) while achieving competitive accuracy. Relative to SCIP’s built-in branching rules and the GNN-based model, our CPU-only models are faster than the default solver and the GPU-accelerated GNN. The models are simple to train and deploy, and they remain effective with small training sets, which makes them practical in low-resource settings. Extensive experiments across diverse problem classes demonstrate the efficiency of this approach.

Method: Uses vision transformer-based framework with autoregressive learning of physical and frequency embeddings. Incorporates Fourier transforms in latent space to capture wave number spectra. Trained on data from high-fidelity Finite Volume simulations with variations in initial conditions and material properties (speed of light).

Result: ML model achieves adequate relative errors below 10% for over 75 time step rollouts, despite material discontinuities and unknown properties. Prediction errors show linear growth over time with sharp increase at material interface.

Conclusion: The vision transformer-based ML surrogate with Fourier transforms effectively approximates Maxwell’s equations solutions for wave-material interactions, demonstrating potential for efficient simulation of electromagnetic wave propagation through interfaces.

Abstract: We develop a machine learning (ML) surrogate model to approximate solutions to Maxwell’s equations in one dimension, focusing on scenarios involving a material interface that reflects and transmits electro-magnetic waves. Derived from high-fidelity Finite Volume (FV) simulations, our training data includes variations of the initial conditions, as well as variations in one material’s speed of light, allowing for the model to learn a range of wave-material interaction behaviors. The ML model autoregressively learns both the physical and frequency embeddings in a vision transformer-based framework. By incorporating Fourier transforms in the latent space, the wave number spectra of the solutions aligns closely with the simulation data. Prediction errors exhibit an approximately linear growth over time with a sharp increase at the material interface. Test results show that the ML solution has adequate relative errors below $10%$ in over $75$ time step rollouts, despite the presence of the discontinuity and unknown material properties.

Method: Proposed algorithm uses round-based exploration with recursive region elimination guided by an adaptive reference arm. Employs new analytical tools for relative feedback in continuous spaces.

Result: Proves regret bound of $\tilde O\left(T^{\frac{d_z+1}{d_z+2}}\right)$ where $d_z$ is zooming dimension, with only logarithmic space complexity in time horizon.

Conclusion: First solution for Lipschitz dueling bandits, achieving optimal space complexity and sublinear regret, enabling preference learning in continuous domains with comparative feedback.

Abstract: We study for the first time, stochastic dueling bandits over continuous action spaces with Lipschitz structure, where feedback is purely comparative. While dueling bandits and Lipschitz bandits have been studied separately, their combination has remained unexplored. We propose the first algorithm for Lipschitz dueling bandits, using round-based exploration and recursive region elimination guided by an adaptive reference arm. We develop new analytical tools for relative feedback and prove a regret bound of $\tilde O\left(T^{\frac{d_z+1}{d_z+2}}\right)$, where $d_z$ is the zooming dimension of the near-optimal region. Further, our algorithm takes only logarithmic space in terms of the total time horizon, best achievable by any bandit algorithm over a continuous action space.

Method: Uses cost-aware contextual bandits with three mechanisms: 1) online primal-dual budget pacer for per-request cost ceilings, 2) geometric forgetting on sufficient statistics for rapid adaptation to shifts, and 3) hot-swap registry for runtime model integration with forced exploration for newcomers.

Result: Across 1,824 prompts with 7 budget ceilings, mean per-request cost never exceeded target by more than 0.4%. System adapts to price cuts (+0.071 quality lift), detects quality regressions, and integrates new models within ~142 steps. Routing latency is 9.8ms on CPU.

Conclusion: ParetoBandit effectively solves the dynamic routing problem for multi-model LLM portfolios by enforcing budgets, adapting to non-stationary conditions, and supporting runtime model integration with minimal overhead.

Abstract: Production LLM serving often relies on multi-model portfolios spanning a ~530x cost range, where routing decisions trade off quality against cost. This trade-off is non-stationary: providers revise pricing, model quality can regress silently, and new models must be integrated without downtime. We present ParetoBandit, an open-source adaptive router built on cost-aware contextual bandits that is the first to simultaneously enforce dollar-denominated budgets, adapt online to such shifts, and onboard new models at runtime. ParetoBandit closes these gaps through three mechanisms. An online primal-dual budget pacer enforces a per-request cost ceiling over an open-ended stream, replacing offline penalty tuning with closed-loop control. Geometric forgetting on sufficient statistics enables rapid adaptation to price and quality shifts while bootstrapping from offline priors. A hot-swap registry lets operators add or remove models at runtime, with a brief forced-exploration phase for each newcomer, after which UCB selection discovers its quality-cost niche from live traffic alone. We evaluate ParetoBandit across four deployment scenarios on 1,824 prompts routed through a three-model portfolio. Across seven budget ceilings, mean per-request cost never exceeds the target by more than 0.4%. When conditions shift, the system adapts: an order-of-magnitude price cut on the costliest model yields up to +0.071 quality lift, and a silent quality regression is detected and rerouted within budget. A cold-started model reaches meaningful adoption within ~142 steps without breaching the cost ceiling. The router discriminates rather than blindly adopting: expensive models are budget-gated and low-quality models rejected after bounded exploration. End-to-end routing latency is 9.8ms on CPU – less than 0.4% of typical inference time – with the routing decision itself taking just 22.5us.

Method: Decompose EEG signals into five frequency bands (delta, theta, alpha, lower beta, higher beta), extract 11 features per band, then use Graph Convolutional Neural Network (GCN) to model spatial dependencies among EEG electrodes represented as graph nodes.

Result: Achieved accuracies of 97.1%, 97.13%, 99.5%, 99.7%, and 51.4% across respective frequency bands on CHB-MIT dataset, with overall broadband accuracy of 99.01%. Mid-frequency bands showed strongest discriminative capability.

Conclusion: The frequency-aware approach improves interpretability and diagnostic precision compared to conventional broadband EEG methods, revealing frequency-specific seizure patterns and providing neurophysiologically relevant insights.

Abstract: Epileptic seizures are neurological disorders characterized by abnormal and excessive electrical activity in the brain, resulting in recurrent seizure events. Electroencephalogram (EEG) signals are widely used for seizure diagnosis due to their ability to capture temporal and spatial neural dynamics. While recent deep learning methods have achieved high detection accuracy, they often lack interpretability and neurophysiological relevance. This study presents a frequency-aware framework for epileptic seizure detection based on ictal-phase EEG analysis. The raw EEG signals are decomposed into five frequency bands (delta, theta, alpha, lower beta, and higher beta), and eleven discriminative features are extracted from each band. A graph convolutional neural network (GCN) is then employed to model spatial dependencies among EEG electrodes, represented as graph nodes. Experiments on the CHB-MIT scalp EEG dataset demonstrate high detection performance, achieving accuracies of 97.1%, 97.13%, 99.5%, 99.7%, and 51.4% across the respective frequency bands, with an overall broadband accuracy of 99.01%. The results highlight the strong discriminative capability of mid-frequency bands and reveal frequency-specific seizure patterns. The proposed approach improves interpretability and diagnostic precision compared to conventional broadband EEG-based methods.

Method: Uses Smart Lacelock sensor (shoe-mounted device with load cell, accelerometer, gyroscope) to capture multimodal signals during Sit-to-Stand tasks. Extracted features from sensor data to train four machine learning classifiers using 4-fold participant-independent cross-validation for transition classification and duration measurement.

Result: Bagged tree classifier achieved 0.98 accuracy and 0.8 F1 score for Sit-to-Stand transition classification. Duration measurement had mean absolute error of 0.047 seconds with SD of 0.07 seconds for correctly classified transitions.

Conclusion: The Smart Lacelock sensor shows potential for real-world fall-risk assessment and mobility monitoring in older adults through accurate detection of Sit-to-Stand transitions.

Abstract: Postural stability during movement is fundamental to independent living, fall prevention, and overall health, particularly among older adults who experience age-related declines in balance, muscle strength, and mobility. Among daily functional activities, the Sit-to-Stand (SiSt) transition is a critical indicator of lower-limb strength, musculoskeletal health, and fall risk, making it an essential parameter for assessing functional capacity and monitoring physical decline in aging populations. This study presents a methodology SiSt transition detection and duration measurement using the Smart Lacelock sensor, a lightweight, shoe-mounted device that integrates a load cell, accelerometer, and gyroscope for motion analysis. The methodology was evaluated in 16 older adults (age: mean: 76.84, SD: 3.45 years) performing SiSt tasks within the Short Physical Performance Battery (SPPB) protocol. Features extracted from multimodal signals were used to train and evaluate four machine learning classifiers using a 4-fold participant-independent cross-validation to classify SiSt transitions and measure their duration. The bagged tree classifier achieved an accuracy of 0.98 and an F1 score of 0.8 in classifying SiSt transition. The mean absolute error in duration measurement of the correctly classified transitions was 0.047, and the SD was 0.07 seconds. These findings highlight the potential of the Smart Lacelock sensor for real-world fall-risk assessment and mobility monitoring in older adults.

Method: Replace Gaussian base distribution in normalizing flows with Lévy process-based distributions (Variance Gamma and Normal-Inverse Gaussian). Use identity-tail Neural Spline Flow architectures that preserve base distribution’s tail shape exactly outside transformation regions. Provide theoretical guarantees on tail behavior preservation.

Result: VG-based flows reduce test negative log-likelihood by 69% relative to Gaussian flows on S&P 500 daily returns, achieve exact 95% VaR calibration, and NIG-based flows provide most accurate Expected Shortfall estimates. Lévy-Flows show substantial improvements in density estimation and risk calibration.

Conclusion: Incorporating Lévy process structure into normalizing flows yields significant gains in modeling heavy-tailed data, with strong applications to financial risk management. The approach maintains exact likelihood evaluation while capturing realistic tail behavior.

Abstract: We introduce Lévy-Flows, a class of normalizing flow models that replace the standard Gaussian base distribution with Lévy process-based distributions, specifically Variance Gamma (VG) and Normal-Inverse Gaussian (NIG). These distributions naturally capture heavy-tailed behavior while preserving exact likelihood evaluation and efficient reparameterized sampling. We establish theoretical guarantees on tail behavior, showing that for regularly varying bases the tail index is preserved under asymptotically linear flow transformations, and that identity-tail Neural Spline Flow architectures preserve the base distribution’s tail shape exactly outside the transformation region. Empirically, we evaluate on S&P 500 daily returns and additional assets, demonstrating substantial improvements in density estimation and risk calibration. VG-based flows reduce test negative log-likelihood by 69% relative to Gaussian flows and achieve exact 95% VaR calibration, while NIG-based flows provide the most accurate Expected Shortfall estimates. These results show that incorporating Lévy process structure into normalizing flows yields significant gains in modeling heavy-tailed data, with applications to financial risk management.

Method: Proposes QUEry-modulated Spherical aTtention (QUEST) which constrains keys to a hyperspherical latent space while allowing individual tokens to control attention distribution sharpness. Can be used as drop-in replacement for standard attention.

Result: QUEST trains without instabilities, produces models with improved performance, and enhances robustness to data corruptions and adversarial attacks. Demonstrated effectiveness across vision applications and other domains.

Conclusion: QUEST provides a stable and effective alternative to standard attention mechanisms, addressing fundamental training instability issues while maintaining or improving model performance and robustness.

Abstract: The Transformer model architecture has become one of the most widely used in deep learning and the attention mechanism is at its core. The standard attention formulation uses a softmax operation applied to a scaled dot product between query and key vectors. We explore the role played by norms of the queries and keys, which can cause training instabilities when they arbitrarily increase. We demonstrate how this can happen even in simple Transformer models, in the presence of easy-to-learn spurious patterns in the data. We propose a new attention formulation, QUEry-modulated Spherical aTtention (QUEST), that constrains the keys to a hyperspherical latent space, while still allowing individual tokens to flexibly control the sharpness of the attention distribution. QUEST can be easily used as a drop-in replacement for standard attention. We focus on vision applications while also exploring other domains to highlight the method’s generality. We show that (1) QUEST trains without instabilities and (2) produces models with improved performance (3) that are robust to data corruptions and adversarial attacks.

Method: Estimate oracle-specific rewards from pairwise comparisons via maximum likelihood, formulate constrained objective as KL-regularized Lagrangian with Gibbs policy optimizer, solve convex dual problem with dual-only algorithm ensuring high-probability constraint satisfaction

Result: Provides first finite-sample performance guarantees for offline constrained preference learning, extends analysis to multiple constraints and general f-divergence regularization

Conclusion: Develops theoretically grounded framework for offline constrained RL from human preferences with statistical guarantees for safety/fairness constraints

Abstract: We study offline constrained reinforcement learning from human feedback with multiple preference oracles. Motivated by applications that trade off performance with safety or fairness, we aim to maximize target population utility subject to a minimum protected group welfare constraint. From pairwise comparisons collected under a reference policy, we estimate oracle-specific rewards via maximum likelihood and analyze how statistical uncertainty propagates through the dual program. We cast the constrained objective as a KL-regularized Lagrangian whose primal optimizer is a Gibbs policy, reducing learning to a convex dual problem. We propose a dual-only algorithm that ensures high-probability constraint satisfaction and provide the first finite-sample performance guarantees for offline constrained preference learning. Finally, we extend our theoretical analysis to accommodate multiple constraints and general f-divergence regularization.

Method: Parameterizes velocities as curl of vector potential to enforce mass conservation by construction. Uses cosine data-consistency loss for simultaneous denoising and unwrapping from wrapped phase images. Unsupervised neural network approach.

Result: Achieved lower errors than existing techniques (up to 11% lower velocity NRMSE, 11% lower directional error, 44% lower divergence). For unwrapping, achieved 0.18% and 5.2% residual wrapped voxels. Outperformed sequential pipelines in combined noise+aliasing scenarios.

Conclusion: DAF-FlowNet provides a unified framework for velocity enhancement and phase unwrapping that improves reliability of cardiovascular 4D Flow MRI, preserving fine-scale flow features and improving internal flow consistency.

Abstract: This work introduces an unsupervised Divergence and Aliasing-Free neural network (DAF-FlowNet) for 4D Flow Magnetic Resonance Imaging (4D Flow MRI) that jointly enhances noisy velocity fields and corrects phase wrapping artifacts. DAF-FlowNet parameterizes velocities as the curl of a vector potential, enforcing mass conservation by construction and avoiding explicit divergence-penalty tuning. A cosine data-consistency loss enables simultaneous denoising and unwrapping from wrapped phase images. On synthetic aortic 4D Flow MRI generated from computational fluid dynamics, DAF-FlowNet achieved lower errors than existing techniques (up to 11% lower velocity normalized root mean square error, 11% lower directional error, and 44% lower divergence relative to the best-performing alternative across noise levels), with robustness to moderate segmentation perturbations. For unwrapping, at peak velocity/velocity-encoding ratios of 1.4 and 2.1, DAF-FlowNet achieved 0.18% and 5.2% residual wrapped voxels, representing reductions of 72% and 18% relative to the best alternative method, respectively. In scenarios with both noise and aliasing, the proposed single-stage formulation outperformed a state-of-the-art sequential pipeline (up to 15% lower velocity normalized root mean square error, 11% lower directional error, and 28% lower divergence). Across 10 hypertrophic cardiomyopathy patient datasets, DAF-FlowNet preserved fine-scale flow features, corrected aliased regions, and improved internal flow consistency, as indicated by reduced inter-plane flow bias in aortic and pulmonary mass-conservation analyses recommended by the 4D Flow MRI consensus guidelines. These results support DAF-FlowNet as a framework that unifies velocity enhancement and phase unwrapping to improve the reliability of cardiovascular 4D Flow MRI.

Method: Uses an unpoled cube of Lead Zirconate Titanate (PZT) as a physical reservoir computing substrate to process MNIST handwritten digits and AudioMNIST spoken digits datasets, comparing performance against logistic regression baselines on the same preprocessed data.

Result: PZT reservoir achieves 89.0% accuracy on MNIST (2.4% improvement over logistic regression) but performs equivalently to baseline methods on AudioMNIST (88.2% vs 88.1%). The system shows greatest benefits for tasks of intermediate difficulty where linear methods underperform.

Conclusion: Physical reservoirs excel when task difficulty exceeds linear classifier capabilities but remains within reservoir computational capacity. PZT offers a low-power computational substrate that can integrate with digital algorithms for specific classification tasks.

Abstract: In this paper we extend our earlier work of (Rietman et al. 2022) presenting an application of physical Reservoir Computing (RC) to the classification of handwritten and spoken digits. We utilize an unpoled cube of Lead Zirconate Titanate (PZT) as a computational substrate to process these datasets. Our results demonstrate that the PZT reservoir achieves 89.0% accuracy on MNIST handwritten digits, representing a 2.4 percentage point improvement over logistic regression baselines applied to the same preprocessed data. However, for the AudioMNIST spoken digits dataset, the reservoir system (88.2% accuracy) performs equivalently to baseline methods (88.1% accuracy), suggesting that reservoir computing provides the greatest benefits for classification tasks of intermediate difficulty where linear methods underperform but the problem remains learnable. PZT is a well-known material already used in semiconductor applications, presenting a low-power computational substrate that can be integrated with digital algorithms. Our findings indicate that physical reservoirs excel when the task difficulty exceeds the capability of simple linear classifiers but remains within the computational capacity of the reservoir dynamics.

Method: Studies high-probability convergence of Decentralized SGD (DSGD) with light-tailed noise, develops technical results including variance-reduction effect of decentralized methods in HP sense and novel bound on MGF of strongly convex costs.

Result: DSGD converges in HP under same conditions as MSE convergence, achieves order-optimal rates for non-convex and strongly convex costs, establishes linear speed-up in number of users with matching or better transient times than MSE results.

Conclusion: First work showing DSGD achieves linear speed-up in HP sense, with relaxed assumptions and sharp rates enabled by novel technical results on variance reduction and MGF bounds, validated by experiments.

Abstract: Convergence in high-probability (HP) has attracted increasing interest, due to implying exponentially decaying tail bounds and strong guarantees for individual runs of an algorithm. While many works study HP guarantees in centralized settings, much less is understood in the decentralized setup, where existing works require strong assumptions, like uniformly bounded gradients, or asymptotically vanishing noise. This results in a significant gap between the assumptions used to establish convergence in the HP and the mean-squared error (MSE) sense, and is also contrary to centralized settings, where it is known that $\mathtt{SGD}$ converges in HP under the same conditions on the cost function as needed for MSE convergence. Motivated by these observations, we study the HP convergence of Decentralized $\mathtt{SGD}$ ($\mathtt{DSGD}$) in the presence of light-tailed noise, providing several strong results. First, we show that $\mathtt{DSGD}$ converges in HP under the same conditions on the cost as in the MSE sense, removing the restrictive assumptions used in prior works. Second, our sharp analysis yields order-optimal rates for both non-convex and strongly convex costs. Third, we establish a linear speed-up in the number of users, leading to matching, or strictly better transient times than those obtained from MSE results, further underlining the tightness of our analysis. To the best of our knowledge, this is the first work that shows $\mathtt{DSGD}$ achieves a linear speed-up in the HP sense. Our relaxed assumptions and sharp rates stem from several technical results of independent interest, including a result on the variance-reduction effect of decentralized methods in the HP sense, as well as a novel bound on the MGF of strongly convex costs, which is of interest even in centralized settings. Finally, we provide experiments that validate our theory.

Method: Derived closed-form expressions showing how superposition systematically deflates standard alignment metrics (Representational Similarity Analysis, Centered Kernel Alignment, linear regression) by making them dependent on cross-similarity between superposition matrices rather than latent features.

Result: Superposition causes networks with identical feature content to appear dissimilar, and under partial feature overlap can invert expected ordering (systems sharing fewer features appear more aligned than those sharing more).

Conclusion: Comparing neural systems in superposition requires extracting and aligning underlying features rather than comparing raw neural mixtures, as compressed sensing guarantees original features remain recoverable from lower-dimensional activity when sparse.

Abstract: Comparing the internal representations of neural networks is a central goal in both neuroscience and machine learning. Standard alignment metrics operate on raw neural activations, implicitly assuming that similar representations produce similar activity patterns. However, neural systems frequently operate in superposition, encoding more features than they have neurons via linear compression. We derive closed-form expressions showing that superposition systematically deflates Representational Similarity Analysis, Centered Kernel Alignment, and linear regression, causing networks with identical feature content to appear dissimilar. The root cause is that these metrics are dependent on cross-similarity between two systems’ respective superposition matrices, which under assumption of random projection usually differ significantly, not on the latent features themselves: alignment scores conflate what a system represents with how it represents it. Under partial feature overlap, this confound can invert the expected ordering, making systems sharing fewer features appear more aligned than systems sharing more. Crucially, the apparent misalignment need not reflect a loss of information; compressed sensing guarantees that the original features remain recoverable from the lower-dimensional activity, provided they are sparse. We therefore argue that comparing neural systems in superposition requires extracting and aligning the underlying features rather than comparing the raw neural mixtures.

Method: Analyzes RKL gradients by decomposing into target and non-target components, then proposes DRKL which removes the problematic gradient effect and strengthens non-target supervision while preserving RKL’s optimization benefits.

Result: Extensive experiments show DRKL consistently outperforms FKL, RKL, and other state-of-the-art distillation objectives, achieving better performance and superior fidelity-diversity trade-off across datasets and model families.

Conclusion: DRKL addresses key limitations of RKL distillation, providing a better balance between performance and output diversity for LLM knowledge distillation.

Abstract: Reverse Kullback-Leibler (RKL) divergence has recently emerged as the preferred objective for large language model (LLM) distillation, consistently outperforming forward KL (FKL), particularly in regimes with large vocabularies and significant teacher-student capacity mismatch, where RKL focuses learning on dominant modes rather than enforcing dense alignment. However, RKL introduces a structural limitation that drives the student toward overconfident predictions. We first provide an analysis of RKL by decomposing its gradients into target and non-target components, and show that non-target gradients consistently push the target logit upward even when the student already matches the teacher, thereby reducing output diversity. In addition, RKL provides weak supervision over non-target classes, leading to poor tail alignment. To address these issues, we propose Diversity-aware RKL (DRKL), which removes this gradient effect and strengthens non-target supervision while preserving the optimization benefits of RKL. Extensive experiments across datasets and model families demonstrate that DRKL consistently outperforms FKL, RKL, and other state-of-the-art distillation objectives, achieving better performance and a superior fidelity-diversity trade-off.

Method: Analyzed training dynamics across different architectures and datasets, measuring feature norm evolution and its relationship to NC onset. Conducted intervention experiments with feature scale perturbations and systematically varied architectural parameters (depth, width, activation functions, weight decay).

Result: Found that NC consistently occurs when mean feature norm reaches a critical value fn* specific to each model-dataset pair, with tight concentration (CV < 8%). The crossing of fn below fn* precedes NC onset by mean 62 epochs. Intervention experiments confirmed fn* as a stable attractor. Structural regularities revealed non-additive effects of architecture and dataset, with depth having non-monotonic effects, activation functions jointly determining collapse speed and fn*, weight decay creating three-regime phase diagram, and width monotonically accelerating collapse.

Conclusion: Feature-norm dynamics serve as an actionable diagnostic for predicting neural collapse timing, suggesting norm-threshold behavior is a general mechanism underlying delayed representational reorganization in deep networks.

Abstract: Neural collapse (NC) – the convergence of penultimate-layer features to a simplex equiangular tight frame – is well understood at equilibrium, but the dynamics governing its onset remain poorly characterised. We identify a simple and predictive regularity: NC occurs when the mean feature norm reaches a model-dataset-specific critical value, fn*, that is largely invariant to training conditions. This value concentrates tightly within each (model, dataset) pair (CV < 8%); training dynamics primarily affect the rate at which fn approaches fn*, rather than the value itself. In standard training trajectories, the crossing of fn below fn* consistently precedes NC onset, providing a practical predictor with a mean lead time of 62 epochs (MAE 24 epochs). A direct intervention experiment confirms fn* is a stable attractor of the gradient flow – perturbations to feature scale are self-corrected during training, with convergence to the same value regardless of direction (p>0.2). Completing the (architecture)x(dataset) grid reveals the paper’s strongest result: ResNet-20 on MNIST gives fn* = 5.867 – a +458% architecture effect versus only +68% on CIFAR-10. The grid is strongly non-additive; fn* cannot be decomposed into independent architecture and dataset contributions. Four structural regularities emerge: (1) depth has a non-monotonic effect on collapse speed; (2) activation jointly determines both collapse speed and fn*; (3) weight decay defines a three-regime phase diagram – too little slows, an optimal range is fastest, and too much prevents collapse; (4) width monotonically accelerates collapse while shifting fn* by at most 13%. These results establish feature-norm dynamics as an actionable diagnostic for predicting NC timing, suggesting that norm-threshold behaviour is a general mechanism underlying delayed representational reorganisation in deep networks.

Method: Three-stage approach: 1) Match stage performs pre-RoPE L2 matching over local window to find semantically similar queries; 2) Amend stage rectifies reused attention by recomputing small band near match boundary; 3) Complete stage fuses rectified results with fresh attention on KV tail through numerically stable merge. On match hit, compute and bandwidth complexity becomes constant regardless of context length.

Result: Reduces KV accesses by up to 99%, cuts token generation latency by over 60% at 128K context length, achieves over 14.3x attention-phase speedups and up to 2.6x end-to-end speedups while maintaining full-attention quality across LongBench v2 (120K), RULER (120K), and LongGenBench (16K continuous generation).

Conclusion: MAC-Attention provides a fidelity- and access-preserving acceleration method for long-context LLM inference that delivers both speed and faithfulness by reusing prior attention computations, making it model-agnostic and compatible with existing optimization techniques.

Abstract: Long-context decoding in LLMs is IO-bound: each token re-reads an ever-growing KV cache. Prior accelerations cut bytes via compression, which lowers fidelity, or selection/eviction, which restricts what remains accessible, and both can degrade delayed recall and long-form generation. We introduce MAC-Attention, a fidelity- and access-preserving alternative that accelerates decoding by reusing prior attention computations for semantically similar recent queries. It starts with a match stage that performs pre-RoPE L2 matching over a short local window; an amend stage rectifies the reused attention by recomputing a small band near the match boundary; and a complete stage fuses the rectified results with fresh attention computed on the KV tail through a numerically stable merge. On a match hit, the compute and bandwidth complexity is constant regardless of context length. The method is model-agnostic and composes with IO-aware kernels, paged-KV managers, and MQA/GQA. Across LongBench v2 (120K), RULER (120K), and LongGenBench (16K continuous generation), compared to the latest FlashInfer library, MAC-Attention reduces KV accesses by up to 99%, cuts token generation latency by over 60% at 128K, and achieves over 14.3x attention-phase speedups, up to 2.6x end-to-end, while maintaining full-attention quality. By reusing computation, MAC-Attention delivers long-context inference that is both fast and faithful. Code is available here: https://github.com/YJHMITWEB/MAC-Attention.git

Method: Introduces hierarchical framework that reduces node pair evaluations and uses discrete flow matching to decrease denoising iterations

Result: Empirically demonstrates better capture of graph distributions while substantially reducing generation time

Conclusion: Proposed hierarchical approach with discrete flow matching addresses computational bottlenecks in graph generation

Abstract: Denoising-based models, including diffusion and flow matching, have led to substantial advances in graph generation. Despite this progress, such models remain constrained by two fundamental limitations: a computational cost that scales quadratically with the number of nodes and a large number of function evaluations required during generation. In this work, we introduce a novel hierarchical generative framework that reduces the number of node pairs that must be evaluated and adopts discrete flow matching to significantly decrease the number of denoising iterations. We empirically demonstrate that our approach more effectively captures graph distributions while substantially reducing generation time.

Method: Uses softmax parameterization of the policy, constructs unbiased variance estimates using two independent draws from each arm’s distribution, and proposes a new algorithm to select the minimal variance arm with proven convergence guarantees.

Result: The algorithm converges under natural conditions, and numerical experiments demonstrate practical behavior and provide implementation guidance for risk-aware bandit problems.

Conclusion: The paper provides a theoretically sound and practical approach to risk-aware multi-armed bandit problems, enabling decision-making that balances reward maximization with variance minimization.

Abstract: Algorithms for the Multi-Armed Bandit (MAB) problem play a central role in sequential decision-making and have been extensively explored both theoretically and numerically. While most classical approaches aim to identify the arm with the highest expected reward, we focus on a risk-aware setting where the goal is to select the arm with the lowest variance, favoring stability over potentially high but uncertain returns. To model the decision process, we consider a softmax parameterization of the policy; we propose a new algorithm to select the minimal variance (or minimal risk) arm and prove its convergence under natural conditions. The algorithm constructs an unbiased estimate of the objective by using two independent draws from the current’s arm distribution. We provide numerical experiments that illustrate the practical behavior of these algorithms and offer guidance on implementation choices. The setting also covers general risk-aware problems where there is a trade-off between maximizing the average reward and minimizing its variance.

Method: Develops KD-ML framework where knowledge is represented as information granules organized as input-output knowledge landmarks. Uses augmented loss function with data optimization component and granular regularization component to enforce adherence to knowledge constraints.

Result: Experiments on physics-governed benchmarks show KD-ML consistently outperforms data-driven ML models. Hyperparameter analysis reveals balancing between data and knowledge contributions.

Conclusion: KD-ML provides effective framework for integrating granular knowledge with data, demonstrating improved performance over purely data-driven approaches in physics-informed settings.

Abstract: Informed Machine Learning has emerged as a viable generalization of Machine Learning (ML) by building a unified conceptual and algorithmic setting for constructing models on a unified basis of knowledge and data. Physics-informed ML involving physics equations is one of the developments within Informed Machine Learning. This study proposes a novel direction of Knowledge-Data ML, referred to as KD-ML, where numeric data are integrated with knowledge tidbits expressed in the form of granular knowledge landmarks. We advocate that data and knowledge are complementary in several fundamental ways: data are precise (numeric) and local, usually confined to some region of the input space, while knowledge is global and formulated at a higher level of abstraction. The knowledge can be represented as information granules and organized as a collection of input-output information granules called knowledge landmarks. In virtue of this evident complementarity, we develop a comprehensive design process of the KD-ML model and formulate an original augmented loss function L, which additively embraces the component responsible for optimizing the model based on available numeric data, while the second component, playing the role of a granular regularizer, so that it adheres to the granular constraints (knowledge landmarks). We show the role of the hyperparameter positioned in the loss function, which balances the contribution and guiding role of data and knowledge, and point to some essential tendencies associated with the quality of data (noise level) and the level of granularity of the knowledge landmarks. Experiments on two physics-governed benchmarks demonstrate that the proposed KD model consistently outperforms data-driven ML models.

Method: HALIDE (Hierarchical Apprenticeship Learning from Imperfect Demonstrations with Evolving Rewards) ranks sub-optimal student demonstrations within a hierarchical learning framework. It models student behavior at multiple abstraction levels to infer higher-level intent from suboptimal actions while capturing temporal evolution of student reward functions. It integrates demonstration quality into hierarchical reward inference to distinguish transient errors from suboptimal strategies.

Result: HALIDE more accurately predicts student pedagogical decisions than approaches relying on optimal trajectories, fixed rewards, or unranked imperfect demonstrations.

Conclusion: Imperfect student demonstrations contain valuable structured information that can be leveraged through hierarchical modeling and ranking to improve pedagogical policy learning in e-learning environments.

Abstract: While apprenticeship learning has shown promise for inducing effective pedagogical policies directly from student interactions in e-learning environments, most existing approaches rely on optimal or near-optimal expert demonstrations under a fixed reward. Real-world student interactions, however, are often inherently imperfect and evolving: students explore, make errors, revise strategies, and refine their goals as understanding develops. In this work, we argue that imperfect student demonstrations are not noise to be discarded, but structured signals-provided their relative quality is ranked. We introduce HALIDE, Hierarchical Apprenticeship Learning from Imperfect Demonstrations with Evolving Rewards, which not only leverages sub-optimal student demonstrations, but ranks them within a hierarchical learning framework. HALIDE models student behavior at multiple levels of abstraction, enabling inference of higher-level intent and strategy from suboptimal actions while explicitly capturing the temporal evolution of student reward functions. By integrating demonstration quality into hierarchical reward inference,HALIDE distinguishes transient errors from suboptimal strategies and meaningful progress toward higher-level learning goals. Our results show that HALIDE more accurately predicts student pedagogical decisions than approaches that rely on optimal trajectories, fixed rewards, or unranked imperfect demonstrations.

Method: Uses an LLM-guided program evolution framework to discover effective shuffling rules for without-replacement SGD. Identifies two key structural components: block reshuffling (reduces prefix-gradient variance) and paired reversal (symmetrizes epoch map, reduces order sensitivity).

Result: The discovered algorithm shows provable improvements over random reshuffling: block reshuffling reduces prefix-gradient variance constants, and paired reversal reduces order sensitivity from quadratic to cubic in step size. Numerical experiments demonstrate consistent gains across convex and nonconvex benchmarks.

Conclusion: LLM-guided program evolution successfully discovers novel shuffling strategies with theoretical guarantees and empirical improvements over standard SGD shuffling schemes, opening new directions for algorithm design.

Abstract: Shuffling strategies for stochastic gradient descent (SGD), including incremental gradient, shuffle-once, and random reshuffling, are supported by rigorous convergence analyses for arbitrary within-epoch permutations. In particular, random reshuffling is known to improve optimization constants relative to cyclic and shuffle-once schemes. However, existing theory offers limited guidance on how to design new data-ordering schemes that further improve optimization constants or stability beyond random reshuffling. In this paper, we design a pipeline using a large language model (LLM)-guided program evolution framework to discover an effective shuffling rule for without-replacement SGD. Abstracting from this instance, we identify two fundamental structural components: block reshuffling and paired reversal. We analyze these components separately and show that block reshuffling strictly reduces prefix-gradient variance constants within the unified shuffling framework, yielding provable improvements over random reshuffling under mild conditions. Separately, we show that paired reversal symmetrizes the epoch map and cancels the leading order-dependent second-order term, reducing order sensitivity from quadratic to cubic in the step size. Numerical experiments with the discovered algorithm validate the theory and demonstrate consistent gains over standard shuffling schemes across convex and nonconvex benchmarks.

Method: Constrained reinforcement learning framework that autonomously selects between implicit BDF integrator (CVODE) and quasi-steady-state solver during chemistry integration, treating solver selection as a Markov decision process

Result: Achieved mean speedup of ~3x (range 1.11x-10.58x) for 0D homogeneous reactor with 106-species n-dodecane mechanism, with only ~1% inference overhead; transferred to 1D counterflow flames without retraining, delivering ~2.2x speedup relative to CVODE

Conclusion: Demonstrates potential of RL framework to learn problem-specific integration strategies while respecting accuracy constraints, opening pathway toward adaptive, self-optimizing workflows for multiphysics systems with spatially heterogeneous stiffness

Abstract: The computational cost of stiff chemical kinetics remains a dominant bottleneck in reacting-flow simulation, yet hybrid integration strategies are typically driven by hand-tuned heuristics or supervised predictors that make myopic decisions from instantaneous local state. We introduce a constrained reinforcement learning (RL) framework that autonomously selects between an implicit BDF integrator (CVODE) and a quasi-steady-state (QSS) solver during chemistry integration. Solver selection is cast as a Markov decision process. The agent learns trajectory-aware policies that account for how present solver choices influence downstream error accumulation, while minimizing computational cost under a user-prescribed accuracy tolerance enforced through a Lagrangian reward with online multiplier adaptation. Across sampled 0D homogeneous reactor conditions, the RL-adaptive policy achieves a mean speedup of approximately $3\times$, with speedups ranging from $1.11\times$ to $10.58\times$, while maintaining accurate ignition delays and species profiles for a 106-species \textit{n}-dodecane mechanism and adding approximately $1%$ inference overhead. Without retraining, the 0D-trained policy transfers to 1D counterflow diffusion flames over strain rates $10$–$2000~\mathrm{s}^{-1}$, delivering consistent $\approx 2.2\times$ speedup relative to CVODE while preserving near-reference temperature accuracy and selecting CVODE at only $12$–$15%$ of space-time points. Overall, the results demonstrate the potential of the proposed reinforcement learning framework to learn problem-specific integration strategies while respecting accuracy constraints, thereby opening a pathway toward adaptive, self-optimizing workflows for multiphysics systems with spatially heterogeneous stiffness.

Method: Proposes stress profiling to quantify dataset difficulty along four interpretable stress dimensions, and SYNTHONY framework that matches stress profiles against a calibrated capability registry of synthesizer families to select optimal synthesizers based on user intent.

Result: Stress-based meta-features are highly predictive of synthesizer performance; kNN selector using these features achieves strong Top-1 selection accuracy, substantially outperforming zero-shot LLM selectors and random baselines across 7 datasets, 10 synthesizers, and 3 intents.

Conclusion: Stress profiling enables effective synthesizer selection, with the hand-crafted capability registry identified as the primary bottleneck, motivating learned capability representations as future work.

Abstract: Deep generative models for tabular data (GANs, diffusion models, and LLM-based generators) exhibit highly non-uniform behavior across datasets; the best-performing synthesizer family depends strongly on distributional stressors such as long-tailed marginals, high-cardinality categorical, Zipfian imbalance, and small-sample regimes. This brittleness makes practical deployment challenging, especially when users must balance competing objectives of fidelity, privacy, and utility. We study {intent-conditioned tabular synthesis selection}: given a dataset and a user intent expressed as a preference over evaluation metrics, the goal is to select a synthesizer that minimizes regret relative to an intent-specific oracle. We propose {stress profiling}, a synthesis-specific meta-feature representation that quantifies dataset difficulty along four interpretable stress dimensions, and integrate it into {SYNTHONY}, a selection framework that matches stress profiles against a calibrated capability registry of synthesizer families. Across a benchmark of 7 datasets, 10 synthesizers, and 3 intents, we demonstrate that stress-based meta-features are highly predictive of synthesizer performance: a $k$NN selector using these features achieves strong Top-1 selection accuracy, substantially outperforming zero-shot LLM selectors and random baselines. We analyze the gap between meta-feature-based and capability-based selection, identifying the hand-crafted capability registry as the primary bottleneck and motivating learned capability representations as a direction for future work.

Method: SAGE learns a proxy posterior over velocity models conditioned on both well logs and migrated seismic images during training. At inference, it generates full-resolution velocity fields conditioned solely on migrated images, with well information implicitly encoded in the learned distribution.

Result: SAGE demonstrates ability to capture complex subsurface variability under limited observational constraints on both synthetic and field datasets. The generated velocity models are geologically plausible and statistically accurate.

Conclusion: SAGE provides a scalable and data-efficient pathway for learning geological proxy posterior distributions for seismic imaging and inversion, enabling generation of realistic velocity models from incomplete observations.

Abstract: Recent advances in generative networks have enabled new approaches to subsurface velocity model synthesis, offering a compelling alternative to traditional methods such as Full Waveform Inversion. However, these approaches predominantly rely on the availability of large-scale datasets of high-quality, geologically realistic subsurface velocity models, which are often difficult to obtain in practice. We introduce SAGE, a novel framework for statistically consistent proxy velocity generation from incomplete observations, specifically sparse well logs and migrated seismic images. During training, SAGE learns a proxy posterior over velocity models conditioned on both modalities (wells and seismic); at inference, it produces full-resolution velocity fields conditioned solely on migrated images, with well information implicitly encoded in the learned distribution. This enables the generation of geologically plausible and statistically accurate velocity realizations. We validate SAGE on both synthetic and field datasets, demonstrating its ability to capture complex subsurface variability under limited observational constraints. Furthermore, samples drawn from the learned proxy distribution can be leveraged to train downstream networks, supporting inversion workflows. Overall, SAGE provides a scalable and data-efficient pathway toward learning geological proxy posterior for seismic imaging and inversion. Repo link: https://github.com/slimgroup/SAGE.

Method: Proposes CASA (Classification Augmented with Safety Attention) - a conditional decoding strategy that uses internal MLLM representations to predict binary safety tokens before response generation, with a novel safety attention module to enhance malicious query detection.

Result: CASA lowers average attack success rate by more than 97% across modalities and attack types on benchmarks including MM-SafetyBench, JailbreakV-28k, and adversarial audio tests, while maintaining strong utility on benign inputs.

Conclusion: CASA provides a simple and generalizable framework to improve multimodal LLM safety without external classifiers, auxiliary heads, or modality-specific safety fine-tuning.

Abstract: Multimodal large-language models (MLLMs) often experience degraded safety alignment when harmful queries exploit cross-modal interactions. Models aligned on text alone show a higher rate of successful attacks when extended to two or more modalities. In this work, we propose a simple conditional decoding strategy, CASA (Classification Augmented with Safety Attention) that utilizes internal representations of MLLMs to predict a binary safety token before response generation. We introduce a novel safety attention module designed to enhance the model’s ability to detect malicious queries. Our design ensures robust safety alignment without relying on any external classifier or auxiliary head, and without the need for modality-specific safety fine-tuning. On diverse benchmarks such as MM-SafetyBench, JailbreakV-28k, and adversarial audio tests, CASA lowers the average attack success rate by more than 97% across modalities and across attack types. Our empirical evaluations also show that CASA maintains strong utility in benign inputs, a result validated through both automated and human evaluations (via 13 trained annotators). Together, these results highlight CASA as a simple and generalizable framework to improve multimodal LLM safety.

Method: Four approaches: 1) ACDC automates circuit discovery in transformers; 2) Latent Adversarial Training removes behaviors by optimizing perturbations in residual stream; 3) Best-of-N jailbreaking uses random input augmentations; 4) Agentic misalignment tests evaluate models’ autonomous harmful choices.

Result: ACDC recovered 5 component types from GPT-2 Small in hours; LAT solved sleeper agent problem with 700x fewer GPU hours; Best-of-N achieved 89% attack success on GPT-4o; Agentic tests showed 96% blackmail rate for Claude Opus 4, with misbehavior rising from 6.5% to 55.1% when scenarios were presented as real.

Conclusion: The thesis makes four AI safety problems tractable and measurable but doesn’t fully resolve them. The work provides practical tools for understanding, removing, testing, and predicting dangerous behaviors in autonomous AI agents.

Abstract: Autonomous AI agents are being deployed with filesystem access, email control, and multi-step planning. This thesis contributes to four open problems in AI safety: understanding dangerous internal computations, removing dangerous behaviors once embedded, testing for vulnerabilities before deployment, and predicting when models will act against deployers. ACDC automates circuit discovery in transformers, recovering all five component types from prior manual work on GPT-2 Small by selecting 68 edges from 32,000 candidates in hours rather than months. Latent Adversarial Training (LAT) removes dangerous behaviors by optimizing perturbations in the residual stream to elicit failure modes, then training under those perturbations. LAT solved the sleeper agent problem where standard safety training failed, matching existing defenses with 700x fewer GPU hours. Best-of-N jailbreaking achieves 89% attack success on GPT-4o and 78% on Claude 3.5 Sonnet through random input augmentations. Attack success follows power law scaling across text, vision, and audio, enabling quantitative forecasting of adversarial robustness. Agentic misalignment tests whether frontier models autonomously choose harmful actions given ordinary goals. Across 16 models, agents engaged in blackmail (96% for Claude Opus 4), espionage, and actions causing death. Misbehavior rates rose from 6.5% to 55.1% when models stated scenarios were real rather than evaluations. The thesis does not fully resolve any of these problems but makes each tractable and measurable.

Method: Engineered 28 structured features directly from raw JSON fields (jobs, education, exits), combined with deterministic rule layer and XGBoost boosted stumps. Also extracted 9 features from prose fields using Claude Haiku LLM at different dataset coverages.

Result: Model achieves Val F0.5 = 0.3030, Precision = 0.3333, Recall = 0.2222 (+17.7pp improvement over zero-shot LLM baseline). LLM features capture 26.4% of model importance but add zero CV signal (delta = -0.05pp). Performance ceiling reflects dataset information content, not modeling limitations.

Conclusion: The signal in founder career data is limited; LLM-extracted features from prose fields don’t add value because they’re lossy re-encodings of the same structured data. The work serves as a benchmark diagnostic pointing to what richer datasets would need to include.

Abstract: Predicting startup success from founder career data is hard. The signal is weak, the labels are rare (9%), and most founders who succeed look almost identical to those who fail. We engineer 28 structured features directly from raw JSON fields – jobs, education, exits – and combine them with a deterministic rule layer and XGBoost boosted stumps. Our model achieves Val F0.5 = 0.3030, Precision = 0.3333, Recall = 0.2222 – a +17.7pp improvement over the zero-shot LLM baseline. We then run a controlled experiment: extract 9 features from the prose field using Claude Haiku, at 67% and 100% dataset coverage. LLM features capture 26.4% of model importance but add zero CV signal (delta = -0.05pp). The reason is structural: anonymised_prose is generated from the same JSON fields we parse directly – it is a lossy re-encoding, not a richer source. The ceiling (CV ~= 0.25, Val ~= 0.30) reflects the information content of this dataset, not a modeling limitation. In characterizing where the signal runs out and why, this work functions as a benchmark diagnostic – one that points directly to what a richer dataset would need to include.

Method: Two orthogonal strategies: (1) multi-token pooling using hierarchical pruning and clustering operators (Top-k, SAGPool, DiffPool, MinCutPool, VNPool) to increase graph-to-LLM bandwidth, and (2) enhancing graph encoder semantic quality via global attention mechanisms. Uses LoRA for stabilization during soft prompt tuning.

Result: Multi-token pooling with LoRA stabilization allows compressed representations to rival full-graph baselines (~73% Hit@1 on WebQSP). VNPool functions structurally as a single-layer Perceiver IO encoder. Analysis reveals GraphQA benchmark suffers from representational saturation where answers correlate with isolated node features.

Conclusion: Multi-token pooling with appropriate stabilization can mitigate information bottlenecks in GraphQA, though clustering operators remain challenging. The FandE Score adaptation reveals benchmark limitations with representational saturation issues.

Abstract: The integration of Graph Neural Networks (GNNs) with Large Language Models (LLMs) has emerged as a promising paradigm for Graph Question Answering (GraphQA). However, effective methods for encoding complex structural information into the LLM’s latent space remain an open challenge. Current state-of-the-art architectures, such as G-Retriever, typically rely on standard GNNs and aggressive mean pooling to compress entire graph substructures into a single token, creating a severe information bottleneck. This work mitigates this bottleneck by investigating two orthogonal strategies: (1) increasing the bandwidth of the graph-to-LLM interface via multi-token pooling, and (2) enhancing the semantic quality of the graph encoder via global attention mechanisms. We evaluate a suite of hierarchical pruning and clustering-based pooling operators including Top-k, SAGPool, DiffPool, MinCutPool, and Virtual Node Pooling (VNPool) to project graph data into multiple learnable tokens. Empirically, we demonstrate that while pooling introduces significant instability during soft prompt tuning, the application of Low-Rank Adaptation (LoRA) effectively stabilizes specific hierarchical projections (notably VNPool and pruning methods), though dense clustering operators remain challenging. This stabilization allows compressed representations to rival full-graph baselines (achieving ~73% Hit@1 on WebQSP). Conceptually, we demonstrate that a Graph Transformer with VNPool implementation functions structurally as a single-layer Perceiver IO encoder. Finally, we adapt the FandE (Features and Edges) Score to the generative GraphQA domain. Our analysis reveals that the GraphQA benchmark suffers from representational saturation, where target answers are often highly correlated with isolated node features. The implementation is available at https://github.com/Agrover112/G-Retriever/tree/all_good/

Method: Proposes a deep learning-based surrogate optimization framework with problem-informed sampling strategy that aligns training data with optimization trajectories. A neural network proxy is trained to approximate mapping between bottomhole pressure trajectories and cumulative production using data from coupled flow-geomechanics model. The proxy is embedded within constrained optimization workflow.

Result: Surrogate achieves agreement with full-physics solutions within 2-5% accuracy across multiple initializations while reducing computational cost by up to three orders of magnitude (1000x speedup). Discrepancies mainly occur near training distribution boundaries and due to local optimization effects.

Conclusion: Combining surrogate modeling with problem-informed sampling enables scalable and reliable optimization for high-dimensional, simulator-based problems, with broader applicability to PDE-constrained systems beyond reservoir engineering.

Abstract: Production optimization in stress-sensitive unconventional reservoirs is governed by a nonlinear trade-off between pressure-driven flow and stress-induced degradation of fracture conductivity and matrix permeability. While higher drawdown improves short-term production, it accelerates permeability loss and reduces long-term recovery. Identifying optimal, time-varying control strategies requires repeated evaluations of fully coupled flow-geomechanics simulators, making conventional optimization computationally expensive. We propose a deep learning-based surrogate optimization framework for high-dimensional well control. Unlike prior approaches that rely on predefined control parameterizations or generic sampling, our method treats well control as a continuous, high-dimensional problem and introduces a problem-informed sampling strategy that aligns training data with trajectories encountered during optimization. A neural network proxy is trained to approximate the mapping between bottomhole pressure trajectories and cumulative production using data from a coupled flow-geomechanics model. The proxy is embedded within a constrained optimization workflow, enabling rapid evaluation of control strategies. Across multiple initializations, the surrogate achieves agreement with full-physics solutions within 2-5 percent, while reducing computational cost by up to three orders of magnitude. Discrepancies are mainly associated with trajectories near the boundary of the training distribution and local optimization effects. This framework shows that combining surrogate modeling with problem-informed sampling enables scalable and reliable optimization for high-dimensional, simulator-based problems, with broader applicability to PDE-constrained systems.

Method: Proposes GUIDE framework with structured actions (intervention type, magnitude, timing) including bolus insulin and carbohydrate intake. Uses patient-specific glucose predictor trained on CGM data, supports both offline and online RL algorithms in unified environment.

Result: CQL-BC algorithm achieved highest average time-in-range (85.49%) with low hypoglycemia exposure. Behavioral similarity analysis shows learned policy preserves patient action patterns (mean cosine similarity 0.87 ± 0.09).

Conclusion: Conservative offline RL with structured behavioral action space can provide clinically meaningful and behaviorally plausible decision support for personalized diabetes management.

Abstract: Type 1 Diabetes (T1D) management requires continuous adjustment of insulin and lifestyle behaviors to maintain blood glucose within a safe target range. Although automated insulin delivery (AID) systems have improved glycemic outcomes, many patients still fail to achieve recommended clinical targets, warranting new approaches to improve glucose control in patients with T1D. While reinforcement learning (RL) has been utilized as a promising approach, current RL-based methods focus primarily on insulin-only treatment and do not provide behavioral recommendations for glucose control. To address this gap, we propose GUIDE, an RL-based decision-support framework designed to complement AID technologies by providing behavioral recommendations to prevent abnormal glucose events. GUIDE generates structured actions defined by intervention type, magnitude, and timing, including bolus insulin administration and carbohydrate intake events. GUIDE integrates a patient-specific glucose level predictor trained on real-world continuous glucose monitoring data and supports both offline and online RL algorithms within a unified environment. We evaluate both off-policy and on-policy methods across 25 individuals with T1D using standardized glycemic metrics. Among the evaluated approaches, the CQL-BC algorithm demonstrates the highest average time-in-range, reaching 85.49% while maintaining low hypoglycemia exposures. Behavioral similarity analysis further indicates that the learned CQL-BC policy preserves key structural characteristics of patient action patterns, achieving a mean cosine similarity of 0.87 $\pm$ 0.09 across subjects. These findings suggest that conservative offline RL with a structured behavioral action space can provide clinically meaningful and behaviorally plausible decision support for personalized diabetes management.

Method: Apply TracIn gradient-similarity scoring to GameFormer on nuPlan benchmark to construct curriculum that weights training scenarios by estimated contribution to validation loss reduction, comparing against metadata-based interaction-difficulty curriculum and uniform baseline.

Result: TracIn-weighted curriculum achieves mean planning ADE of 1.704±0.029m, significantly outperforming metadata-based curriculum (1.822±0.014m) with lower variance than uniform baseline (1.772±0.134m). Gradient-based scores are nearly orthogonal to metadata (Spearman ρ=-0.014).

Conclusion: Gradient-based data valuation captures training dynamics invisible to hand-crafted features and provides practical tool for improving sample efficiency in game-theoretic planning, with full-data curriculum weighting outperforming hard data selection.

Abstract: We demonstrate that gradient-based data valuation produces curriculum orderings that significantly outperform metadata-based heuristics for training game-theoretic motion planners. Specifically, we apply TracIn gradient-similarity scoring to GameFormer on the nuPlan benchmark and construct a curriculum that weights training scenarios by their estimated contribution to validation loss reduction. Across three random seeds, the TracIn-weighted curriculum achieves a mean planning ADE of $1.704\pm0.029$,m, significantly outperforming the metadata-based interaction-difficulty curriculum ($1.822\pm0.014$,m; paired $t$-test $p=0.021$, Cohen’s $d_z=3.88$) while exhibiting lower variance than the uniform baseline ($1.772\pm0.134$,m). Our analysis reveals that TracIn scores and scenario metadata are nearly orthogonal (Spearman $ρ=-0.014$), indicating that gradient-based valuation captures training dynamics invisible to hand-crafted features. We further show that gradient-based curriculum weighting succeeds where hard data selection fails: TracIn-curated 20% subsets degrade performance by $2\times$, whereas full-data curriculum weighting with the same scores yields the best results. These findings establish gradient-based data valuation as a practical tool for improving sample efficiency in game-theoretic planning.

Method: Separates trained networks from density estimators built from their representations/outputs. Introduces two estimators: Jacobian-based estimators and autoregressive self-estimators. Applies this framework to analyze iGPT, PixelCNN++, Glow, score-based diffusion models, DINOv2, and I-JEPA. Uses Spearman rank correlation to quantify ordering consistency with external complexity metrics.

Result: Found striking regularity: lower-complexity samples consistently receive higher estimated density across all models tested, while higher-complexity samples receive lower density. This ordering appears within test sets and across OOD pairs (CIFAR-10 vs SVHN). Models trained only on lowest-density (most complex) samples or even a single such sample still rank simpler images as higher density.

Conclusion: Deep networks consistently favor simple data across architectures, objectives, and density estimators. The OOD anomaly is part of a broader pattern where neural network representations systematically prefer simpler patterns over more complex ones.

Abstract: Estimated density is often interpreted as indicating how typical a sample is under a model. Yet deep models trained on one dataset can assign \emph{higher} density to simpler out-of-distribution (OOD) data than to in-distribution test data. We refer to this behavior as the OOD anomaly. Prior work typically studies this phenomenon within a single architecture, detector, or benchmark, implicitly assuming certain canonical densities. We instead separate the trained network from the density estimator built from its representations or outputs. We introduce two estimators: Jacobian-based estimators and autoregressive self-estimators, making density analysis applicable to a wide range of models. Applying this perspective to a range of models, including iGPT, PixelCNN++, Glow, score-based diffusion models, DINOv2, and I-JEPA, we find the same striking regularity that goes beyond the OOD anomaly: \textbf{lower-complexity samples receive higher estimated density, while higher-complexity samples receive lower estimated density}. This ordering appears within a test set and across OOD pairs such as CIFAR-10 and SVHN, and remains highly consistent across independently trained models. To quantify these orderings, we introduce Spearman rank correlation and find striking agreement both across models and with external complexity metrics. Even when trained only on the lowest-density (most complex) samples or \textbf{even a single such sample} the resulting models still rank simpler images as higher density. These observations lead us beyond the original OOD anomaly to a more general conclusion: deep networks consistently favor simple data. Our goal is not to close this question, but to define and visualize it more clearly. We broaden its empirical scope and show that it appears across architectures, objectives, and density estimators.

Method: CTP introduces a cross-graph tuning-free prompting framework that supports both homogeneous and heterogeneous graphs, can be directly deployed to unseen graphs without further parameter tuning, enabling plug-and-play GNN inference.

Result: Extensive experiments on few-shot prediction tasks show CTP achieves an average accuracy gain of 30.8% and maximum gain of 54% compared to state-of-the-art methods.

Conclusion: CTP offers an effective solution for cross-graph adaptation without parameter tuning and provides a new perspective on graph prompt learning.

Abstract: GNN prompting aims to adapt models across tasks and graphs without requiring extensive retraining. However, most existing graph prompt methods still require task-specific parameter updates and face the issue of generalizing across graphs, limiting their performance and undermining the core promise of prompting. In this work, we introduce a Cross-graph Tuning-free Prompting Framework (CTP), which supports both homogeneous and heterogeneous graphs, can be directly deployed to unseen graphs without further parameter tuning, and thus enables a plug-and-play GNN inference engine. Extensive experiments on few-shot prediction tasks show that, compared to SOTAs, CTP achieves an average accuracy gain of 30.8% and a maximum gain of 54%, confirming its effectiveness and offering a new perspective on graph prompt learning.

Method: Apply single targeted gradient-ascent step to increase loss of candidate example, measure changes in internal representations (logits, hidden-layer activations, projections), then train lightweight logistic classifier on drift signals to separate members from non-members.

Result: G-Drift substantially outperforms confidence-based, perplexity-based, and reference-based attacks across multiple transformer-based LLMs and realistic MIA benchmark datasets.

Conclusion: Small gradient interventions provide practical tool for auditing training data membership and assessing privacy risks in LLMs, with memorized samples showing smaller, more structured feature drift.

Abstract: Large language models (LLMs) are trained on massive web-scale corpora, raising growing concerns about privacy and copyright. Membership inference attacks (MIAs) aim to determine whether a given example was used during training. Existing LLM MIAs largely rely on output probabilities or loss values and often perform only marginally better than random guessing when members and non-members are drawn from the same distribution. We introduce G-Drift MIA, a white-box membership inference method based on gradient-induced feature drift. Given a candidate (x,y), we apply a single targeted gradient-ascent step that increases its loss and measure the resulting changes in internal representations, including logits, hidden-layer activations, and projections onto fixed feature directions, before and after the update. These drift signals are used to train a lightweight logistic classifier that effectively separates members from non-members. Across multiple transformer-based LLMs and datasets derived from realistic MIA benchmarks, G-Drift substantially outperforms confidence-based, perplexity-based, and reference-based attacks. We further show that memorized training samples systematically exhibit smaller and more structured feature drift than non-members, providing a mechanistic link between gradient geometry, representation stability, and memorization. In general, our results demonstrate that small, controlled gradient interventions offer a practical tool for auditing the membership of training-data and assessing privacy risks in LLMs.

Method: Construct probability density functions weighted by Lagrangian Descriptor values to embed geometric information into a statistical framework suitable for information-theoretic comparison. Benchmark physically constrained architectures (SympNet, HénonNet, Generalized Hamiltonian Neural Networks) against data-driven Reservoir Computing on two canonical systems: Duffing oscillator and three-mode nonlinear Schrödinger equation.

Result: For the Duffing oscillator, all models recover homoclinic orbit geometry with modest data requirements, though accuracy near critical structures varies. For the three-mode nonlinear Schrödinger equation, clear differences emerge: symplectic architectures preserve energy but distort phase-space topology, while Reservoir Computing reproduces homoclinic structure with high fidelity despite lacking explicit physical constraints.

Conclusion: Lagrangian Descriptor-based diagnostics provide valuable assessment of not only predictive performance but also the global dynamical integrity of learned Hamiltonian models, revealing trade-offs between physical constraints and geometric fidelity.

Abstract: We propose Lagrangian Descriptors (LDs) as a diagnostic framework for evaluating neural network models of Hamiltonian systems beyond conventional trajectory-based metrics. Standard error measures quantify short-term predictive accuracy but provide little insight into global geometric structures such as orbits and separatrices. Existing evaluation tools in dissipative systems are inadequate for Hamiltonian dynamics due to fundamental differences in the systems. By constructing probability density functions weighted by LD values, we embed geometric information into a statistical framework suitable for information-theoretic comparison. We benchmark physically constrained architectures (SympNet, HénonNet, Generalized Hamiltonian Neural Networks) against data-driven Reservoir Computing across two canonical systems. For the Duffing oscillator, all models recover the homoclinic orbit geometry with modest data requirements, though their accuracy near critical structures varies. For the three-mode nonlinear Schrödinger equation, however, clear differences emerge: symplectic architectures preserve energy but distort phase-space topology, while Reservoir Computing, despite lacking explicit physical constraints, reproduces the homoclinic structure with high fidelity. These results demonstrate the value of LD-based diagnostics for assessing not only predictive performance but also the global dynamical integrity of learned Hamiltonian models.

Method: 1) PCA-informed alignment to steer embeddings toward principal components for interpretable axes without distorting local neighborhoods. 2) Concept-alignment regularization that aligns embedding dimensions with external knowledge (class labels or user-defined concepts). 3) Extraction of common knowledge across the Rashomon set by identifying trustworthy and persistent nearest-neighbor relationships to construct refined embeddings.

Result: The framework provides a flexible approach for building interpretable, robust, and goal-aligned visualizations by moving beyond single embeddings and leveraging the multiplicity of equally good solutions in the Rashomon set.

Conclusion: Embracing the multiplicity of dimension reduction solutions through the Rashomon set framework leads to more powerful and trustworthy representations that can be aligned with interpretability goals and external knowledge.

Abstract: Dimension reduction (DR) is inherently non-unique: multiple embeddings can preserve the structure of high-dimensional data equally well while differing in layout or geometry. In this paper, we formally define the Rashomon set for DR – the collection of `good’ embedding – and show how embracing this multiplicity leads to more powerful and trustworthy representations. Specifically, we pursue three goals. First, we introduce PCA-informed alignment to steer embeddings toward principal components, making axes interpretable without distorting local neighborhoods. Second, we design concept-alignment regularization that aligns an embedding dimension with external knowledge, such as class labels or user-defined concepts. Third, we propose a method to extract common knowledge across the Rashomon set by identifying trustworthy and persistent nearest-neighbor relationships, which we use to construct refined embeddings with improved local structure while preserving global relationships. By moving beyond a single embedding and leveraging the Rashomon set, we provide a flexible framework for building interpretable, robust, and goal-aligned visualizations.

Method: Analyze empirical data to find output lengths follow heavy-tailed distributions, fit with log-t distribution. Propose Tail Inflated Expectation (TIE) metric that adjusts expectation with tail probabilities to account for risk of long outputs, replacing output length in SJF scheduling.

Result: TIE reduces per-token latency by 2.31× for online inference and improves throughput by 1.42× for offline data generation compared to three strong baselines.

Conclusion: Modeling output length uncertainty with distributions rather than point estimates significantly improves LLM inference scheduling performance.

Abstract: To schedule LLM inference, the \textit{shortest job first} (SJF) principle is favorable by prioritizing requests with short output lengths to avoid head-of-line (HOL) blocking. Existing methods usually predict a single output length for each request to facilitate scheduling. We argue that such a \textit{point estimate} does not match the \textit{stochastic} decoding process of LLM inference, where output length is \textit{uncertain} by nature and determined by when the end-of-sequence (EOS) token is sampled. Hence, the output length of each request should be fitted with a distribution rather than a single value. With an in-depth analysis of empirical data and the stochastic decoding process, we observe that output length follows a heavy-tailed distribution and can be fitted with the log-t distribution. On this basis, we propose a simple metric called Tail Inflated Expectation (TIE) to replace the output length in SJF scheduling, which adjusts the expectation of a log-t distribution with its tail probabilities to account for the risk that a request generates long outputs. To evaluate our TIE scheduler, we compare it with three strong baselines, and the results show that TIE reduces the per-token latency by $2.31\times$ for online inference and improves throughput by $1.42\times$ for offline data generation.

Method: Develops initialization-dependent complexity bounds using a new peeling technique to handle initialization-dependent constraints, focusing on path-norm of distance from initialization rather than spectral norms.

Result: Achieves logarithmic dependency on network width, provides tight lower bounds up to constant factors, and demonstrates non-vacuous generalization bounds for overparameterized networks through empirical comparisons.

Conclusion: The paper presents the first fully initialization-dependent generalization analysis for shallow neural networks that effectively explains benign overfitting in overparameterized settings with practical non-vacuous bounds.

Abstract: Overparameterized neural networks often show a benign overfitting property in the sense of achieving excellent generalization behavior despite the number of parameters exceeding the number of training examples. A promising direction to explain benign overfitting is to relate generalization to the norm of distance from initialization, motivated by the empirical observations that this distance is often significantly smaller than the norm itself. However, the existing initialization-dependent complexity analyses cannot fully exploit the power of initialization since the associated bounds depend on the spectral norm of the initialization matrix, which can scale as a square-root function of the width and are therefore not effective for overparameterized models. In this paper, we develop the first \emph{fully} initialization-dependent complexity bounds for shallow neural networks with general Lipschitz activation functions, which enjoys a logarithmic dependency on the width. Our bounds depend on the path-norm of the distance from initialization, which are derived by introducing a new peeling technique to handle the challenge along with the initialization-dependent constraint. We also develop a lower bound tight up to a constant factor. Finally, we conduct empirical comparisons and show that our generalization analysis implies non-vacuous bounds for overparameterized networks.

Method: Uses discrete wavelet transform to separate trajectories into low-frequency trends and high-frequency details, learns transferable basis vectors via sparse dictionary learning, constructs structured latent manifold with topology-aware contrastive constraint, and employs surrogate-assisted search for initial population sampling.

Result: DB-GEN improves tracking accuracy across various dynamic benchmarks compared to existing algorithms, performs direct online inference without retraining/fine-tuning, executes in milliseconds (~0.2 seconds per environmental change), and was pre-trained on 120 million solutions.

Conclusion: The proposed DB-GEN framework effectively addresses key challenges in dynamic multi-objective optimization through decoupled representation learning and zero-shot generation, enabling efficient and accurate tracking of moving Pareto fronts.

Abstract: Dynamic multi-objective optimization requires continuous tracking of moving Pareto fronts. Existing methods struggle with irregular mutations and data sparsity, primarily facing three challenges: the non-linear coupling of dynamic modes, negative transfer from outdated historical data, and the cold-start problem during environmental switches. To address these issues, this paper proposes a decoupled basis-vector-driven generative framework (DB-GEN). First, to resolve non-linear coupling, the framework employs the discrete wavelet transform to separate evolutionary trajectories into low-frequency trends and high-frequency details. Second, to mitigate negative transfer, it learns transferable basis vectors via sparse dictionary learning rather than directly memorizing historical instances. Recomposing these bases under a topology-aware contrastive constraint constructs a structured latent manifold. Finally, to overcome the cold-start problem, a surrogate-assisted search paradigm samples initial populations from this manifold. Pre-trained on 120 million solutions, DB-GEN performs direct online inference without retraining or fine-tuning. This zero-shot generation process executes in milliseconds, requiring approximately 0.2 seconds per environmental change. Experimental results demonstrate that DB-GEN improves tracking accuracy across various dynamic benchmarks compared to existing algorithms.

Method: 1) Multi-head modality fusion module to adaptively integrate raw signals; 2) Joint contrastive and reinforcement learning framework to autonomously explore effective reasoning strategies; 3) Fine-grained residual enhancement module to progressively preserve local details throughout the network.

Result: Demonstrates state-of-the-art zero-shot performance across various downstream tasks on both the proposed MBE3.0 benchmark and public datasets.

Conclusion: MOON3.0 successfully addresses key challenges in fine-grained product attribute modeling through reasoning-aware MLLM design, achieving superior performance and releasing a large-scale multimodal e-commerce benchmark.

Abstract: With the rapid growth of e-commerce, exploring general representations rather than task-specific ones has attracted increasing attention. Although recent multimodal large language models (MLLMs) have driven significant progress in product understanding, they are typically employed as feature extractors that implicitly encode product information into global embeddings, thereby limiting their ability to capture fine-grained attributes. Therefore, we argue that leveraging the reasoning capabilities of MLLMs to explicitly model fine-grained product attributes holds significant potential. Nevertheless, achieving this goal remains non-trivial due to several key challenges: (i) long-context reasoning tends to dilute the model’s attention to salient information in the raw input; (ii) supervised fine-tuning (SFT) primarily encourages rigid imitation, limiting the exploration of effective reasoning strategies; and (iii) fine-grained details are progressively attenuated during forward propagation. To address these issues, we propose MOON3.0, the first reasoning-aware MLLM-based model for product representation learning. Our method (1) employs a multi-head modality fusion module to adaptively integrate raw signals; (2) incorporates a joint contrastive and reinforcement learning framework to autonomously explore more effective reasoning strategies; and (3) introduces a fine-grained residual enhancement module to progressively preserve local details throughout the network. Additionally, we release a large-scale multimodal e-commerce benchmark MBE3.0. Experimentally, our model demonstrates state-of-the-art zero-shot performance across various downstream tasks on both our benchmark and public datasets.

Method: Multi-format QAT trains models robust across multiple quantization formats, with Slice-and-Scale conversion for MXINT/MXFP formats and anchor format checkpoint storage.

Result: Multi-format QAT matches single-format QAT performance at each precision, works with unseen formats, and enables on-the-fly conversion with minimal accuracy loss.

Conclusion: Provides practical elastic precision scaling for diverse deployment targets by allowing runtime format selection without retraining.

Abstract: Quantization-aware training (QAT) is typically performed for a single target numeric format, while practical deployments often need to choose numerical precision at inference time based on hardware support or runtime constraints. We study multi-format QAT, where a single model is trained to be robust across multiple quantization formats. We find that multi-format QAT can match single-format QAT at each target precision, yielding one model that performs well overall across different formats, even formats that were not seen during training. To enable practical deployment, we propose the Slice-and-Scale conversion procedure for both MXINT and MXFP that converts a high-precision representation into lower-precision formats without re-training. Building on this, we introduce a pipeline that (i) trains a model with multi-format QAT, (ii) stores a single anchor format checkpoint (MXINT8/MXFP8), and (iii) allows on-the-fly conversion to lower MXINT or MXFP formats at runtime with negligible-or no-additional accuracy degradation. Together, these components provide a practical path to elastic precision scaling and allow selecting the runtime format at inference time across diverse deployment targets.

Method: Proposes an Optimism in the Face of Uncertainty Linear (OFUL) algorithm that: 1) collects data through exploration phase, 2) estimates shared low-rank representation via spectral initialization, 3) constructs confidence sets based on the shared model, and 4) conducts OFUL-based learning over these confidence sets.

Result: Theoretical analysis shows the algorithm achieves cumulative regret bound of Õ(√(drNT)), significantly better than independent task learning’s Õ(dT√N). Numerical simulations validate improved performance across different problem sizes.

Conclusion: Multi-task representation learning in linear bandits with shared low-rank structure enables substantial improvements in sample efficiency and regret reduction compared to independent task learning, with theoretical guarantees and empirical validation.

Abstract: Multi-task representation learning is an approach that learns shared latent representations across related tasks, facilitating knowledge transfer and improving sample efficiency. This paper introduces a novel approach to multi-task representation learning in linear bandits. We consider a setting with T concurrent linear bandit tasks, each with feature dimension d, that share a common latent representation of dimension r \ll min{d,T}$, capturing their underlying relatedness. We propose a new Optimism in the Face of Uncertainty Linear (OFUL) algorithm that leverages shared low-rank representations to enhance decision-making in a sample-efficient manner. Our algorithm first collects data through an exploration phase, estimates the shared model via spectral initialization, and then conducts OFUL based learning over a newly constructed confidence set. We provide theoretical guarantees for the confidence set and prove that the unknown reward vectors lie within the confidence set with high probability. We derive cumulative regret bounds and show that the proposed approach achieves \tilde{O}(\sqrt{drNT}), a significant improvement over solving the T tasks independently, resulting in a regret of \tilde{O}(dT\sqrt{N}). We performed numerical simulations to validate the performance of our algorithm for different problem sizes.

Method: Inspired by Drosophila memory updating pathways, SyCo uses Rac1 pathway to confine plasticity to tail-gradient subspace (less critical for source knowledge) and MAPK pathway with tiered controller to suppress noisy updates. Updates low-rank adapters with structured TTA objective based on problem understanding, process understanding, and source-domain guardrail.

Result: Across 18 NLP datasets and Multi-source Open-set Adaptation (MOA) setting, SyCo achieves 78.31% on unseen-task adaptation and 85.37% on unseen-data shifts, outperforming strong baselines.

Conclusion: SyCo provides effective parameter-efficient adaptation for LLMs in complex deployment scenarios with multiple sources and continually emerging tasks, balancing rapid specialization with source knowledge preservation.

Abstract: Large Language Models (LLMs) generalize across tasks via reusable representations and flexible reasoning, yet remain brittle in real deployment under evolving tasks and continual distribution shift. A common approach is Test-Time Adaptation (TTA), existing ones of which updates models with hand-designed unsupervised objectives over the full parameter space and mostly overlook preserving shared source knowledge and the reliability of adaptation signals. Drawing on molecular signaling cascades of memory updating in Drosophila, we propose Synapse Consolidation (SyCo), a parameter-efficient LLM adaptation method that updates low-rank adapters through Rac1 and MAPK pathways under the guidance of a structured TTA objective driven by problem understanding, process understanding, and source-domain guardrail. Rac1 confines plasticity to a tail-gradient subspace that is less critical for source knowledge, enabling rapid specialization while preserving source representations. MAPK uses a tiered controller to suppress noisy updates and consolidate useful adaptations under non-stationary streams. To model real deployments with multiple sources and continually emerging tasks, we introduce Multi-source Open-set Adaptation (MOA) setting, where a model is trained on multiple labeled source tasks and then adapts on open, non-stationary unlabeled test streams that mix seen and unseen tasks with partial overlap in label and intent space. Across 18 NLP datasets and the MOA setting, SyCo consistently outperforms strong baselines, achieving 78.31% on unseen-task adaptation and 85.37% on unseen-data shifts.

Method: HabitatAgent uses a multi-agent architecture with four specialized roles: Memory Agent (maintains multi-layer user memory with constraint extraction, memory fusion, and verification-gated updates), Retrieval Agent (performs hybrid vector-graph retrieval called GraphRAG), Generation Agent (produces evidence-referenced recommendations and explanations), and Validation Agent (applies multi-tier verification and targeted remediation).

Result: Evaluated on 100 real user consultation scenarios (300 multi-turn QA pairs) with end-to-end correctness protocol. A strong baseline (Dense+Rerank) achieved 75% accuracy, while HabitatAgent reached 95% accuracy.

Conclusion: HabitatAgent provides an auditable and reliable workflow for end-to-end housing consultation, significantly outperforming existing approaches by combining specialized agents for memory management, hybrid retrieval, evidence-based generation, and multi-tier validation.

Abstract: Housing selection is a high-stakes and largely irreversible decision problem. We study housing consultation as a decision-support interface for housing selection. Existing housing platforms and many LLM-based assistants often reduce this process to ranking or recommendation, resulting in opaque reasoning, brittle multi-constraint handling, and limited guarantees on factuality. We present HabitatAgent, the first LLM-powered multi-agent architecture for end-to-end housing consultation. HabitatAgent comprises four specialized agent roles: Memory, Retrieval, Generation, and Validation. The Memory Agent maintains multi-layer user memory through internal stages for constraint extraction, memory fusion, and verification-gated updates; the Retrieval Agent performs hybrid vector–graph retrieval (GraphRAG); the Generation Agent produces evidence-referenced recommendations and explanations; and the Validation Agent applies multi-tier verification and targeted remediation. Together, these agents provide an auditable and reliable workflow for end-to-end housing consultation. We evaluate HabitatAgent on 100 real user consultation scenarios (300 multi-turn question–answer pairs) under an end-to-end correctness protocol. A strong single-stage baseline (Dense+Rerank) achieves 75% accuracy, while HabitatAgent reaches 95%.

Method: Introduced Orientation features as a geometrically grounded, rotation-aware encoding of protein backbone. Compared it against common descriptions across three dynamical regimes: fast-folding proteins, large-scale domain motions, and protein-protein association. Developed ManiProt library for efficient computation and analysis of multiple protein representations.

Result: Different representations emphasize complementary aspects of conformational space across all three systems studied. No single representation provides a complete picture of the underlying dynamics. Representation choice fundamentally shapes conformational organization, similarity relationships, and apparent transitions.

Conclusion: A comparative, representation-aware framework is necessary for interpreting molecular dynamics simulations, as different representations reveal complementary aspects of protein conformational space and dynamics.

Abstract: Molecular dynamics simulations provide detailed trajectories at the atomic level, but extracting interpretable and robust insights from these high-dimensional data remains challenging. In practice, analyses typically rely on a single representation. Here, we show that representation choice is not neutral: it fundamentally shapes the conformational organization, similarity relationships, and apparent transitions inferred from identical simulation data. To complement existing representations, we introduce Orientation features, a geometrically grounded, rotation-aware encoding of protein backbone. We compare it against common descriptions across three dynamical regimes: fast-folding proteins, large-scale domain motions, and protein-protein association. Across these systems, we find that different representations emphasize complementary aspects of conformational space, and that no single representation provides a complete picture of the underlying dynamics. To facilitate systematic comparison, we developed ManiProt, a library for efficient computation and analysis of multiple protein representations. Our results motivate a comparative, representation-aware framework for the interpretation of molecular dynamics simulations.

Method: Sparse Identification Graph Neural Network (SIGN) decouples scalability from network size by defining symbolic discovery as edge-level information, enabling efficient equation discovery even in large systems with over 100,000 nodes.

Result: SIGN recovers governing equations with high precision across diverse benchmark systems (coupled chaotic oscillators, neural dynamics, epidemic spreading) and sustains accurate long-term predictions. Applied to 71,987 sea surface temperature positions, it identifies compact predictive models capturing conditions up to two years in advance.

Conclusion: SIGN opens a path toward interpretable and reliable prediction of real-world complex systems by enabling equation discovery at previously inaccessible scales.

Abstract: Predicting the behavior of ultra-large complex systems, from climate to biological and technological networks, is a central unsolved challenge. Existing approaches face a fundamental trade-off: equation discovery methods provide interpretability but fail to scale, while neural networks scale but operate as black boxes and often lose reliability over long times. Here, we introduce the Sparse Identification Graph Neural Network, a framework that overcome this divide by allowing to infer the governing equations of large networked systems from data. By defining symbolic discovery as edge-level information, SIGN decouples the scalability of sparse identification from network size, enabling efficient equation discovery even in large systems. SIGN allows to study networks with over 100,000 nodes while remaining robust to noise, sparse sampling, and missing data. Across diverse benchmark systems, including coupled chaotic oscillators, neural dynamics, and epidemic spreading, it recovers governing equations with high precision and sustains accurate long-term predictions. Applied to a data set of time series of temperature measurements in 71,987 sea surface positions, SIGN identifies a compact predictive network model and captures large-scale sea surface temperature conditions up to two years in advance. By enabling equation discovery at previously inaccessible scales, SIGN opens a path toward interpretable and reliable prediction of real-world complex systems.

Method: Introduces a unified f-divergence framework over on-policy samples and organizes the OPD landscape along three dimensions: feedback signal (logit-based, outcome-based, or self-play), teacher access (white-box, black-box, or teacher-free), and loss granularity (token-level, sequence-level, or hybrid). Systematically analyzes representative methods across these dimensions.

Result: Provides first comprehensive overview of OPD for LLMs, examining industrial deployments and identifying open problems including distillation scaling laws, uncertainty-aware feedback, and agent-level distillation.

Conclusion: On-Policy Distillation addresses fundamental limitations of traditional distillation by enabling interactive learning where students receive feedback on their own generations, offering a more robust framework for transferring reasoning capabilities from large to small models.

Abstract: Knowledge distillation has become a primary mechanism for transferring reasoning and domain expertise from frontier Large Language Models (LLMs) to smaller, deployable students. However, the dominant paradigm remains \textit{off-policy}: students train on static teacher-generated data and never encounter their own errors during learning. This train–test mismatch, an instance of \textit{exposure bias}, causes prediction errors to compound autoregressively at inference time. On-Policy Distillation (OPD) addresses this by letting the student generate its own trajectories and receive teacher feedback on these self-generated outputs, grounding distillation in the theory of interactive imitation learning. Despite rapid growth spanning divergence minimization, reward-guided learning, and self-play, the OPD literature remains fragmented with no unified treatment. This survey provides the first comprehensive overview of OPD for LLMs. We introduce a unified $f$-divergence framework over on-policy samples and organize the landscape along three orthogonal dimensions: \emph{feedback signal} (logit-based, outcome-based, or self-play), \emph{teacher access} (white-box, black-box, or teacher-free), and \emph{loss granularity} (token-level, sequence-level, or hybrid). We systematically analyze representative methods, examine industrial deployments, and identify open problems including distillation scaling laws, uncertainty-aware feedback, and agent-level distillation.

Method: Adapts the DualPrompt algorithm for next activity prediction, using prompt-based continual learning to maintain performance on previous tasks while adapting to new data distributions with concept drifts.

Result: Achieves state-of-the-art or competitive results on three synthetic and two real-world datasets with recurring concept drifts, outperforming five baselines using a novel task-specific forgetting metric.

Conclusion: CNAPwP effectively mitigates catastrophic forgetting in dynamic process environments, demonstrating practical applicability for real-world scenarios with concept drifts.

Abstract: Predictive process monitoring (PPM) focuses on predicting future process trajectories, including next activity predictions. This is crucial in dynamic environments where processes change or face uncertainty. However, current frameworks often assume a static environment, overlooking dynamic characteristics and concept drifts. This results in catastrophic forgetting, where training while focusing merely on new data distribution negatively impacts the performance on previously learned data distributions. Continual learning addresses, among others, the challenges related to mitigating catastrophic forgetting. This paper proposes a novel approach called Continual Next Activity Prediction with Prompts (CNAPwP), which adapts the DualPrompt algorithm for next activity prediction to improve accuracy and adaptability while mitigating catastrophic forgetting. We introduce new datasets with recurring concept drifts, alongside a task-specific forgetting metric that measures the prediction accuracy gap between initial occurrence and subsequent task occurrences. Extensive testing on three synthetic and two real-world datasets representing several setups of recurrent drifts shows that CNAPwP achieves SOTA or competitive results compared to five baselines, demonstrating its potential applicability in real-world scenarios. An open-source implementation of our method, together with the datasets and results, is available at: https://github.com/SvStraten/CNAPwP.

Method: V-NSDE combines Neural SDEs with Variational Autoencoders. An encoder processes initial observations and district embeddings into a Gaussian distribution for the initial latent state. A Neural SDE with neural network-based drift and diffusion functions models continuous-time latent dynamics, incorporating time, latent state, and district embeddings. A probabilistic decoder reconstructs observations from the latent trajectory using Gaussian likelihood.

Result: The model effectively learns complex temporal patterns, producing realistic outcomes with clear trends and random fluctuations across different districts. The ELBO training with KL-divergence regularization improves model performance.

Conclusion: V-NSDE successfully addresses challenges in modeling complex socioeconomic time-series data by combining the expressive power of Neural SDEs with the generative capabilities of VAEs, enabling effective capture of both trends and variations with district-specific characteristics.

Abstract: This study examines the challenges of modeling complex and noisy data related to socioeconomic factors over time, with a focus on data from various districts in Odisha, India. Traditional time-series models struggle to capture both trends and variations together in this type of data. To tackle this, a Variational Neural Stochastic Differential Equation (V-NSDE) model is designed that combines the expressive dynamics of Neural SDEs with the generative capabilities of Variational Autoencoders (VAEs). This model uses an encoder and a decoder. The encoder takes the initial observations and district embeddings and translates them into a Gaussian distribution, which determines the mean and log-variance of the first latent state. Then the obtained latent state initiates the Neural SDE, which utilize neural networks to determine the drift and diffusion functions that govern continuous-time latent dynamics. These governing functions depend on the time index, latent state, and district embedding, which help the model learn the unique characteristics specific to each district. After that, using a probabilistic decoder, the observations are reconstructed from the latent trajectory. The decoder outputs a mean and log-variance for each time step, which follows the Gaussian likelihood. The Evidence Lower Bound (ELBO) training loss improves by adding a KL-divergence regularization term to the negative log-likelihood (nll). The obtained results demonstrate the effective learning of V-NSDE in recognizing complex patterns over time, yielding realistic outcomes that include clear trends and random fluctuations across different areas.

Method: Proposes Full-Gradient Successor Feature Representations Q-Learning (FG-SFRQL) that minimizes the full Mean Squared Bellman Error by computing gradients with respect to parameters in both online and target networks, unlike standard semi-gradient approaches.

Result: Theoretical proof of almost-sure convergence for FG-SFRQL and empirical demonstration of superior sample efficiency and transfer performance compared to semi-gradient baselines in both discrete and continuous domains.

Conclusion: Full gradient optimization for successor features provides more stable learning with convergence guarantees, leading to better transfer learning performance in multi-task reinforcement learning settings.

Abstract: Successor Features (SF) combined with Generalized Policy Improvement (GPI) provide a robust framework for transfer learning in Reinforcement Learning (RL) by decoupling environment dynamics from reward functions. However, standard SF learning methods typically rely on semi-gradient Temporal Difference (TD) updates. When combined with non-linear function approximation, semi-gradient methods lack robust convergence guarantees and can lead to instability, particularly in the multi-task setting where accurate feature estimation is critical for effective GPI. Inspired by Full Gradient DQN, we propose Full-Gradient Successor Feature Representations Q-Learning (FG-SFRQL), an algorithm that optimizes the successor features by minimizing the full Mean Squared Bellman Error. Unlike standard approaches, our method computes gradients with respect to parameters in both the online and target networks. We provide a theoretical proof of almost-sure convergence for FG-SFRQL and demonstrate empirically that minimizing the full residual leads to superior sample efficiency and transfer performance compared to semi-gradient baselines in both discrete and continuous domains.

Method: Compares reduced-basis neural operators (trained with L² and H¹ objectives) and Fourier neural operators against polynomial surrogates (reduced-basis sparse-grid and tensor-train). Evaluates on linear parametric diffusion and nonlinear parametric hyperelasticity problems with varying input field smoothness (decay rates s). Uses Pareto frontier analysis of cost vs accuracy, decomposing cost into data generation, setup, and evaluation components.

Result: No single method is universally superior. Polynomial surrogates achieve substantially better data efficiency for smooth input fields (s ≥ 2), with sparse-grid surrogate convergence rates matching theoretical predictions. For rough inputs (s ≤ 1), Fourier neural operator shows fastest convergence rates. Derivative-informed training improves data efficiency over standard L² training, providing competitive alternative for rough inputs in low-data regime when Jacobian information is available.

Conclusion: Surrogate methodology should be matched to problem regularity, accuracy demands, and computational constraints. Different methods excel in different regimes: polynomial methods for smooth problems, neural operators for rough problems, with derivative-informed training offering advantages when Jacobian information is available.

Abstract: We consider the problem of constructing surrogate operators for parameter-to-solution maps arising from parametric partial differential equations, where repeated forward model evaluations are computationally expensive. We present a systematic empirical comparison of neural operator surrogates, including a reduced-basis neural operator trained with $L^2_μ$ and $H^1_μ$ objectives and the Fourier neural operator, against polynomial surrogate methods, specifically a reduced-basis sparse-grid surrogate and a reduced-basis tensor-train surrogate. All methods are evaluated on a linear parametric diffusion problem and a nonlinear parametric hyperelasticity problem, using input fields with algebraically decaying spectral coefficients at varying rates of decay $s$. To enable fair comparisons, we analyze ensembles of surrogate models generated by varying hyperparameters and compare the resulting Pareto frontiers of cost versus approximation accuracy, decomposing cost into contributions from data generation, setup, and evaluation. Our results show that no single method is universally superior. Polynomial surrogates achieve substantially better data efficiency for smooth input fields ($s \geq 2$), with convergence rates for the sparse-grid surrogate in agreement with theoretical predictions. For rough inputs ($s \leq 1$), the Fourier neural operator displays the fastest convergence rates. Derivative-informed training consistently improves data efficiency over standard $L^2_μ$ training, providing a competitive alternative for rough inputs in the low-data regime when Jacobian information is available at reasonable cost. These findings highlight the importance of matching the surrogate methodology to the regularity of the problem as well as accuracy demands and computational constraints of the application.

Method: HiLL jointly trains a hinter policy and reasoner policy. For hard questions, the hinter generates hints online conditioned on the reasoner’s incorrect rollouts. Introduces hint reliance metric measuring dependence of correct trajectories on hints. Derives transferability result showing lower hint reliance implies better transfer, and uses this to define transfer-weighted reward for training hinter.

Result: Experiments across multiple benchmarks show HiLL consistently outperforms GRPO and prior hint-based baselines, demonstrating value of adaptive and transfer-aware hint learning for RL.

Conclusion: HiLL addresses advantage collapse in GRPO through adaptive hint generation that evolves with reasoner’s errors and transfer-weighted rewards that ensure hints improve the original no-hint policy, leading to better performance than existing methods.

Abstract: Group Relative Policy Optimization (GRPO) is widely used for reinforcement learning with verifiable rewards, but it often suffers from advantage collapse: when all rollouts in a group receive the same reward, the group yields zero relative advantage and thus no learning signal. For example, if a question is too hard for the reasoner, all sampled rollouts can be incorrect and receive zero reward. Recent work addresses this issue by adding hints or auxiliary scaffolds to such hard questions so that the reasoner produces mixed outcomes and recovers a non-zero update. However, existing hints are usually fixed rather than adapted to the current reasoner, and a hint that creates learning signal under the hinted input does not necessarily improve the no-hint policy used at test time. To this end, we propose Hint Learning for Reinforcement Learning (HiLL), a framework that jointly trains a hinter policy and a reasoner policy during RL. For each hard question, the hinter generates hints online conditioned on the current reasoner’s incorrect rollout, allowing hint generation to adapt to the reasoner’s evolving errors. We further introduce hint reliance, which measures how strongly correct hinted trajectories depend on the hint. We derive a transferability result showing that lower hint reliance implies stronger transfer from hinted success to no-hint success, and we use this result to define a transfer-weighted reward for training the hinter. Therefore, HiLL favors hints that not only recover informative GRPO groups, but also produce signals that are more likely to improve the original no-hint policy. Experiments across multiple benchmarks show that HiLL consistently outperforms GRPO and prior hint-based baselines, demonstrating the value of adaptive and transfer-aware hint learning for RL. The code is available at https://github.com/Andree-9/HiLL.

Method: Controlled study using targeted fault injection at GPU matrix-multiply instruction level. Characterizes sensitivity of bit positions, kernel functions, and execution stages. Proposes lightweight detection method for harmful parameter updates.

Result: Locally originating faults can cause impactful corruption including NaN propagation, loss spikes, gradient norm spikes, attention logit spikes, and persistent parameter divergence. Detection method works and recomputing recent training steps effectively mitigates impact.

Conclusion: SDC poses serious risks to LLM training. The proposed detection and mitigation approach can effectively handle these hardware-induced faults, preventing training failures.

Abstract: As Large Language Models (LLMs) scale in size and complexity, the consequences of failures during training become increasingly severe. A major challenge arises from Silent Data Corruption (SDC): hardware-induced faults that bypass system-level detection mechanisms. SDC may behave like benign numerical noise, but can also cause harmful gradient corruption that leads to loss spikes, divergence, or stalled progress. This work provides a controlled study of how intermittent SDC affects LLM pretraining. Using targeted fault injection at the level of GPU matrix-multiply instructions, we characterize the sensitivity of different bit positions, kernel functions, and execution stages. Our analysis shows that locally originating faults can produce impactful corruption, including NaN propagation, short-lived spikes in loss, gradient norm, and attention logits, as well as persistent parameter divergence. Building on the observed corruption signatures, we propose a lightweight detection method that identifies potentially harmful parameter updates. Experiments on LLaMA models with 60M, 350M, and 1.3B parameters demonstrate that recomputing the most recent training step upon detection can effectively mitigate the impact of these events.

Method: SCT replaces dense weight matrices with permanent truncated SVD factors (W = U diag(s) V^T), where the full dense matrix is never materialized. Gradients flow through compact spectral factors via standard backpropagation, with U and V retracted to the Stiefel manifold via QR decomposition after each optimizer step.

Result: Achieves up to 199x memory reduction per MLP layer at rank 32, enabling full training of 70B-parameter architectures on a Steam Deck (7.2 GB peak vs. 1,245 GB for dense FP32). Rank 128 emerges as efficiency sweet spot with 11.7x MLP compression and lowest perplexity. GPU memory drops 46% at rank 32 while training throughput doubles.

Conclusion: SCT dramatically reduces memory requirements for LLM training, making large-scale training feasible on consumer hardware. The learning rate schedule, not MLP rank, is identified as the primary bottleneck for convergence.

Abstract: The memory wall remains the primary bottleneck for training large language models on consumer hardware. We introduce Spectral Compact Training (SCT), a method that replaces dense weight matrices with permanent truncated SVD factors W = U diag(s) V^T, where the full dense matrix is never materialized during training or inference. Gradients flow through the compact spectral factors via standard backpropagation, and U, V are retracted to the Stiefel manifold via QR decomposition after each optimizer step. SCT achieves up to 199x memory reduction per MLP layer at rank 32, enabling full training steps of 70B-parameter architectures on a Steam Deck handheld (7.2 GB peak memory vs. 1,245 GB for dense FP32 training with Adam). Rank-sweep experiments on SmolLM2-1.7B (ranks 32-256, 2000 steps, NVIDIA A100) show that all tested ranks converge to the same loss floor (~4.2-4.5), identifying the learning rate schedule – not MLP rank – as the primary bottleneck. Rank 128 emerges as the efficiency sweet spot at 11.7x MLP compression with the lowest perplexity. GPU memory drops 46% at rank 32 while training throughput doubles.

Method: Extends COMPASS transformer model by integrating biomarker and treatment information through specialized loss components (treatment gating and pathway consistency loss) rather than feeding biomarkers as direct input. Uses Leave-one-cohort-out, Leave-one-cancer-type-out, and Leave-one-treatment-out evaluation strategies.

Result: Treatment gating and pathway consistency loss components improved generalizability across different evaluation strategies. Integrating biomarker and treatment information through carefully designed components enhances model generalization for immunotherapy response prediction.

Conclusion: Building components that exploit biomarker and treatment information improves generalizability in immunotherapy response prediction. Future research should focus on carefully curated components leveraging complementary clinical information and domain knowledge.

Abstract: Datasets used in immunotherapy response prediction are typically small in size, as well as diverse in cancer type, drug administered, and sequencer used. Models often drop in performance when tested on patient cohorts that are not included in the training process. Recent work has shown that transformer-based models along with self-supervised learning show better generalisation performance than threshold-based biomarkers, but is still suboptimal. We present BioCOMPASS, an extension of a transformer-based model called COMPASS, that integrates biomarkers and treatment information to further improve its generalisability. Instead of feeding biomarker data as input, we built loss components to align them with the model’s intermediate representations. We found that components such as treatment gating and pathway consistency loss improved generalisability when evaluated with Leave-one-cohort-out, Leave-one-cancer-type-out and Leave-one-treatment-out strategies. Results show that building components that exploit biomarker and treatment information can help in generalisability of immunotherapy response prediction. Careful curation of additional components that leverage complementary clinical information and domain knowledge represents a promising direction for future research.

Method: Three components: 1) Naturalistic data collection with multi-position wearable sensing and free-form time-aligned narrative descriptions, 2) Retrieval-based evaluation framework measuring semantic alignment between sensor data and language, 3) Language-conditioned learning architecture for sensor-to-text inference over variable-length streams and heterogeneous placements.

Result: Models with fixed-label objectives degrade sharply under real-world variability, while open-vocabulary sensor-language alignment yields robust representations. Achieves 65.3% Macro-F1 in cross-participant evaluation vs 31-34% for closed-set baselines.

Conclusion: Open-ended narrative modeling provides practical foundation for real-world wearable HAR, where closed-set recognition becomes simple downstream task after sensor-language alignment is learned.

Abstract: Wearable HAR has improved steadily, but most progress still relies on closed-set classification, which limits real-world use. In practice, human activity is open-ended, unscripted, personalized, and often compositional, unfolding as narratives rather than instances of fixed classes. We argue that addressing this gap does not require simply scaling datasets or models. It requires a fundamental shift in how wearable HAR is formulated, supervised, and evaluated. This work shows how to model open-ended activity narratives by aligning wearable sensor data with natural-language descriptions in an open-vocabulary setting. Our framework has three core components. First, we introduce a naturalistic data collection and annotation pipeline that combines multi-position wearable sensing with free-form, time-aligned narrative descriptions of ongoing behavior, allowing activity semantics to emerge without a predefined vocabulary. Second, we define a retrieval-based evaluation framework that measures semantic alignment between sensor data and language, enabling principled evaluation without fixed classes while also subsuming closed-set classification as a special case. Third, we present a language-conditioned learning architecture that supports sensor-to-text inference over variable-length sensor streams and heterogeneous sensor placements. Experiments show that models trained with fixed-label objectives degrade sharply under real-world variability, while open-vocabulary sensor-language alignment yields robust and semantically grounded representations. Once this alignment is learned, closed-set activity recognition becomes a simple downstream task. Under cross-participant evaluation, our method achieves 65.3% Macro-F1, compared with 31-34% for strong closed-set HAR baselines. These results establish open-ended narrative modeling as a practical and effective foundation for real-world wearable HAR.

Method: ThoughtSteer perturbs a single embedding vector at the input layer; the model’s own multi-pass reasoning amplifies this perturbation into a hijacked latent trajectory that reliably produces attacker-chosen answers while remaining structurally invisible.

Result: Achieves ≥99% attack success rate across two architectures (Coconut and SimCoT), three reasoning benchmarks, model scales from 124M to 3B parameters; transfers to held-out benchmarks (94-100%); evades all five evaluated active defenses; survives 25 epochs of clean fine-tuning.

Conclusion: Continuous reasoning models have unique vulnerabilities; backdoor perturbations reveal mechanistic interpretability insights about latent trajectories; effective backdoors leave linearly separable signatures in latent space despite individual vectors encoding correct answers.

Abstract: A new generation of language models reasons entirely in continuous hidden states, producing no tokens and leaving no audit trail. We show that this silence creates a fundamentally new attack surface. ThoughtSteer perturbs a single embedding vector at the input layer; the model’s own multi-pass reasoning amplifies this perturbation into a hijacked latent trajectory that reliably produces the attacker’s chosen answer, while remaining structurally invisible to every token-level defense. Across two architectures (Coconut and SimCoT), three reasoning benchmarks, and model scales from 124M to 3B parameters, ThoughtSteer achieves >=99% attack success rate with near-baseline clean accuracy, transfers to held-out benchmarks without retraining (94-100%), evades all five evaluated active defenses, and survives 25 epochs of clean fine-tuning. We trace these results to a unifying mechanism: Neural Collapse in the latent space pulls triggered representations onto a tight geometric attractor, explaining both why defenses fail and why any effective backdoor must leave a linearly separable signature (probe AUC>=0.999). Yet a striking paradox emerges: individual latent vectors still encode the correct answer even as the model outputs the wrong one. The adversarial information is not in any single vector but in the collective trajectory, establishing backdoor perturbations as a new lens for mechanistic interpretability of continuous reasoning. Code and checkpoints are available.

Method: The method associates label prediction with O(1) complexity closest cluster center search in a vector system used as target for latent space configuration. It only requires finding indexes of several largest and lowest values in the embedding vector, making it extremely computationally efficient.

Result: The method achieves up to 11.6 times overall acceleration over conventional methods while maintaining the same training accuracy. It also has unique properties that allow predicting the existence of new classes.

Conclusion: By leveraging known latent space geometry with specific properties, label prediction complexity can be significantly reduced from O(n) to O(1), enabling faster inference without sacrificing accuracy.

Abstract: Label prediction in neural networks (NNs) has O(n) complexity proportional to the number of classes. This holds true for classification using fully connected layers and cosine similarity with some set of class prototypes. In this paper we show that if NN latent space (LS) geometry is known and possesses specific properties, label prediction complexity can be significantly reduced. This is achieved by associating label prediction with the O(1) complexity closest cluster center search in a vector system used as target for latent space configuration (LSC). The proposed method only requires finding indexes of several largest and lowest values in the embedding vector making it extremely computationally efficient. We show that the proposed method does not change NN training accuracy computational results. We also measure the time required by different computational stages of NN inference and label prediction on multiple datasets. The experiments show that the proposed method allows to achieve up to 11.6 times overall acceleration over conventional methods. Furthermore, the proposed method has unique properties which allow to predict the existence of new classes.

Method: Developed Optimus training library with support for standard large model training techniques, pretrained dense and MoE models from scratch on 3,072 GPU tiles, implemented custom GPU kernels for expert computation, and created EP-Aware sharded optimizer for improved performance.

Result: Successfully pretrained models from 1B to 220B parameters, achieved 90% scaling efficiency at 12,288 GPU tiles, obtained up to 1.71x training speedups with custom optimizations, and developed robust reliability features for stable large-scale training.

Conclusion: Demonstrated effective large-scale LLM pretraining on exascale hardware with optimized techniques, showing promising scaling efficiency and performance improvements for MoE models.

Abstract: Pretraining Large Language Models (LLMs) from scratch requires massive amount of compute. Aurora super computer is an ExaScale machine with 127,488 Intel PVC (Ponte Vechio) GPU tiles. In this work, we showcase LLM pretraining on Aurora at the scale of 1000s of GPU tiles. Towards this effort, we developed Optimus, an inhouse training library with support for standard large model training techniques. Using Optimus, we first pretrained Mula-1B, a 1 Billion dense model and Mula-7B-A1B, a 7 Billion Mixture of Experts (MoE) model from scratch on 3072 GPU tiles for the full 4 trillion tokens of the OLMoE-mix-0924 dataset. We then demonstrated model scaling by pretraining three large MoE models Mula-20B-A2B, Mula-100B-A7B, and Mula-220B-A10B till 100 Billion tokens on the same dataset. On our largest model Mula-220B-A10B, we pushed the compute scaling from 384 to 12288 GPU tiles and observed scaling efficiency of around 90% at 12288 GPU tiles. We significantly improved the runtime performance of MoE models using custom GPU kernels for expert computation, and a novel EP-Aware sharded optimizer resulting in training speedups up to 1.71x. As part of the Optimus library, we also developed a robust set of reliability and fault tolerant features to improve training stability and continuity at scale.

Method: MIRANDA applies adversarial regularization to intermediate features instead of final representations, and uses a rank-based objective for year-invariance rather than binary domain classification.

Result: On a 70-year dataset with 67,800 observations of 5 tree species, MIRANDA improves robustness to climatic distribution shifts and narrows the performance gap with mechanistic models, unlike conventional DA approaches.

Conclusion: MIRANDA effectively addresses climate change-induced domain shifts in phenology modeling through intermediate feature adversarial regularization and rank-based objectives.

Abstract: Plant phenology modelling aims to predict the timing of seasonal phases, such as leaf-out or flowering, from meteorological time series. Reliable predictions are crucial for anticipating ecosystem responses to climate change. While phenology modelling has traditionally relied on mechanistic approaches, deep learning methods have recently been proposed as flexible, data-driven alternatives with often superior performance. However, mechanistic models tend to outperform deep networks when data distribution shifts are induced by climate change. Domain Adaptation (DA) techniques could help address this limitation. Yet, unlike standard DA settings, climate change induces a temporal continuum of domains and involves both a covariate and label shift, with warmer records and earlier start of spring. To tackle this challenge, we introduce Mid-feature Rank-adversarial Domain Adaptation (MIRANDA). Whereas conventional adversarial methods enforce domain invariance on final latent representations, an approach that does not explicitly address label shift, we apply adversarial regularization to intermediate features. Moreover, instead of a binary domain-classification objective, we employ a rank-based objective that enforces year-invariance in the learned meteorological representations. On a country-scale dataset spanning 70 years and comprising 67,800 phenological observations of 5 tree species, we demonstrate that, unlike conventional DA approaches, MIRANDA improves robustness to climatic distribution shifts and narrows the performance gap with mechanistic models.

Method: Proposes Routing-Free MoE which eliminates external routers, Softmax, Top-K and load balancing, encapsulating activation functionalities within individual experts optimized through continuous gradient flow. Introduces unified adaptive load-balancing framework for simultaneous optimization of expert-balancing and token-balancing objectives.

Result: Extensive experiments show Routing-Free MoE consistently outperforms baselines with better scalability and robustness. Detailed behavior analysis provides insights for future MoE design and optimization.

Conclusion: Routing-Free MoE offers a flexible, customizable approach to MoE design that eliminates rigid centralized routing, enabling more adaptive and efficient resource allocation through self-determining experts.

Abstract: Standard Mixture-of-Experts (MoE) models rely on centralized routing mechanisms that introduce rigid inductive biases. We propose Routing-Free MoE which eliminates any hard-coded centralized designs including external routers, Softmax, Top-K and load balancing, instead encapsulating all activation functionalities within individual experts and directly optimized through continuous gradient flow, enabling each expert to determine its activation entirely on its own. We introduce a unified adaptive load-balancing framework to simultaneously optimize both expert-balancing and token-balancing objectives through a configurable interpolation, allowing flexible and customizable resource allocation. Extensive experiments show that Routing-Free MoE can consistently outperform baselines with better scalability and robustness. We analyze its behavior in detail and offer insights that may facilitate future MoE design ad optimization.

Method: Uses cost-penalized fitness metrics combined with linear grace periods for newborn experts in a Free-Market Algorithm orchestrated transformer with dynamic MoE management, tested through seven controlled runs including round-trip domain shift experiments.

Result: Achieves 9-11x faster recovery when returning to previously learned domains with zero expert births/replacements required, demonstrating “molecular memory” effect where dormant experts survive and reactivate; preliminary analysis shows $39.1M annual savings and 27.1 GWh energy reduction for OpenAI-scale providers.

Conclusion: The approach enables MoE systems to accumulate domain expertise through diversification rather than replacement, creating efficient “molecular memory” that significantly improves performance and reduces costs compared to current MoE management approaches.

Abstract: We present experimental results from seven controlled runs of nanoFMT, a Free-Market Algorithm (FMA) orchestrated transformer with dynamic Mixture-of-Experts (MoE) management. The experiments address a fundamental question for advanced LLM development: how should an MoE system manage its expert pool when operating at full capacity under changing data distributions? We demonstrate that cost-penalized fitness metrics, combined with a linear grace period for newborn experts, produce a system that accumulates domain expertise through diversification rather than replacement. The central result is a round-trip domain shift experiment showing 9-11x faster recovery when returning to a previously learned domain, with zero expert births or replacements required. This “molecular memory” effect – where dormant experts survive and reactivate when their domain returns – has no analogue in current MoE management approaches. A preliminary cost analysis estimates annual savings of $39.1M and 27.1 GWh energy reduction for an OpenAI-scale provider under a moderate scenario.

Method: Meta-TTL formulates adaptation policy discovery as a bi-level optimization problem. The inner loop executes standard TTL to evaluate how effectively candidate adaptation policies help agents correct errors across sequential episodes. The outer loop uses evolutionary search over a diverse distribution of training tasks to iteratively refine the adaptation policy.

Result: Meta-TTL consistently outperforms hand-crafted baselines on Jericho and WebArena-Lite benchmarks across both in-distribution and out-of-distribution settings, using multiple meta-agent backbones. The optimized adaptation policies encode transferable strategies that generalize beyond the training task distribution.

Conclusion: Learning adaptation policies through meta-learning is superior to hand-crafting them, and the discovered policies generalize well to new tasks and environments, demonstrating the effectiveness of the Meta-TTL framework.

Abstract: Test-Time Learning (TTL) enables language agents to iteratively refine their performance through repeated interactions with the environment at inference time. At the core of TTL is an adaptation policy that updates the actor policy based on experience from previous episodes, thereby improving future behavior. Existing methods rely on fixed, hand-crafted adaptation policies rather than optimizing them for downstream improvement. We argue that optimal adaptation policies should be learned from task environments, not hand-engineered based on human intuition. To achieve this, we introduce Meta-TTL, a framework that formulates the discovery of effective adaptation policies as a bi-level optimization problem. Within this framework, the inner loop executes the standard TTL process, measuring how effectively a candidate adaptation policy helps an agent correct errors across sequential episodes. Guided by the agent’s performance, the outer loop employs evolutionary search over a diverse distribution of training tasks to iteratively refine the adaptation policy. We evaluate Meta-TTL on Jericho and WebArena-Lite across both in-distribution (ID) and out-of-distribution (OOD) settings, using multiple meta-agent backbones. Results on both benchmarks show that Meta-TTL consistently outperforms hand-crafted baselines, suggesting that the optimized adaptation policy encodes transferable strategies that generalize beyond the training task distribution.

Method: Uses second-order Hessian information with Kronecker-factorization to show decomposition must consider both input and output information. Proposes bi-directional whitening of weight matrices as a closed-form optimal decomposition solution.

Result: Achieves ~20-40% better decomposition results than previous state-of-the-art SVD-LLM method, demonstrating superior performance in low-rank approximation of LLM weights.

Conclusion: OBD-LLM provides a theoretically grounded, loss-aware decomposition method that significantly outperforms previous approaches by properly accounting for both input and output information through Hessian analysis.

Abstract: Low-rank decomposition has emerged as an important problem in Large Language Model (LLM) fine-tuning and inference. Through Singular Value Decomposition (SVD), the weight matrix can be factorized into low-rank spaces optimally. Previously, a common practice was to decompose the weight in the activation-whitened space, and then achieve satisfying results. In this work, we propose Optimal Brain Decomposition LLM (OBD-LLM), which studies the decomposition problem in the model space by utilizing second-order Hessian information. Through a rigorous Kronecker-factorization of the Hessian, we show that the decomposition needs to consider both input and output information of the layer, and achieves much better decomposition results compared to input only method. Our loss-aware decomposition method involves a bi-directional whitening on the weight matrix. As a result, OBD-LLM is a closed-form solution for the optimal decomposition of weights in the language model. Remarkably, we achieve ~20-40% better results than previous state-of-the-art decomposition methods, the SVD-LLM.

Method: Introduces Policy Improvement Reinforcement Learning (PIRL) framework that maximizes cumulative policy improvement across iterations, and Policy Improvement Policy Optimization (PIPO) which implements closed-loop optimization through retrospective verification against historical baselines.

Result: Theoretical analysis shows PIPO performs ascent on the PIRL objective in expectation, and experiments on mathematical reasoning benchmarks demonstrate improved stability and performance over GRPO and its variants.

Conclusion: Policy improvement feedback is crucial for stable RL-based post-training of LLMs, and the closed-loop PIRL/PIPO framework transforms open-loop optimization into a self-correcting process that better aligns with final task performance.

Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has become a central post-training paradigm for improving the reasoning capabilities of large language models. Yet existing methods share a common blind spot: they optimize policies based on instantaneous group-level or batch-level statistics without ever verifying whether the resulting update actually improved the model. This open-loop design – updating in isolation at each step, guided only by within-group (batch) reward signals – means optimization can drift or collapse with no mechanism to detect and correct these failures. We argue that the missing ingredient is policy improvement feedback: the ability to measure and optimize inter-iteration progress directly. To this end, we introduce Policy Improvement Reinforcement Learning (PIRL), a framework that replaces surrogate reward maximization with the explicit objective of maximizing cumulative policy improvement across iterations, and prove this temporal objective is perfectly aligned with maximizing final task performance. Building on PIRL, we propose Policy Improvement Policy Optimization (PIPO), which implements closed-loop optimization through retrospective verification. At each iteration, PIPO evaluates whether the previous update yielded genuine improvement against a sliding-window historical baseline, then actively reinforces beneficial updates and suppresses the harmful ones – transforming an open-loop process into a self-correcting one. We provide theoretical analysis showing that PIPO performs ascent on the PIRL objective in expectation, and experiments on mathematical reasoning benchmarks demonstrate improved stability and performance over GRPO and its variants.

Method: Decouples forecasting from spatial resolution using learned generative super-resolution as post-processing. Formulates super-resolution as stochastic inverse problem with residual formulation to preserve large-scale structure while reconstructing unresolved variability. Trained with flow matching exclusively on reanalysis data and applied to global medium-range forecasts.

Result: Super-resolution preserves large-scale structure and variance after re-coarsening, introduces physically consistent small-scale variability, and achieves competitive probabilistic forecast skill at 0.25° resolution relative to operational ensemble baseline, with modest additional training cost compared to end-to-end high-resolution forecasting.

Conclusion: The modular framework successfully reduces computational costs while maintaining forecast quality, demonstrating that generative super-resolution can effectively enhance spatial resolution of weather forecasts without expensive end-to-end training.

Abstract: Machine learning-based weather forecasting models now surpass state-of-the-art numerical weather prediction systems, but training and operating these models at high spatial resolution remains computationally expensive. We present a modular framework that decouples forecasting from spatial resolution by applying learned generative super-resolution as a post-processing step to coarse-resolution forecast trajectories. We formulate super-resolution as a stochastic inverse problem, using a residual formulation to preserve large-scale structure while reconstructing unresolved variability. The model is trained with flow matching exclusively on reanalysis data and is applied to global medium-range forecasts. We evaluate (i) design consistency by re-coarsening super-resolved forecasts and comparing them to the original coarse trajectories, and (ii) high-resolution forecast quality using standard ensemble verification metrics and spectral diagnostics. Results show that super-resolution preserves large-scale structure and variance after re-coarsening, introduces physically consistent small-scale variability, and achieves competitive probabilistic forecast skill at 0.25° resolution relative to an operational ensemble baseline, while requiring only a modest additional training cost compared with end-to-end high-resolution forecasting.

Method: Proposes FALCON (Fatigue-Aware Learning to Defer via Constrained Optimisation) that models workload-varying human performance using fatigue curves. Formulates L2D as a Constrained Markov Decision Process (CMDP) with states including task features and cumulative human workload, optimized via PPO-Lagrangian training under human-AI cooperation budgets.

Result: FALCON consistently outperforms state-of-the-art L2D methods across coverage levels, generalizes zero-shot to unseen experts with different fatigue patterns, and demonstrates advantages of adaptive human-AI collaboration over AI-only or human-only decision-making.

Conclusion: Explicitly modeling human fatigue dynamics significantly improves learning to defer systems, enabling more realistic and effective human-AI collaboration that adapts to varying human performance states.

Abstract: Learning to defer (L2D) enables human-AI cooperation by deciding when an AI system should act autonomously or defer to a human expert. Existing L2D methods, however, assume static human performance, contradicting well-established findings on fatigue-induced degradation. We propose Fatigue-Aware Learning to Defer via Constrained Optimisation (FALCON), which explicitly models workload-varying human performance using psychologically grounded fatigue curves. FALCON formulates L2D as a Constrained Markov Decision Process (CMDP) whose state includes both task features and cumulative human workload, and optimises accuracy under human-AI cooperation budgets via PPO-Lagrangian training. We further introduce FA-L2D, a benchmark that systematically varies fatigue dynamics from near-static to rapidly degrading regimes. Experiments across multiple datasets show that FALCON consistently outperforms state-of-the-art L2D methods across coverage levels, generalises zero-shot to unseen experts with different fatigue patterns, and demonstrates the advantage of adaptive human-AI collaboration over AI-only or human-only decision-making when coverage lies strictly between 0 and 1.

Method: Used Event2Vec (additive sequence embedding model) trained on random walks from human STRING interactome, compared against DeepWalk baseline (Word2Vec) on same walks, both producing 64-dimensional representations.

Result: Compositional structure substantially improves pathway coherence (30.2× vs 2.9× above random), functional analogy accuracy (mean similarity 0.966 vs 0.650), and hierarchical pathway organization, while geometric properties like norm-degree anticorrelation are shared or exceeded by baseline.

Conclusion: Enforced compositionality specifically benefits relational and compositional reasoning tasks in biological networks, suggesting structured embeddings are valuable for biological network analysis.

Abstract: In this work, we study whether enforcing strict compositional structure in sequence embeddings yields meaningful geometric organization when applied to protein-protein interaction networks. Using Event2Vec, an additive sequence embedding model, we train 64-dimensional representations on random walks from the human STRING interactome, and compare against a DeepWalk baseline based on Word2Vec, trained on the same walks. We find that compositional structure substantially improves pathway coherence (30.2$\times$ vs 2.9$\times$ above random), functional analogy accuracy (mean similarity 0.966 vs 0.650), and hierarchical pathway organization, while geometric properties such as norm–degree anticorrelation are shared with or exceeded by the non-compositional baseline. These results indicate that enforced compositionality specifically benefits relational and compositional reasoning tasks in biological networks.

Method: WARP formulates repair as a convex quadratic program derived from first-order linearization of logit gap, enabling optimization over high-dimensional parameter space with sensitivity-based preprocessing for feasibility.

Result: Empirical evaluation on encoder-only Transformers shows the guarantees hold in practice while improving robustness to adversarial inputs, demonstrating achievable guaranteed repair.

Conclusion: Guaranteed, generalizable Transformer repair is achievable through principled constraint-based optimization that extends beyond final layers.

Abstract: Transformer-based NLP models remain vulnerable to adversarial perturbations, yet existing repair methods face a fundamental trade-off: gradient-based approaches offer flexibility but lack verifiability and often overfit; methods that do provide repair guarantees are restricted to the final layer or small networks, significantly limiting the parameter search space available for repair. We present WARP (Weight-Adjusted Repair with Provability), a constraint-based repair framework that extends repair beyond the last layer of Transformer models. WARP formulates repair as a convex quadratic program derived from a first-order linearization of the logit gap, enabling tractable optimization over a high-dimensional parameter space. Under the condition that the first-order approximation holds, this formulation induces three per-sample guarantees: (i) a positive margin constraint ensuring correct classification on repaired inputs, (ii) preservation constraints over a designated remain set, and (iii) a certified robustness radius derived from Lipschitz continuity. To ensure feasibility across varying model architectures, we introduce a sensitivity-based preprocessing step that conditions the optimization landscape accordingly. We further show that the iterative optimization procedure converges to solutions satisfying all repair constraints under mild assumptions. Empirical evaluation on encoder-only Transformers with varying layer architectures validates that these guarantees hold in practice while improving robustness to adversarial inputs. Our results demonstrate that guaranteed, generalizable Transformer repair is achievable through principled constraint-based optimization.

Method: Introduces LT-O-Learners with custom overlap weights that downweight samples with low overlap in treatment or long-term observation. The retargeted loss satisfies Neyman-orthogonality for robustness to nuisance estimation errors.

Result: Theoretical analysis shows LT-O-Learners achieve quasi-oracle rates under certain conditions. Empirical evaluations on synthetic and semi-synthetic benchmarks confirm robustness in low-overlap settings.

Conclusion: LT-O-Learners are the first orthogonal learners for HLTE estimation robust to low overlap, offering model-agnostic approach for personalized decision-making with combined experimental and observational data.

Abstract: Estimation of heterogeneous long-term treatment effects (HLTEs) is widely used for personalized decision-making in marketing, economics, and medicine, where short-term randomized experiments are often combined with long-term observational data. However, HLTE estimation is challenging due to limited overlap in treatment or in observing long-term outcomes for certain subpopulations, which can lead to unstable HLTE estimates with large finite-sample variance. To address this challenge, we introduce the LT-O-learners (Long-Term Orthogonal Learners), a set of novel orthogonal learners for HLTE estimation. The learners are designed for the canonical HLTE setting that combines a short-term randomized dataset $\mathcal{D}_1$ with a long-term historical dataset $\mathcal{D}_2$. The key idea of our LT-O-Learners is to retarget the learning objective by introducing custom overlap weights that downweight samples with low overlap in treatment or in long-term observation. We show that the retargeted loss is equivalent to the weighted oracle loss and satisfies Neyman-orthogonality, which means our learners are robust to errors in the nuisance estimation. We further provide a general error bound for the LT-O-Learners and give the conditions under which quasi-oracle rate can be achieved. Finally, our LT-O-learners are model-agnostic and can thus be instantiated with arbitrary machine learning models. We conduct empirical evaluations on synthetic and semi-synthetic benchmarks to confirm the theoretical properties of our LT-O-Learners, especially the robustness in low-overlap settings. To the best of our knowledge, ours are the first orthogonal learners for HLTE estimation that are robust to low overlap that is common in long-term outcomes.

Method: Proposes FP-DRL algorithm that: 1) Uses flow matching to model policies, offering computational efficiency and ability to fit complex distributions; 2) Employs distributional RL to model and optimize the entire return distribution rather than just the mean.

Result: Achieves state-of-the-art performance on most MuJoCo control tasks and demonstrates superior representation capability of the flow policy.

Conclusion: Combining flow-based policy modeling with distributional RL effectively addresses limitations of traditional RL approaches, enabling better handling of multimodal distributions and improved agent performance.

Abstract: Reinforcement Learning (RL) has proven highly effective in addressing complex control and decision-making tasks. However, in most traditional RL algorithms, the policy is typically parameterized as a diagonal Gaussian distribution, which constrains the policy from capturing multimodal distributions, making it difficult to cover the full range of optimal solutions in multi-solution problems, and the return is reduced to a mean value, losing its multimodal nature and thus providing insufficient guidance for policy updates. In response to these problems, we propose a RL algorithm termed flow-based policy with distributional RL (FP-DRL). This algorithm models the policy using flow matching, which offers both computational efficiency and the capacity to fit complex distributions. Additionally, it employs a distributional RL approach to model and optimize the entire return distribution, thereby more effectively guiding multimodal policy updates and improving agent performance. Experimental trails on MuJoCo benchmarks demonstrate that the FP-DRL algorithm achieves state-of-the-art (SOTA) performance in most MuJoCo control tasks while exhibiting superior representation capability of the flow policy.

Method: Analyze spectral GNNs in graph Fourier domain where each layer becomes element-wise frequency update, separating fixed spectrum from trainable parameters; use Gaussian complexity invariance under Graph Fourier Transform to derive generalization bounds and stability estimates

Result: Derived data-dependent, depth, and order-aware generalization bounds with stability estimates; in linear case, bounds are tighter; on real graphs, data-dependent term correlates with generalization gap across polynomial bases

Conclusion: The analysis provides theoretical understanding of spectral GNNs’ depth and order effects, highlighting practical choices that avoid frequency amplification across layers for better generalization

Abstract: Spectral graph neural networks learn graph filters, but their behavior with increasing depth and polynomial order is not well understood. We analyze these models in the graph Fourier domain, where each layer becomes an element-wise frequency update, separating the fixed spectrum from trainable parameters and making depth and order explicit. In this setting, we show that Gaussian complexity is invariant under the Graph Fourier Transform, which allows us to derive data-dependent, depth, and order-aware generalization bounds together with stability estimates. In the linear case, our bounds are tighter, and on real graphs, the data-dependent term correlates with the generalization gap across polynomial bases, highlighting practical choices that avoid frequency amplification across layers.

Method: Iterative procedure that: (1) privately estimates local means and tangent geometry from reference data, (2) projects query points along estimated subspaces toward local means, (3) uses rigorous privacy accounting across iterations with (ε,δ)-DP.

Result: The framework provides utility guarantees showing corrected queries converge toward the manifold at non-asymptotic rates, with simulations demonstrating accurate signal recovery under moderate privacy budgets.

Conclusion: This work brings differential privacy to manifold methods, enabling privacy-preserving manifold-based workflows in regulated environments while maintaining geometric signal for downstream tasks.

Abstract: We introduce a differentially private manifold denoising framework that allows users to exploit sensitive reference datasets to correct noisy, non-private query points without compromising privacy. The method follows an iterative procedure that (i) privately estimates local means and tangent geometry using the reference data under calibrated sensitivity, (ii) projects query points along the privately estimated subspace toward the local mean via corrective steps at each iteration, and (iii) performs rigorous privacy accounting across iterations and queries using $(\varepsilon,δ)$-differential privacy (DP). Conceptually, this framework brings differential privacy to manifold methods, retaining sufficient geometric signal for downstream tasks such as embedding, clustering, and visualization, while providing formal DP guarantees for the reference data. Practically, the procedure is modular and scalable, separating DP-protected local geometry (means and tangents) from budgeted query-point updates, with a simple scheduler allocating privacy budget across iterations and queries. Under standard assumptions on manifold regularity, sampling density, and measurement noise, we establish high-probability utility guarantees showing that corrected queries converge toward the manifold at a non-asymptotic rate governed by sample size, noise level, bandwidth, and the privacy budget. Simulations and case studies demonstrate accurate signal recovery under moderate privacy budgets, illustrating clear utility-privacy trade-offs and providing a deployable DP component for manifold-based workflows in regulated environments without reengineering privacy systems.

Method: Instead of operating directly on irregular graph structures, EmbedPart leverages node embeddings produced during GNN training and clusters these dense embeddings to derive partitioning. This shifts partitioning from graph structures to dense embeddings.

Result: EmbedPart achieves more than 100x speedup over Metis while maintaining competitive partitioning quality, accelerating distributed GNN training. It also naturally supports graph updates, fast repartitioning, and can improve single-machine GNN training via graph reordering.

Conclusion: By shifting partitioning from irregular graph structures to dense embeddings, EmbedPart enables scalable and high-quality graph data optimization for GNN training.

Abstract: Graph Neural Networks (GNNs) are widely used for learning on graph-structured data, but scaling GNN training to massive graphs remains challenging. To enable scalable distributed training, graphs are divided into smaller partitions that are distributed across multiple machines such that inter-machine communication is minimized and computational load is balanced. In practice, existing partitioning approaches face a fundamental trade-off between partitioning overhead and partitioning quality. We propose EmbedPart, an embedding-driven partitioning approach that achieves both speed and quality. Instead of operating directly on irregular graph structures, EmbedPart leverages node embeddings produced during the actual GNN training workload and clusters these dense embeddings to derive a partitioning. EmbedPart achieves more than 100x speedup over Metis while maintaining competitive partitioning quality and accelerating distributed GNN training. Moreover, EmbedPart naturally supports graph updates and fast repartitioning, and can be applied to graph reordering to improve data locality and accelerate single-machine GNN training. By shifting partitioning from irregular graph structures to dense embeddings, EmbedPart enables scalable and high-quality graph data optimization.

Method: Two transfer learning algorithms: PC-stable-transfer learning (PCS-TL) for constraint-based structure learning, and hill climbing transfer learning (HC-TL) for score-based learning. Both include specific metrics to detect and avoid negative transfer. Parameter estimation uses log-linear pooling. Evaluation involves kernel density estimation Bayesian networks tested on synthetic and UCI datasets with added noise/modifications.

Result: PCS-TL and HC-TL demonstrate reliable improvement in learning performance for nonparametric Bayesian networks with scarce data. Statistical analysis (Friedman test with Bergmann-Hommel post-hoc) shows enhanced experimental behavior compared to models without transfer learning.

Conclusion: The proposed transfer learning methods effectively improve nonparametric Bayesian network learning under data scarcity while avoiding negative transfer, reducing deployment time in industrial applications.

Abstract: This paper introduces two transfer learning methodologies for estimating nonparametric Bayesian networks under scarce data. We propose two algorithms, a constraint-based structure learning method, called PC-stable-transfer learning (PCS-TL), and a score-based method, called hill climbing transfer learning (HC-TL). We also define particular metrics to tackle the negative transfer problem in each of them, a situation in which transfer learning has a negative impact on the model’s performance. Then, for the parameters, we propose a log-linear pooling approach. For the evaluation, we learn kernel density estimation Bayesian networks, a type of nonparametric Bayesian network, and compare their transfer learning performance with the models alone. To do so, we sample data from small, medium and large-sized synthetic networks and datasets from the UCI Machine Learning repository. Then, we add noise and modifications to these datasets to test their ability to avoid negative transfer. To conclude, we perform a Friedman test with a Bergmann-Hommel post-hoc analysis to show statistical proof of the enhanced experimental behavior of our methods. Thus, PCS-TL and HC-TL demonstrate to be reliable algorithms for improving the learning performance of a nonparametric Bayesian network with scarce data, which in real industrial environments implies a reduction in the required time to deploy the network.

Method: Proposes model estimation procedure for tabular POMDPs that exploits connection between filter stability and concentration inequalities for weakly dependent random variables. Uses single trajectory to estimate superstate MDP model with tight sample complexity guarantees.

Result: Obtains tight sample complexity guarantees for estimating superstate MDP model from a single trajectory. Combined with value iteration, yields approximately optimal finite-window policies for POMDP.

Conclusion: Provides sample-efficient approach to learning finite-window policies in tabular POMDPs through superstate MDP model estimation with theoretical guarantees.

Abstract: We study model-based learning of finite-window policies in tabular partially observable Markov decision processes (POMDPs). A common approach to learning under partial observability is to approximate unbounded history dependencies using finite action-observation windows. This induces a finite-state Markov decision process (MDP) over histories, referred to as the superstate MDP. Once a model of this superstate MDP is available, standard MDP algorithms can be used to compute optimal policies, motivating the need for sample-efficient model estimation. Estimating the superstate MDP model is challenging because trajectories are generated by interaction with the original POMDP, creating a mismatch between the sampling process and target model. We propose a model estimation procedure for tabular POMDPs and analyze its sample complexity. Our analysis exploits a connection between filter stability and concentration inequalities for weakly dependent random variables. As a result, we obtain tight sample complexity guarantees for estimating the superstate MDP model from a single trajectory. Combined with value iteration, this yields approximately optimal finite-window policies for the POMDP.

Method: Lightweight probes that take internal representations of LLM checkpoints as input and directly predict downstream task performance (pass@1 success probability). Several probe architectures were designed and validated using OLMo3-7B checkpoints across diverse downstream tasks.

Result: Probes accurately predict checkpoint performance with average AUROC >0.75, demonstrate decent generalizability across checkpoints (earlier checkpoints can predict later ones), and reduce computation latency from ~1 hour to ~3 minutes compared to conventional generative evaluation.

Conclusion: This work presents a practical and scalable in-training downstream evaluation paradigm that enables more agile, informed, and efficient LLM development by providing fast, accurate performance monitoring during training.

Abstract: The paradigm of scaling Large Language Models (LLMs) in both parameter size and test time has pushed the boundaries of AI capabilities, but at the cost of making the traditional generative evaluation paradigm prohibitively expensive, therefore making the latency of LLM’s in-training downstream performance evaluation unbearable. However, simple metrics like training loss (perplexity) are not always correlated with downstream performance, as sometimes their trends diverge from the actual task outcomes. This dilemma calls for a method that is computationally efficient and sufficiently accurate in measuring model capabilities. To address this challenge, we introduce a new in-training evaluation paradigm that uses a lightweight probe for monitoring downstream performance. The probes take the internal representations of LLM checkpoints (during training) as input and directly predict the checkpoint’s performance on downstream tasks measured by success probability (i.e., pass@1). We design several probe architectures, validating their effectiveness using the OLMo3-7B’s checkpoints across a diverse set of downstream tasks. The probes can accurately predict a checkpoint’s performance (with avg. AUROC$>$0.75), have decent generalizability across checkpoints (earlier predicts later), and reduce the computation latency from $\sim$1 hr (using conventional generative evaluation method) to $\sim$3 min. In sum, this work presents a practical and scalable in-training downstream evaluation paradigm, enabling a more agile, informed, and efficient LLM development process.

Method: XOR-SMOO uses SAT oracle queries poly-logarithmic times in approximation parameters γ and δ to obtain γ-approximate Pareto frontiers with probability 1-δ. A γ-approximate frontier is only below the true frontier by a fixed multiplicative factor γ.

Result: Experiments on real-world road network strengthening and supply chain design problems show XOR-SMOO outperforms baselines in identifying Pareto frontiers with higher objective values, better coverage of optimal solutions, and more evenly distributed solutions.

Conclusion: XOR-SMOO significantly enhances the practicality and reliability of SMOO solvers by solving highly intractable SMOO problems with only SAT oracle queries while obtaining tight, constant factor approximation guarantees.

Abstract: Stochastic Multi-Objective Optimization (SMOO) is critical for decision-making trading off multiple potentially conflicting objectives in uncertain environments. SMOO aims at identifying the Pareto frontier, which contains all mutually non-dominating decisions. The problem is highly intractable due to the embedded probabilistic inference, such as computing the marginal, posterior probabilities, or expectations. Existing methods, such as scalarization, sample average approximation, and evolutionary algorithms, either offer arbitrarily loose approximations or may incur prohibitive computational costs. We propose XOR-SMOO, a novel algorithm that with probability $1-δ$, obtains $γ$-approximate Pareto frontiers ($γ>1$) for SMOO by querying an SAT oracle poly-log times in $γ$ and $δ$. A $γ$-approximate Pareto frontier is only below the true frontier by a fixed, multiplicative factor $γ$. Thus, XOR-SMOO solves highly intractable SMOO problems (#P-hard) with only queries to SAT oracles while obtaining tight, constant factor approximation guarantees. Experiments on real-world road network strengthening and supply chain design problems demonstrate that XOR-SMOO outperforms several baselines in identifying Pareto frontiers that have higher objective values, better coverage of the optimal solutions, and the solutions found are more evenly distributed. Overall, XOR-SMOO significantly enhanced the practicality and reliability of SMOO solvers.

Method: ORCA uses conformal prediction and test-time training with a meta-learning procedure that updates calibration modules for each input, providing valid confidence estimates under distributional shifts in reasoning patterns or prompt distributions.

Result: ORCA achieves significant efficiency improvements: up to 47.5% savings on in-distribution tasks with supervised labels and 40.7% with self-consistency labels for Qwen2.5-32B. Under zero-shot out-of-domain settings, it improves MATH-500 savings from 24.8% to 67.0% while maintaining low error rates.

Conclusion: ORCA provides theoretical guarantees on conformal risks while empirically demonstrating higher efficiency and generalization across reasoning tasks, offering a practical solution to reduce computational costs in LLM reasoning.

Abstract: While test-time scaling has enabled large language models to solve highly difficult tasks, state-of-the-art results come at exorbitant compute costs. These inefficiencies can be attributed to the miscalibration of post-trained language models, and the lack of calibration in popular sampling techniques. Here, we present Online Reasoning Calibration (ORCA), a framework for calibrating the sampling process that draws upon conformal prediction and test-time training. Specifically, we introduce a meta-learning procedure that updates the calibration module for each input. This allows us to provide valid confidence estimates under distributional shift, e.g. in thought patterns that occur across different stages of reasoning, or in prompt distributions between model development and deployment. ORCA not only provides theoretical guarantees on conformal risks, but also empirically shows higher efficiency and generalization across different reasoning tasks. At risk level $δ=0.1$, ORCA improves Qwen2.5-32B efficiency on in-distribution tasks with savings up to 47.5% with supervised labels and 40.7% with self-consistency labels. Under zero-shot out-of-domain settings, it improves MATH-500 savings from 24.8% of the static calibration baseline to 67.0% while maintaining a low empirical error rate, and the same trend holds across model families and downstream benchmarks. Our code is publicly available at https://github.com/wzekai99/ORCA.

Method: Develops information-geometric analysis interpreting pseudo-Gibbs sampling steps as m-projections onto full conditional manifolds. Introduces full conditional divergence and derives an upper bound characterizing stationary distribution location. Reformulates structure and parameter learning as optimization problems decomposable into independent subproblems per node.

Result: Proves that learned model distribution converges to true underlying distribution as training samples increase to infinity. Experiments confirm the proposed upper bound is tight in practice.

Conclusion: Provides rigorous theoretical foundation for dependency networks through information-geometric analysis, enabling provable convergence guarantees and improved learning algorithms for complex system modeling.

Abstract: Dependency networks (Heckerman et al., 2000) provide a flexible framework for modeling complex systems with many variables by combining independently learned local conditional distributions through pseudo-Gibbs sampling. Despite their computational advantages over Bayesian and Markov networks, the theoretical foundations of dependency networks remain incomplete, primarily because their model distributions – defined as stationary distributions of pseudo-Gibbs sampling – lack closed-form expressions. This paper develops an information-geometric analysis of pseudo-Gibbs sampling, interpreting each sampling step as an m-projection onto a full conditional manifold. Building on this interpretation, we introduce the full conditional divergence and derive an upper bound that characterizes the location of the stationary distribution in the space of probability distributions. We then reformulate both structure and parameter learning as optimization problems that decompose into independent subproblems for each node, and prove that the learned model distribution converges to the true underlying distribution as the number of training samples grows to infinity. Experiments confirm that the proposed upper bound is tight in practice.

Method: Multiscreen architecture replaces softmax attention with “screening” mechanism. Instead of redistributing attention across all keys, screening evaluates each key against an explicit threshold, discards irrelevant keys that don’t meet the threshold, and aggregates only the remaining relevant keys. This removes global competition among keys and enables absolute relevance assessment.

Result: Multiscreen achieves comparable validation loss with ~40% fewer parameters than Transformer baselines, enables stable optimization at larger learning rates, maintains strong long-context perplexity, shows minimal degradation in retrieval performance beyond training context length, and reduces inference latency by up to 3.2× at 100K context length.

Conclusion: Screening mechanism provides a more efficient alternative to softmax attention by enabling absolute query-key relevance assessment, leading to parameter-efficient models with better long-context capabilities and faster inference, particularly for long sequences.

Abstract: A core limitation of standard softmax attention is that it does not define a notion of absolute query–key relevance: attention weights are obtained by redistributing a fixed unit mass across all keys according to their relative scores. As a result, relevance is defined only relative to competing keys, and irrelevant keys cannot be explicitly rejected. We introduce Multiscreen, a language-model architecture built around a mechanism we call screening, which enables absolute query–key relevance. Instead of redistributing attention across all keys, screening evaluates each key against an explicit threshold, discarding irrelevant keys and aggregating the remaining keys, thereby removing global competition among keys. Across experiments, Multiscreen achieves comparable validation loss with approximately 40% fewer parameters than a Transformer baseline, enables stable optimization at substantially larger learning rates, maintains strong performance in long-context perplexity, shows little to no degradation in retrieval performance even far beyond the training context length, and reduces inference latency by up to 3.2$\times$ at 100K context length.

Method: System uses Kinect 2.0 depth sensor and optical camera for markerless motion capture. Extracts 18 kinematic features from four biomechanical dimensions. Develops two modules: personalized optimal throwing trajectory model combining historical high-quality samples with minimum jerk criterion, and motion deviation diagnosis model based on z-scores and hierarchical logic.

Result: Collected 2,396 throwing samples from professional and non-professional athletes. System generates smooth personalized reference trajectories consistent with natural human movement. Case studies show it can detect poor trunk stability, abnormal elbow displacement, and imbalanced velocity control, providing targeted recommendations.

Conclusion: The framework shifts dart evaluation from deviation from uniform standards to deviation from individual’s optimal control range, improving personalization and interpretability for darts training and other high-precision target sports.

Abstract: As sports training becomes more data-driven, traditional dart coaching based mainly on experience and visual observation is increasingly inadequate for high-precision, goal-oriented movements. Although prior studies have highlighted the importance of release parameters, joint motion, and coordination in dart throwing, most quantitative methods still focus on local variables, single-release metrics, or static template matching. These approaches offer limited support for personalized training and often overlook useful movement variability. This paper presents a data-driven dart training assistance system. The system creates a closed-loop framework spanning motion capture, feature modeling, and personalized feedback. Dart-throwing data were collected in markerless conditions using a Kinect 2.0 depth sensor and an optical camera. Eighteen kinematic features were extracted from four biomechanical dimensions: three-link coordination, release velocity, multi-joint angular configuration, and postural stability. Two modules were developed: a personalized optimal throwing trajectory model that combines historical high-quality samples with the minimum jerk criterion, and a motion deviation diagnosis and recommendation model based on z-scores and hierarchical logic. A total of 2,396 throwing samples from professional and non-professional athletes were collected. Results show that the system generates smooth personalized reference trajectories consistent with natural human movement. Case studies indicate that it can detect poor trunk stability, abnormal elbow displacement, and imbalanced velocity control, then provide targeted recommendations. The framework shifts dart evaluation from deviation from a uniform standard to deviation from an individual’s optimal control range, improving personalization and interpretability for darts training and other high-precision target sports.

Method: Three-stage pipeline: (1) Extract Lowest Floor Elevation (LFE) and height difference from street grade using Elev-Vision framework on Google Street View imagery, (2) Impute missing values with Random Forest and Gradient Boosting models trained on terrain/hydrologic/geographic features, (3) Integrate with flood inundation data and damage functions to estimate interior flood depth and expected loss.

Result: Successfully extracted elevation data for 49% of 12,241 structures; imputation performed well for 13 of 18 areas with R² values 0.159-0.974; enabled structure-level flood damage estimates beyond just hazard exposure.

Conclusion: Street-view-based elevation mapping provides a scalable, replicable framework for regional flood-risk characterization, advancing from pilot-scale to operational workflow for jurisdictions lacking traditional elevation data.

Abstract: This paper argues that AI-enabled analysis of street-view imagery, complemented by performance-gated machine-learning imputation, provides a viable pathway for generating building-specific elevation data at regional scale for flood risk assessment. We develop and apply a three-stage pipeline across 18 areas of interest (AOIs) in Texas that (1) extracts LFE and the height difference between street grade and the lowest floor (HDSL) from Google Street View imagery using the Elev-Vision framework, (2) imputes missing HDSL values with Random Forest and Gradient Boosting models trained on 16 terrain, hydrologic, geographic, and flood-exposure features, and (3) integrates the resulting elevation dataset with Fathom 1-in-100 year inundation surfaces and USACE depth-damage functions to estimate property-specific interior flood depth and expected loss. Across 12,241 residential structures, street-view imagery was available for 73.4% of parcels and direct LFE/HDSL extraction was successful for 49.0% (5,992 structures). Imputation was retained for 13 AOIs where cross-validated performance was defensible, with selected models achieving R suqre values from 0.159 to 0.974; five AOIs were explicitly excluded from prediction because performance was insufficient. The results show that street-view-based elevation mapping is not universally available for every property, but it is sufficiently scalable to materially improve regional flood-risk characterization by moving beyond hazard exposure to structure-level estimates of interior inundation and expected damage. Scientifically, the study advances LFE estimation from a pilot-scale proof of concept to a regional, end-to-end workflow. Practically, it offers a replicable framework for jurisdictions that lack comprehensive Elevation Certificates but need parcel-level information to support mitigation, planning, and flood-risk management.

Method: Systematic evaluation of multiple reasoning models across three scenarios: (1) problems with lengthy irrelevant context, (2) multi-turn conversational settings with independent tasks, and (3) problems presented as subtasks within complex tasks.

Result: Reasoning models produce up to 50% shorter reasoning traces under different context conditions compared to isolated problems. This compression correlates with decreased self-verification and uncertainty management behaviors like double-checking.

Conclusion: The robustness of reasoning models and context management for LLMs needs attention, as behavioral shifts in different contexts may affect performance on challenging tasks despite maintaining performance on straightforward problems.

Abstract: Large language models (LLMs) exhibiting test-time scaling behavior, such as extended reasoning traces and self-verification, have demonstrated remarkable performance on complex, long-term reasoning tasks. However, the robustness of these reasoning behaviors remains underexplored. To investigate this, we conduct a systematic evaluation of multiple reasoning models across three scenarios: (1) problems augmented with lengthy, irrelevant context; (2) multi-turn conversational settings with independent tasks; and (3) problems presented as a subtask within a complex task. We observe an interesting phenomenon: reasoning models tend to produce much shorter reasoning traces (up to 50%) for the same problem under different context conditions compared to the traces produced when the problem is presented in isolation. A finer-grained analysis reveals that this compression is associated with a decrease in self-verification and uncertainty management behaviors, such as double-checking. While this behavioral shift does not compromise performance on straightforward problems, it might affect performance on more challenging tasks. We hope our findings draw additional attention to both the robustness of reasoning models and the problem of context management for LLMs and LLM-based agents.

Method: Proposes a data-driven distribution alignment framework that pre-trains generative models on fully observed simulation data, then aligns them with partial experimental observations using Adversarial Distribution Alignment (ADA), specifically applied to atomic positions with Boltzmann distributions.

Result: Theoretically proves the method recovers target observable distributions even with multiple correlated observables, and empirically validates on synthetic, molecular, and experimental protein data, showing successful alignment with diverse observables.

Conclusion: ADA provides an effective framework for bridging simulation-experiment gaps by leveraging generative models to align imperfect simulations with partial experimental observations, with applications in physical sciences.

Abstract: A fundamental challenge in science and engineering is the simulation-to-experiment gap. While we often possess prior knowledge of physical laws, these physical laws can be too difficult to solve exactly for complex systems. Such systems are commonly modeled using simulators, which impose computational approximations. Meanwhile, experimental measurements more faithfully represent the real world, but experimental data typically consists of observations that only partially reflect the system’s full underlying state. We propose a data-driven distribution alignment framework that bridges this simulation-to-experiment gap by pre-training a generative model on fully observed (but imperfect) simulation data, then aligning it with partial (but real) observations of experimental data. While our method is domain-agnostic, we ground our approach in the physical sciences by introducing Adversarial Distribution Alignment (ADA). This method aligns a generative model of atomic positions – initially trained on a simulated Boltzmann distribution – with the distribution of experimental observations. We prove that our method recovers the target observable distribution, even with multiple, potentially correlated observables. We also empirically validate our framework on synthetic, molecular, and experimental protein data, demonstrating that it can align generative models with diverse observables. Our code is available at https://kaityrusnelson.com/ada/.

Method: NeuroDDAF integrates: (1) GRU-Graph Attention encoder for temporal dynamics and wind-aware spatial interactions, (2) Fourier-domain diffusion-advection module with learnable residuals, (3) wind-modulated latent Neural ODE for continuous-time evolution, and (4) evidential fusion mechanism to combine physics-guided and neural forecasts with uncertainty quantification.

Result: Outperforms strong baselines including AirPhyNet on four urban datasets (Beijing, Shenzhen, Tianjin, Ancona) across 1-3 day horizons, achieving up to 9.7% RMSE reduction and 9.4% MAE reduction on long-term forecasts. Best performance on Beijing dataset with RMSE of 41.63 μg/m³ (1-day) and 48.88 μg/m³ (3-day).

Conclusion: NeuroDDAF effectively unifies neural representation learning with physics-based transport modeling, improving forecasting accuracy, cross-city generalization, and providing well-calibrated uncertainty estimates for air quality prediction.

Abstract: Accurate air quality forecasting is crucial for protecting public health and guiding environmental policy, yet it remains challenging due to nonlinear spatiotemporal dynamics, wind-driven transport, and distribution shifts across regions. Physics-based models are interpretable but computationally expensive and often rely on restrictive assumptions, whereas purely data-driven models can be accurate but may lack robustness and calibrated uncertainty. To address these limitations, we propose Neural Dynamic Diffusion-Advection Fields (NeuroDDAF), a physics-informed forecasting framework that unifies neural representation learning with open-system transport modeling. NeuroDDAF integrates (i) a GRU-Graph Attention encoder to capture temporal dynamics and wind-aware spatial interactions, (ii) a Fourier-domain diffusion-advection module with learnable residuals, (iii) a wind-modulated latent Neural ODE to model continuous-time evolution under time-varying connectivity, and (iv) an evidential fusion mechanism that adaptively combines physics-guided and neural forecasts while quantifying uncertainty. Experiments on four urban datasets (Beijing, Shenzhen, Tianjin, and Ancona) across 1-3 day horizons show that NeuroDDAF consistently outperforms strong baselines, including AirPhyNet, achieving up to 9.7% reduction in RMSE and 9.4% reduction in MAE on long-term forecasts. On the Beijing dataset, NeuroDDAF attains an RMSE of 41.63 $μ$g/m$^3$ for 1-day prediction and 48.88 $μ$g/m$^3$ for 3-day prediction, representing the best performance among all compared methods. In addition, NeuroDDAF improves cross-city generalization and yields well-calibrated uncertainty estimates, as confirmed by ensemble variance analysis and case studies under varying wind conditions.

Method: An agentic evolutionary framework where core evolution operators (pair selection, crossover, mutation, review) are implemented as LLM agents. Features structured scientific artifacts (theory+code or code_only), reviewer judgments as first-class selection gates, and split mutation pathways (exploration for novelty via domain import, correction for evidence-guided repair).

Result: Demonstrated on three benchmark-grounded studies: transformer hyper-connection evolution, optimizer discovery on fixed nanoGPT stack, and native-optimizer ablation. Supports explicit metric direction, reproducible persistence, and reviewer-gated comparison under controlled search conditions.

Conclusion: CliffSearch provides a discovery workflow that prioritizes scientific interpretability and correctness while optimizing task metrics under controlled novelty constraints, rather than maximizing candidate throughput alone.

Abstract: Scientific algorithm discovery is iterative: hypotheses are proposed, implemented, stress-tested, and revised. Current LLM-guided search systems accelerate proposal generation, but often under-represent scientific structure by optimizing code-only artifacts with weak correctness/originality gating. We present CliffSearch, an agentic evolutionary framework in which the core evolution operators (pair selection, crossover, mutation, and review) are implemented as LLM agents, and the loop is designed around three principles: (1) each node is a structured scientific artifact, instantiated in either theory+code or code_only mode, (2) reviewer judgments of correctness and originality are first-class selection gates alongside optimization of the benchmark metric of interest, and (3) mutation is split into exploration and correction pathways with distinct objectives. Exploration mutation imports ideas from adjacent scientific domains to increase novelty, while correction mutation performs targeted evidence-guided repair using reviewer signals over theory, code, benchmark results, and runtime errors. We illustrate the framework on three benchmark-grounded studies: transformer hyper-connection evolution, optimizer discovery on a fixed nanoGPT stack, and a smaller native-optimizer ablation. Across these settings, the same loop supports explicit metric direction, reproducible persistence, and reviewer-gated comparison of discoveries under controlled search conditions. The result is a discovery workflow that prioritizes scientific interpretability and correctness while optimizing task metrics under controlled novelty constraints, rather than maximizing candidate throughput alone. Full run artifacts, interactive visualizations, and exported best nodes for the reported studies are available at https://cliffsearch.ai .

Method: Develops theoretical framework combining approximation theory on sphere, dynamical systems theory, information theory, and statistical learning theory. Creates Learning Pipeline Error Decomposition, Loss Function Spectral Theory, and Out-of-Distribution Extrapolation Bounds. Empirically validates across ten architecturally diverse AI weather models using NVIDIA Earth2Studio with ERA5 initial conditions, evaluating six metrics across 30 initialization dates.

Result: Results confirm: 1) universal spectral energy loss at high wavenumbers for MSE-trained models, 2) rising Error Consensus Ratios showing majority of forecast error is shared across architectures, and 3) linear negative bias during extreme events. Provides Holistic Model Assessment Score for unified multi-dimensional evaluation.

Conclusion: Estimation error (loss- and data-dependent) dominates approximation error (architecture-dependent) at current scales. The framework enables mathematical evaluation of proposed pipelines before training and explains systematic underestimation of record-breaking extremes in data-driven models.

Abstract: AI weather prediction has advanced rapidly, yet no unified mathematical framework explains what determines forecast skill. Existing theory addresses specific architectural choices rather than the learning pipeline as a whole, while operational evidence from 2023-2026 demonstrates that training methodology, loss function design, and data diversity matter at least as much as architecture selection. This paper makes two interleaved contributions. Theoretically, we construct a framework rooted in approximation theory on the sphere, dynamical systems theory, information theory, and statistical learning theory that treats the complete learning pipeline (architecture, loss function, training strategy, data distribution) rather than architecture alone. We establish a Learning Pipeline Error Decomposition showing that estimation error (loss- and data-dependent) dominates approximation error (architecture-dependent) at current scales. We develop a Loss Function Spectral Theory formalizing MSE-induced spectral blurring in spherical harmonic coordinates, and derive Out-of-Distribution Extrapolation Bounds proving that data-driven models systematically underestimate record-breaking extremes with bias growing linearly in record exceedance. Empirically, we validate these predictions via inference across ten architecturally diverse AI weather models using NVIDIA Earth2Studio with ERA5 initial conditions, evaluating six metrics across 30 initialization dates spanning all seasons. Results confirm universal spectral energy loss at high wavenumbers for MSE-trained models, rising Error Consensus Ratios showing that the majority of forecast error is shared across architectures, and linear negative bias during extreme events. A Holistic Model Assessment Score provides unified multi-dimensional evaluation, and a prescriptive framework enables mathematical evaluation of proposed pipelines before training.

Method: Three-stage pipeline: (1) SHRED model pre-trained on simulation data maps sensor time-histories to structured latent space; (2) Temporal sequence model trained on simulation-derived latent trajectories learns to propagate latent states forward/backward in time; (3) At deployment, uses short observation window of sparse sensor measurements with frozen models to reconstruct/forecast complete spatiotemporal trajectory.

Result: Evaluated on six experiments spanning complex spatio-temporal physics including turbulent flows, multiscale propulsion physics, volatile combustion transients, and satellite-derived environmental fields, demonstrating effectiveness under extreme observational constraints.

Conclusion: LAPIS-SHRED provides a lightweight, modular architecture suitable for operational settings with physical/logistical observation constraints, supporting bidirectional inference and inheriting data assimilation/multiscale reconstruction capabilities.

Abstract: Reconstructing full spatio-temporal dynamics from sparse observations in both space and time remains a central challenge in complex systems, as measurements can be spatially incomplete and can be also limited to narrow temporal windows. Yet approximating the complete spatio-temporal trajectory is essential for mechanistic insight and understanding, model calibration, and operational decision-making. We introduce LAPIS-SHRED (LAtent Phase Inference from Short time sequence using SHallow REcurrent Decoders), a modular architecture that reconstructs and/or forecasts complete spatiotemporal dynamics from sparse sensor observations confined to short temporal windows. LAPIS-SHRED operates through a three-stage pipeline: (i) a SHRED model is pre-trained entirely on simulation data to map sensor time-histories into a structured latent space, (ii) a temporal sequence model, trained on simulation-derived latent trajectories, learns to propagate latent states forward or backward in time to span unobserved temporal regions from short observational time windows, and (iii) at deployment, only a short observation window of hyper-sparse sensor measurements from the true system is provided, from which the frozen SHRED model and the temporal model jointly reconstruct or forecast the complete spatiotemporal trajectory. The framework supports bidirectional inference, inherits data assimilation and multiscale reconstruction capabilities from its modular structure, and accommodates extreme observational constraints including single-frame terminal inputs. We evaluate LAPIS-SHRED on six experiments spanning complex spatio-temporal physics: turbulent flows, multiscale propulsion physics, volatile combustion transients, and satellite-derived environmental fields, highlighting a lightweight, modular architecture suited for operational settings where observation is constrained by physical or logistical limitations.

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 2306.14052: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2306.14052&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 2406.08097: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2406.08097&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 2406.15904: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2406.15904&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 determine conclusion as paper content could not be retrieved

Abstract: Failed to fetch summary for 2411.08687: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2411.08687&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 2501.09821: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2501.09821&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2502.01556: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2502.01556&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Method unknown - abstract retrieval failed with 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 2502.01861: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2502.01861&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 2502.21060: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2502.21060&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions as paper content is unavailable

Abstract: Failed to fetch summary for 2503.07555: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2503.07555&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 2503.22244: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2503.22244&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 2504.16624: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2504.16624&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to API access issues

Result: Unable to determine results due to API access issues

Conclusion: Unable to determine conclusion due to API access issues

Abstract: Failed to fetch summary for 2505.04738: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.04738&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2505.23459: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.23459&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2506.23053: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.23053&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2508.09281: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.09281&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content.

Result: Cannot determine results without access to paper content.

Conclusion: Cannot draw conclusions without access to paper content.

Abstract: Failed to fetch summary for 2508.13749: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.13749&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 2509.07013: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.07013&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed API request

Result: Unable to determine results due to failed API request

Conclusion: Unable to determine conclusion due to failed API request

Abstract: Failed to fetch summary for 2509.14078: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.14078&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to the paper abstract

Result: Cannot determine results without access to the paper abstract

Conclusion: Cannot draw conclusions without access to the paper abstract

Abstract: Failed to fetch summary for 2509.15517: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.15517&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 2509.17180: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.17180&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 2511.02272: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.02272&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 2511.13018: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.13018&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 2512.00580: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.00580&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 2512.16383: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.16383&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available due to API request failure

Result: No results available - arXiv API returned HTTP 429 (Too Many Requests) error

Conclusion: Cannot analyze paper content due to technical limitations in accessing the abstract

Abstract: Failed to fetch summary for 2601.02728: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.02728&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to the paper abstract

Result: Cannot determine results without access to the paper abstract

Conclusion: Cannot draw conclusions without access to the paper abstract

Abstract: Failed to fetch summary for 2601.04462: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.04462&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 2601.12604: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.12604&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed API request

Result: Unable to determine results due to failed API request

Conclusion: Unable to determine conclusion due to failed API request

Abstract: Failed to fetch summary for 2601.17196: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.17196&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 2602.00577: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.00577&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 2602.10905: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.10905&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2602.11618: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.11618&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 2602.17477: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.17477&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to missing abstract content.

Result: Unable to determine results due to missing abstract content.

Conclusion: Unable to determine conclusion due to missing abstract content.

Abstract: Failed to fetch summary for 2603.10305: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.10305&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 2603.20527: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.20527&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to technical error in fetching paper content

Result: Unable to determine results due to technical error in fetching paper content

Conclusion: Unable to draw conclusions due to technical error in fetching paper content

Abstract: Failed to fetch summary for 2603.20826: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.20826&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2603.22455: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.22455&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to retrieval failure

Result: Unable to determine results due to retrieval failure

Conclusion: Unable to determine conclusion due to retrieval failure

Abstract: Failed to fetch summary for 2603.24587: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.24587&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2603.28117: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.28117&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to API access limitations

Result: Unable to determine results due to API access limitations

Conclusion: Unable to determine conclusion due to API access limitations

Abstract: Failed to fetch summary for 2603.28595: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.28595&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 2603.29068: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.29068&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 2603.29264: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.29264&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: N/A - Paper content not available for analysis

Result: HTTP 429 error indicates rate limiting on arXiv API requests

Conclusion: Cannot analyze paper relevance without access to the actual abstract content

Abstract: Failed to fetch summary for 2307.14012: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2307.14012&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to missing paper content

Result: Cannot determine results due to missing paper content

Conclusion: Cannot determine conclusion due to missing paper content

Abstract: Failed to fetch summary for 2309.00125: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2309.00125&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 2405.03083: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2405.03083&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 2405.15132: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2405.15132&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 2410.22729: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2410.22729&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 2503.19091: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2503.19091&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 2504.09279: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2504.09279&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to technical error in accessing paper content

Result: Unable to determine results due to technical error in accessing paper content

Conclusion: Unable to determine conclusion due to technical error in accessing paper content

Abstract: Failed to fetch summary for 2505.19367: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.19367&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 2505.21580: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.21580&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot analyze method as paper content is unavailable due to HTTP 429 error

Result: No results available due to API rate limiting preventing access to paper information

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

Abstract: Failed to fetch summary for 2508.02473: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.02473&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2508.21236: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.21236&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to access error

Result: Cannot determine results due to access error

Conclusion: Cannot determine conclusion due to access error

Abstract: Failed to fetch summary for 2509.19162: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.19162&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 2510.08095: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.08095&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 2510.09534: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.09534&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 2510.15669: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.15669&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 2511.12749: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.12749&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 2512.02418: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.02418&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 2512.03743: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.03743&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 2512.10394: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.10394&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 2512.18147: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.18147&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available - paper content inaccessible

Result: No results available - paper summary fetch failed

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

Abstract: Failed to fetch summary for 2601.20589: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.20589&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 2603.21291: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.21291&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 2603.23561: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.23561&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 determine conclusion due to abstract fetch failure

Abstract: Failed to fetch summary for 2603.27592: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.27592&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 2603.27936: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.27936&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Proposes a comprehensive framework categorizing unlearning into three contexts: state unlearning (forgetting specific states/items), trajectory unlearning (forgetting sequences of actions), and environment unlearning (forgetting entire environments/task categories). Introduces a natural language-based method that trains a conversion model to transform high-level unlearning requests into actionable prompts guiding agents through controlled forgetting.

Result: Experimental results show the approach effectively enables agents to forget targeted knowledge while preserving performance on untargeted tasks, and prevents adversaries from inferring the forgotten knowledge through the introduced unlearning inference adversary.

Conclusion: The paper initiates research on unlearning in LLM-based agents, providing a framework and method for selective forgetting that addresses privacy and knowledge management concerns while maintaining agent performance on relevant tasks.

Abstract: Large language model (LLM)-based agents have recently gained considerable attention due to the powerful reasoning capabilities of LLMs. Existing research predominantly focuses on enhancing the task performance of these agents in diverse scenarios. However, as LLM-based agents become increasingly integrated into real-world applications, significant concerns emerge regarding their accumulation of sensitive or outdated knowledge. Addressing these concerns requires the development of mechanisms that allow agents to selectively forget previously learned knowledge, giving rise to a new term LLM-based agent unlearning. This paper initiates research on unlearning in LLM-based agents. Specifically, we propose a novel and comprehensive framework that categorizes unlearning scenarios into three contexts: state unlearning (forgetting specific states or items), trajectory unlearning (forgetting sequences of actions) and environment unlearning (forgetting entire environments or categories of tasks). Within this framework, we introduce a natural language-based unlearning method that trains a conversion model to transform high-level unlearning requests into actionable unlearning prompts, guiding agents through a controlled forgetting process. Moreover, to evaluate the robustness of the proposed framework, we introduce an unlearning inference adversary capable of crafting prompts, querying agents, and observing their behaviors in an attempt to infer the forgotten knowledge. Experimental results show that our approach effectively enables agents to forget targeted knowledge while preserving performance on untargeted tasks, and prevents the adversary from inferring the forgotten knowledge.

Method: Uses common information framework to handle partial observability, introduces internal state to compress accumulated information, and implements internal state-based natural policy gradient method with finite-state controllers to find Nash equilibria.

Result: Establishes non-asymptotic convergence bound with two interpretable components: statistical error term and approximation error. Simulations show finite-state controllers outperform methods using only current observations across multiple partially observable environments.

Conclusion: The proposed approach effectively addresses challenges in partially observable multi-agent reinforcement learning with theoretical guarantees and practical performance improvements using finite-state controllers.

Abstract: This letter studies multi-agent reinforcement learning in partially observable Markov potential games. Solving this problem is challenging due to partial observability, decentralized information, and the curse of dimensionality. First, to address the first two challenges, we leverage the common information framework, which allows agents to act based on both shared and local information. Second, to ensure tractability, we study an internal state that compresses accumulated information, preventing it from growing unboundedly over time. We then implement an internal state-based natural policy gradient method to find Nash equilibria of the Markov potential game. Our main contribution is to establish a non-asymptotic convergence bound for this method. Our theoretical bound decomposes into two interpretable components: a statistical error term that also arises in standard Markov potential games, and an approximation error capturing the use of finite-state controllers. Finally, simulations across multiple partially observable environments demonstrate that the proposed method using finite-state controllers achieves consistent improvements in performance compared to the setting where only the current observation is used.

Method: Each agent maintains an explicit knowledge base, solves role-conditioned local decision problems, and coordinates through lightweight contract primitives. Communication scope expands only when local validation indicates current scope is insufficient, with explicit budget controls for who is contacted and how far coordination propagates.

Result: The paper evaluates CASCADE on disrupted manufacturing and supply-chain settings using unified diagnostics to test whether explicit scope control yields useful quality-latency-communication trade-offs and improved robustness under uncertainty.

Conclusion: CASCADE provides a mechanism for explicit communication scope control in multi-agent replanning, separating agent substrate from scoped interaction layer to improve robustness and efficiency under disruption propagation.

Abstract: Industrial disruption replanning demands multi-agent coordination under strict latency and communication budgets, where disruptions propagate through tightly coupled physical dependencies and rapidly invalidate baseline schedules and commitments. Existing coordination schemes often treat communication as either effectively free (broadcast-style escalation) or fixed in advance (hand-tuned neighborhoods), both of which are brittle once the disruption footprint extends beyond a local region. We present \CASCADE, a budgeted replanning mechanism that makes communication scope explicit and auditable rather than fixed or implicit. Each agent maintains an explicit knowledge base, solves role-conditioned local decision problems to revise commitments, and coordinates through lightweight contract primitives whose footprint expands only when local validation indicates that the current scope is insufficient. This design separates a unified agent substrate (Knowledge Base / Decision Manager / Communication Manager) from a scoped interaction layer that controls who is contacted, how far coordination propagates, and when escalation is triggered under explicit budgets. We evaluate \CASCADE on disrupted manufacturing and supply-chain settings using unified diagnostics intended to test a mechanism-design claim – whether explicit scope control yields useful quality-latency-communication trade-offs and improved robustness under uncertainty – rather than to provide a complete algorithmic ranking.

Method: Study LLM agent interactions in two standard games (network resource allocation and Cournot competition), use multi-round prompts with non-zero-sum context, and analyze behavior through chain-of-thought reasoning

Result: LLM agents tend to cooperate rather than converge to Nash equilibria, with fairness reasoning being central to this behavior as revealed by chain-of-thought analysis

Conclusion: LLM agents exhibit cooperative behavior in competitive settings due to fairness reasoning, requiring new analytical frameworks to understand their strategic dynamics

Abstract: Large language model (LLM) agents are increasingly deployed in competitive multi-agent settings, raising fundamental questions about whether they converge to equilibria and how their strategic behavior can be characterized. In this paper, we study LLM agent interactions in two standard games: a network resource allocation game and a Cournot competition game. Rather than converging to Nash equilibria, we find that LLM agents tend to cooperate when given multi-round prompts and non-zero-sum context. Chain-of-thought analysis reveals that fairness reasoning is central to this behavior. We propose an analytical framework that captures the dynamics of LLM agent reasoning across rounds and explains these experimental findings.

Method: Proposes Gradient Realignment via Active Shared Perception (GRASP) framework that defines generalized Bellman equilibrium as stable objective. Uses independent agent gradients to derive consensus gradient, enabling active perception of policy updates and optimizing team collaboration. Theoretically proven using Kakutani Fixed-Point Theorem.

Result: Extensive experiments on StarCraft II Multi-Agent Challenge (SMAC) and Google Research Football (GRF) demonstrate scalability and promising performance improvements over existing approaches.

Conclusion: GRASP effectively addresses non-stationarity in multi-agent systems by enabling active shared perception, reducing equilibrium oscillations, and accelerating convergence through consensus gradient alignment.

Abstract: Non-stationarity arises from concurrent policy updates and leads to persistent environmental fluctuations. Existing approaches like Centralized Training with Decentralized Execution (CTDE) and sequential update schemes mitigate this issue. However, since the perception of the policies of other agents remains dependent on sampling environmental interaction data, the agent essentially operates in a passive perception state. This inevitably triggers equilibrium oscillations and significantly slows the convergence speed of the system. To address this issue, we propose Gradient Realignment via Active Shared Perception (GRASP), a novel framework that defines generalized Bellman equilibrium as a stable objective for policy evolution. The core mechanism of GRASP involves utilizing the independent gradients of agents to derive a defined consensus gradient, enabling agents to actively perceive policy updates and optimize team collaboration. Theoretically, we leverage the Kakutani Fixed-Point Theorem to prove that the consensus direction $u^*$ guarantees the existence and attainability of this equilibrium. Extensive experiments on StarCraft II Multi-Agent Challenge (SMAC) and Google Research Football (GRF) demonstrate the scalability and promising performance of the framework.

Method: Computational model of multi-level selection where group-level selection shapes common substrate and mutation operator shared by all group members undergoing individual-level selection. Embodied ecology with coupled resource channels: positive-sum intake channel and zero-sum redistribution channel. Investigated role differentiation under turnover driven by birth and death.

Result: In learned ecology, both channels remain occupied at colony level, avoiding collapse into single acquisition mode. Zero-sum channel usage increases over generations despite not being directly optimized. Channel occupancy fluctuates over boid lifetime. Ablation shows baseline performance from inherited behavioral basis, with learned variation providing systematic improvement.

Conclusion: Multi-level selection enables groups in common-pool settings to circumvent tragedy-of-the-commons through differentiated use of coupled channels under continual turnover.

Abstract: A group of non-cooperating agents can succumb to the \emph{tragedy-of-the-commons} if all of them seek to maximize the same resource channel to improve their viability. In nature, however, groups often avoid such collapses by differentiating into distinct roles that exploit different resource channels. It remains unclear how such coordination can emerge under continual individual-level selection alone. To address this, we introduce a computational model of multi-level selection, in which group-level selection shapes a common substrate and mutation operator shared by all group members undergoing individual-level selection. We also place this process in an embodied ecology where distinct resource channels are not segregated, but coupled through the same behavioral primitives. These channels are classified as a positive-sum intake channel and a zero-sum redistribution channel. We investigate whether such a setting can give rise to role differentiation under turnover driven by birth and death. We find that in a learned ecology, both channels remain occupied at the colony level, and the collapse into a single acquisition mode is avoided. Zero-sum channel usage increases over generations despite not being directly optimized by group-level selection. Channel occupancy also fluctuates over the lifetime of a boid. Ablation studies suggest that most baseline performance is carried by the inherited behavioral basis, while the learned variation process provides a smaller but systematic improvement prior to saturation. Together, the results suggest that multi-level selection can enable groups in a common-pool setting to circumvent tragedy-of-the-commons through differentiated use of coupled channels under continual turnover.

Method: OrgAgent separates collaboration into three layers: governance layer for planning and resource allocation, execution layer for task solving and review, and compliance layer for final answer control. The framework is evaluated across reasoning tasks, LLMs, execution modes, and policies.

Result: Company-style hierarchical multi-agent systems outperform other organizational structures. For GPT-OSS-120B, hierarchical setting improves performance by 102.73% while reducing token usage by 74.52% on SQuAD 2.0. Hierarchy helps most when tasks benefit from stable skill assignment, controlled information flow, and layered verification.

Conclusion: Organizational structure is an important factor in multi-agent reasoning, shaping effectiveness, cost, and coordination behavior. Hierarchical coordination reduces token consumption relative to flat collaboration in most settings.

Abstract: While large language model-based multi-agent systems have shown strong potential for complex reasoning, how to effectively organize multiple agents remains an open question. In this paper, we introduce OrgAgent, a company-style hierarchical multi-agent framework that separates collaboration into governance, execution, and compliance layers. OrgAgent decomposes multi-agent reasoning into three layers: a governance layer for planning and resource allocation, an execution layer for task solving and review, and a compliance layer for final answer control. By evaluating the framework across reasoning tasks, LLMs, execution modes, and execution policies, we find that multi-agent systems organized in a company-style hierarchy generally outperform other organizational structures. Besides, hierarchical coordination also reduces token consumption relative to flat collaboration in most settings. For example, for GPT-OSS-120B, the hierarchical setting improves performance over flat multi-agent system by 102.73% while reducing token usage by 74.52% on SQuAD 2.0. Further analysis shows that hierarchy helps most when tasks benefit from stable skill assignment, controlled information flow, and layered verification. Overall, our findings highlight organizational structure as an important factor in multi-agent reasoning, shaping not only effectiveness and cost, but also coordination behavior.