Method: Proposes Mean Continuation Log-Probability (MCLP) metric using LALM’s in-context learning to predict continuation log-probability of ground-truth speech given generated speech. Uses MCLP as reinforcement learning reward to enhance style alignment. Constructs RP-TTS dataset with scene/character annotations.

Result: Method significantly outperforms strong LALM baselines on both objective and subjective metrics for role-play TTS tasks.

Conclusion: MCLP effectively quantifies stylistic consistency and serves as a useful reward signal for improving LALM-based role-play TTS systems.

Abstract: Recent advances in Large Audio Language Models (LALMs) have extended Text-to-Speech (TTS) to interactive role-play scenarios, which demand high expressiveness and strict adherence to role-play instructions. However, existing models struggle to maintain stylistic consistency with character profiles and scene descriptions across multi-turn dialogues. A critical bottleneck is the lack of objective metrics for quantifying speaking style. To bridge this gap, we propose Mean Continuation Log-Probability (MCLP) as both an evaluation metric and a reward signal, validated on LALM-based Role-Play TTS (RP-TTS) tasks. Critically, we leverage the In-Context Learning capability of pre-trained LALMs to formulate MCLP via a continuation log-probability prediction. This metric quantifies stylistic consistency by measuring the likelihood of the ground-truth speech conditioned on the generated speech. Furthermore, we employ MCLP as a reinforcement learning reward to enhance the style alignment between generated speech and Role-Play instructions. To facilitate evaluation, we construct an RP-TTS dataset with rich scene and character annotations. Experimental results demonstrate that our method significantly outperforms strong LALM baselines on both objective and subjective metrics.

Method: Proposes SDD-APALLM framework that combines raw audio with structured spectrograms to explicitly expose fine-grained time-frequency evidence as accessible acoustic cues, enabling audio LLMs to capture subtle acoustic inconsistencies without compromising semantic understanding.

Result: Experimental results show consistent gains in detection accuracy and robustness, especially in cases where semantic cues are misleading. Improvements stem from coordinated utilization of semantic and acoustic information rather than simple modality aggregation.

Conclusion: The acoustically enhanced framework effectively addresses the limitation of semantic-dominant reasoning in audio LLMs for speech deepfake detection by making fine-grained acoustic evidence more accessible during decision-making.

Abstract: Speech deepfake detection (SDD) focuses on identifying whether a given speech signal is genuine or has been synthetically generated. Existing audio large language model (LLM)-based methods excel in content understanding; however, their predictions are often biased toward semantically correlated cues, which results in fine-grained acoustic artifacts being overlooked during the decisionmaking process. Consequently, fake speech with natural semantics can bypass detectors despite harboring subtle acoustic anomalies; this suggests that the challenge stems not from the absence of acoustic data, but from its inadequate accessibility when semantic-dominant reasoning prevails. To address this issue, we investigate SDD within the audio LLM paradigm and introduce SDD with Auditory Perception-enhanced Audio Large Language Model (SDD-APALLM), an acoustically enhanced framework designed to explicitly expose fine-grained time-frequency evidence as accessible acoustic cues. By combining raw audio with structured spectrograms, the proposed framework empowers audio LLMs to more effectively capture subtle acoustic inconsistencies without compromising their semantic understanding. Experimental results indicate consistent gains in detection accuracy and robustness, especially in cases where semantic cues are misleading. Further analysis reveals that these improvements stem from a coordinated utilization of semantic and acoustic information, as opposed to simple modality aggregation.

Method: DIFFA-2 upgrades the speech encoder, uses dual semantic and acoustic adapters, and employs a four-stage curriculum training: semantic and acoustic alignment, large-scale supervised fine-tuning, and variance-reduced preference optimization using only open-source corpora.

Result: Experiments on MMSU, MMAU, and MMAR benchmarks show DIFFA-2 consistently improves over DIFFA and is competitive with strong autoregressive LALMs under practical training budgets, demonstrating diffusion-based modeling as a viable backbone for large-scale audio understanding.

Conclusion: Diffusion-based large audio language models are a practical alternative to autoregressive models, offering competitive performance with more efficient training and inference characteristics for general audio understanding tasks.

Abstract: Autoregressive (AR) large audio language models (LALMs) such as Qwen-2.5-Omni have achieved strong performance on audio understanding and interaction, but scaling them remains costly in data and computation, and strictly sequential decoding limits inference efficiency. Diffusion large language models (dLLMs) have recently been shown to make effective use of limited training data, and prior work on DIFFA indicates that replacing an AR backbone with a diffusion counterpart can substantially improve audio understanding under matched settings, albeit at a proof-of-concept scale without large-scale instruction tuning, preference alignment, or practical decoding schemes. We introduce DIFFA-2, a practical diffusion-based LALM for general audio understanding. DIFFA-2 upgrades the speech encoder, employs dual semantic and acoustic adapters, and is trained with a four-stage curriculum that combines semantic and acoustic alignment, large-scale supervised fine-tuning, and variance-reduced preference optimization, using only fully open-source corpora. Experiments on MMSU, MMAU, and MMAR show that DIFFA-2 consistently improves over DIFFA and is competitive to strong AR LALMs under practical training budgets, supporting diffusion-based modeling is a viable backbone for large-scale audio understanding. Our code is available at https://github.com/NKU-HLT/DIFFA.git.

Method: Three approaches: persona-based prompting (simulating drunk personas), causal fine-tuning (training on drunk language data), and reinforcement-based post-training. Evaluated on 5 LLMs using JailbreakBench and ConfAIde benchmarks with manual and LLM-based evaluation.

Result: Drunk language inducement significantly increases susceptibility to jailbreaking (even with defenses) and privacy leaks compared to base LLMs and previous approaches, showing correspondence between human-intoxicated behavior and LLM safety failures.

Conclusion: Simple drunk language inducement methods pose significant risks to LLM safety and could serve as counters to safety tuning, highlighting vulnerabilities in anthropomorphized AI systems.

Abstract: Humans are susceptible to undesirable behaviours and privacy leaks under the influence of alcohol. This paper investigates drunk language, i.e., text written under the influence of alcohol, as a driver for safety failures in large language models (LLMs). We investigate three mechanisms for inducing drunk language in LLMs: persona-based prompting, causal fine-tuning, and reinforcement-based post-training. When evaluated on 5 LLMs, we observe a higher susceptibility to jailbreaking on JailbreakBench (even in the presence of defences) and privacy leaks on ConfAIde, where both benchmarks are in English, as compared to the base LLMs as well as previously reported approaches. Via a robust combination of manual evaluation and LLM-based evaluators and analysis of error categories, our findings highlight a correspondence between human-intoxicated behaviour, and anthropomorphism in LLMs induced with drunk language. The simplicity and efficiency of our drunk language inducement approaches position them as potential counters for LLM safety tuning, highlighting significant risks to LLM safety.

Method: Proposes MrRoPE (Mixed-radix RoPE) based on radix system conversion perspective, unifying various RoPE-extension approaches as distinct radix conversion strategies. Introduces two training-free extensions: MrRoPE-Uni (uniform radix conversion) and MrRoPE-Pro (progressive radix conversion).

Result: MrRoPE-Pro sustains over 85% recall in 128K-context Needle-in-a-Haystack test, achieves more than double YaRN’s accuracy on Infinite-Bench retrieval and dialogue subsets without fine-tuning, and effectively raises the upper bound of RoPE’s attainable encoding length.

Conclusion: MrRoPE provides a unified theoretical foundation for RoPE-extension methods, enabling effective training-free long-context generalization and validating the reliability of the radix conversion perspective.

Abstract: Rotary Position Embedding (RoPE)-extension refers to modifying or generalizing the Rotary Position Embedding scheme to handle longer sequences than those encountered during pre-training. However, current extension strategies are highly diverse and lack a unified theoretical foundation. In this paper, we propose MrRoPE (Mixed-radix RoPE), a generalized encoding formulation based on a radix system conversion perspective, which elegantly unifies various RoPE-extension approaches as distinct radix conversion strategies. Based on this theory, we introduce two training-free extensions, MrRoPE-Uni and MrRoPE-Pro, which leverage uniform and progressive radix conversion strategies, respectively, to achieve ’train short, test long’ generalization. Without fine-tuning, MrRoPE-Pro sustains over 85% recall in the 128K-context Needle-in-a-Haystack test and achieves more than double YaRN’s accuracy on Infinite-Bench retrieval and dialogue subsets. Theoretical analysis confirms that MrRoPE-Pro effectively raises the upper bound of RoPE’s attainable encoding length, which further validates the reliability and utility of our theory and methodology.

Method: SDRL samples multiple candidate solutions, constructs debate contexts with diverse reasoning paths, generates second-turn responses conditioned on this context, and jointly optimizes both initial and debate-conditioned responses.

Result: Experiments across multiple base models and reasoning benchmarks show SDRL improves overall multi-agent debate performance while simultaneously strengthening single model reasoning.

Conclusion: SDRL successfully equips LLMs with both strong standalone problem-solving ability and the capability to learn from diverse reasoning trajectories in debate settings.

Abstract: The reasoning abilities of large language models (LLMs) have been substantially improved by reinforcement learning with verifiable rewards (RLVR). At test time, collaborative reasoning through Multi-Agent Debate (MAD) has emerged as a promising approach for enhancing LLM performance. However, current RLVR methods typically train LLMs to solve problems in isolation, without explicitly preparing them to synthesize and benefit from different rationales that arise during debate. In this work, we propose Self-Debate Reinforcement Learning (SDRL), a training framework that equips a single LLM with strong standalone problem-solving ability and the capability to learn from diverse reasoning trajectories in MAD. Given a prompt, SDRL first samples multiple candidate solutions, then constructs a debate context with diverse reasoning paths and generates second-turn responses conditioned on this context. Finally, SDRL jointly optimizes both the initial and debate-conditioned responses, yielding a model that is effective as both a standalone solver and a debate participant. Experiments across multiple base models and reasoning benchmarks show that SDRL improves overall MAD performance while simultaneously strengthening single model reasoning.

Method: Proposes MERMAID framework with agent-driven search, structured knowledge representations, and persistent memory module within a Reason-Action iterative process for dynamic evidence acquisition and cross-claim evidence reuse.

Result: Achieves state-of-the-art performance on three fact-checking benchmarks and two claim-verification datasets using multiple LLMs (GPT, LLaMA, Qwen families), while improving search efficiency.

Conclusion: Demonstrates effectiveness of synergizing retrieval, reasoning, and memory for reliable veracity assessment, showing that dynamic evidence management and reuse improves both accuracy and efficiency.

Abstract: Assessing the veracity of online content has become increasingly critical. Large language models (LLMs) have recently enabled substantial progress in automated veracity assessment, including automated fact-checking and claim verification systems. Typical veracity assessment pipelines break down complex claims into sub-claims, retrieve external evidence, and then apply LLM reasoning to assess veracity. However, existing methods often treat evidence retrieval as a static, isolated step and do not effectively manage or reuse retrieved evidence across claims. In this work, we propose MERMAID, a memory-enhanced multi-agent veracity assessment framework that tightly couples the retrieval and reasoning processes. MERMAID integrates agent-driven search, structured knowledge representations, and a persistent memory module within a Reason-Action style iterative process, enabling dynamic evidence acquisition and cross-claim evidence reuse. By retaining retrieved evidence in an evidence memory, the framework reduces redundant searches and improves verification efficiency and consistency. We evaluate MERMAID on three fact-checking benchmarks and two claim-verification datasets using multiple LLMs, including GPT, LLaMA, and Qwen families. Experimental results show that MERMAID achieves state-of-the-art performance while improving the search efficiency, demonstrating the effectiveness of synergizing retrieval, reasoning, and memory for reliable veracity assessment.

Method: Measured representational straightening in Gemma 2 models across diverse in-context tasks, analyzing neural trajectory straightness in relation to context length and task structure.

Result: Found a dichotomy: continual prediction tasks (natural language, grid world traversal) show increased straightening with context that correlates with improved predictions, while structured prediction tasks (few-shot learning) show inconsistent straightening - only present in phases with explicit structure.

Conclusion: ICL is not monolithic; LLMs dynamically select between strategies like a Swiss Army knife, with representational straightening occurring only for certain task types and structures.

Abstract: Large Language Models (LLMs) have been shown to organize the representations of input sequences into straighter neural trajectories in their deep layers, which has been hypothesized to facilitate next-token prediction via linear extrapolation. Language models can also adapt to diverse tasks and learn new structure in context, and recent work has shown that this in-context learning (ICL) can be reflected in representational changes. Here we bring these two lines of research together to explore whether representation straightening occurs \emph{within} a context during ICL. We measure representational straightening in Gemma 2 models across a diverse set of in-context tasks, and uncover a dichotomy in how LLMs’ representations change in context. In continual prediction settings (e.g., natural language, grid world traversal tasks) we observe that increasing context increases the straightness of neural sequence trajectories, which is correlated with improvement in model prediction. Conversely, in structured prediction settings (e.g., few-shot tasks), straightening is inconsistent – it is only present in phases of the task with explicit structure (e.g., repeating a template), but vanishes elsewhere. These results suggest that ICL is not a monolithic process. Instead, we propose that LLMs function like a Swiss Army knife: depending on task structure, the LLM dynamically selects between strategies, only some of which yield representational straightening.

Method: Across two clinical tasks (MedAlign applicability/correctness and MS subtype abstraction) with multiple open and proprietary models, they measure prompt sensitivity via flip rates and relate it to calibration and selective prediction. They propose a dual-objective prompt optimization loop targeting both accuracy and stability.

Result: Higher accuracy doesn’t guarantee prompt stability, and models can appear well-calibrated yet remain fragile to paraphrases. Explicitly including a stability term in optimization reduces flip rates across tasks and models, sometimes at modest accuracy cost.

Conclusion: Prompt sensitivity should be an explicit objective when validating clinical LLM systems to ensure robustness and reliability in healthcare applications.

Abstract: Large language models used for clinical abstraction are sensitive to prompt wording, yet most work treats prompts as fixed and studies uncertainty in isolation. We argue these should be treated jointly. Across two clinical tasks (MedAlign applicability/correctness and MS subtype abstraction) and multiple open and proprietary models, we measure prompt sensitivity via flip rates and relate it to calibration and selective prediction. We find that higher accuracy does not guarantee prompt stability, and that models can appear well-calibrated yet remain fragile to paraphrases. We propose a dual-objective prompt optimization loop that jointly targets accuracy and stability, showing that explicitly including a stability term reduces flip rates across tasks and models, sometimes at modest accuracy cost. Our results suggest prompt sensitivity should be an explicit objective when validating clinical LLM systems.

Method: SPLA uses second-order Taylor expansions to accurately select relevant blocks for exact attention, then compresses unselected blocks into a compact recurrent state via residual linear attention (RLA) with an optimized subtraction-based formulation to avoid IO overhead.

Result: SPLA surpasses dense attention models on long-context benchmarks like RULER while maintaining competitive general knowledge and reasoning capabilities, closing the performance gap in continual pretraining.

Conclusion: SPLA provides an effective framework for efficient long-context modeling that preserves contextual information without the overhead of accessing unselected blocks during inference.

Abstract: Block-wise sparse attention offers significant efficiency gains for long-context modeling, yet existing methods often suffer from low selection fidelity and cumulative contextual loss by completely discarding unselected blocks. To address these limitations, we introduce Sparse Plus Linear Attention (SPLA), a framework that utilizes a selection metric derived from second-order Taylor expansions to accurately identify relevant blocks for exact attention. Instead of discarding the remaining “long tail,” SPLA compresses unselected blocks into a compact recurrent state via a residual linear attention (RLA) module. Crucially, to avoid IO overhead, we derive an optimized subtraction-based formulation for RLA – calculating the residual as the difference between global and selected linear attention – ensuring that unselected blocks are never explicitly accessed during inference. Our experiments demonstrate that SPLA closes the performance gap in continual pretraining, surpassing dense attention models on long-context benchmarks like RULER while maintaining competitive general knowledge and reasoning capabilities.

Method: SP2DPO replaces DPO’s global temperature beta with instance-specific schedules beta_i determined offline from structured semantic-gap annotations (category, magnitude, confidence) produced by teacher language models. The method is instantiated on the UltraFeedback corpus (59,960 pairs) and maintains standard DPO training with per-pair beta values.

Result: SP2DPO is competitive with tuned global-beta DPO baselines and improves AlpacaEval 2.0 length-controlled win rate on two of four backbones (4B-8B models), while avoiding the need for per-model beta sweeps.

Conclusion: Instance-specific temperature scheduling based on semantic annotations provides a more nuanced approach to preference optimization that can handle heterogeneous preference data effectively without training-time overhead.

Abstract: Direct Preference Optimization (DPO) controls the trade-off between fitting preference labels and staying close to a reference model using a single global temperature beta, implicitly treating all preference pairs as equally informative. Real-world preference corpora are heterogeneous: they mix high-signal, objective failures (for example, safety, factuality, instruction violations) with low-signal or subjective distinctions (for example, style), and also include label noise. We introduce our method, SP2DPO (Semantic Per-Pair DPO), a generalization that replaces the global temperature with an instance-specific schedule beta_i pre-decided offline from structured semantic-gap annotations (category, magnitude, confidence) produced by teacher language models. We instantiate this procedure on the UltraFeedback preference corpus (59,960 pairs), enabling large-scale construction of an auditable beta_i artifact, and incur zero training-time overhead: the inner-loop optimizer remains standard DPO with beta set per pair. We focus our empirical study on AlpacaEval 2.0, reporting both raw win rate and length-controlled win rate. Across four open-weight, instruction-tuned student backbones (4B-8B), SP2DPO is competitive with a tuned global-beta DPO baseline and improves AlpacaEval 2.0 length-controlled win rate on two of four backbones, while avoiding per-model beta sweeps. All code, annotations, and artifacts will be released.

Method: Evaluated single-agent and multi-agent LLM architectures using GPT-5.1 on ASAP 2.0 corpus. Multi-agent system had three specialist agents (Content, Structure, Language) coordinated by a Chairman Agent with rubric-aligned logic including veto rules and score capping. Tested both zero-shot and few-shot conditions.

Result: Multi-agent system significantly better at identifying weak essays, single-agent better on mid-range essays. Both struggle with high-quality essays. Few-shot calibration (just two examples per score level) improves QWK by ~26% for both architectures.

Conclusion: Architectural choice should align with deployment priorities: multi-agent AI suited for diagnostic screening of at-risk students, single-agent models provide cost-effective general assessment. Few-shot calibration is critical for performance.

Abstract: Automated essay scoring (AES) systems increasingly rely on large language models, yet little is known about how architectural choices shape their performance across different essay quality levels. This paper evaluates single-agent and multi-agent LLM architectures for essay grading using the ASAP 2.0 corpus. Our multi-agent system decomposes grading into three specialist agents (Content, Structure, Language) coordinated by a Chairman Agent that implements rubric-aligned logic including veto rules and score capping. We test both architectures in zero-shot and few-shot conditions using GPT-5.1. Results show that the multi-agent system is significantly better at identifying weak essays while the single-agent system performs better on mid-range essays. Both architectures struggle with high-quality essays. Critically, few-shot calibration emerges as the dominant factor in system performance – providing just two examples per score level improves QWK by approximately 26% for both architectures. These findings suggest architectural choice should align with specific deployment priorities, with multi-agent AI particularly suited for diagnostic screening of at-risk students, while single-agent models provide a cost-effective solution for general assessment.

Method: Generate LLM personas based on interpretable WVS-derived variables, then evaluate through three lenses: 1) positioning on Inglehart-Welzel Cultural Map to assess cultural conditioning interpretation, 2) demographic-level consistency with World Values Survey response distributions, and 3) moral profiles from Moral Foundations questionnaire analyzed through culture-to-morality mapping.

Result: The approach enables evaluation of cross-cultural structure and moral variation in synthetic personas, revealing how well LLM-generated personas track human group patterns and cultural configurations.

Conclusion: Culturally-grounded persona generation and analysis provides a framework for assessing how accurately LLMs simulate cross-cultural human behavior and moral value systems.

Abstract: Despite the growing utility of Large Language Models (LLMs) for simulating human behavior, the extent to which these synthetic personas accurately reflect world and moral value systems across different cultural conditionings remains uncertain. This paper investigates the alignment of synthetic, culturally-grounded personas with established frameworks, specifically the World Values Survey (WVS), the Inglehart-Welzel Cultural Map, and Moral Foundations Theory. We conceptualize and produce LLM-generated personas based on a set of interpretable WVS-derived variables, and we examine the generated personas through three complementary lenses: positioning on the Inglehart-Welzel map, which unveils their interpretation reflecting stable differences across cultural conditionings; demographic-level consistency with the World Values Survey, where response distributions broadly track human group patterns; and moral profiles derived from a Moral Foundations questionnaire, which we analyze through a culture-to-morality mapping to characterize how moral responses vary across different cultural configurations. Our approach of culturally-grounded persona generation and analysis enables evaluation of cross-cultural structure and moral variation.

Method: Introduces Bifocal Attention with two positional encoding modalities: Geometric Eyes (standard RoPE) for token-level manipulation and Spectral Eyes (learnable harmonic operators) for tracking long-range recursive depth. Uses Spectral Evolution training protocol that initializes frequencies as static geometric parameters but allows gradient-based evolution into task-optimized harmonic basis.

Result: The approach aims to bridge the Structure Gap by enabling models to better handle algorithmic reasoning tasks requiring deep recursive steps, though specific experimental results are not provided in the abstract.

Conclusion: Bifocal Attention with Spectral Evolution provides a solution to RoPE’s spectral rigidity limitation, potentially improving LLMs’ ability to handle complex recursive and algorithmic reasoning tasks.

Abstract: Rotary Positional Embeddings (RoPE) have become the standard for Large Language Models (LLMs) due to their ability to encode relative positions through geometric rotation. However, we identify a significant limitation we term ‘‘Spectral Rigidity’’: standard RoPE utilizes a fixed geometric decay ($θ^{-i}$) optimized for local syntactic coherence, which fails to capture the long-range, periodic structures inherent in recursive logic and algorithmic reasoning. This results in a ‘‘Structure Gap’’, where models trained on shallow reasoning chains fail to extrapolate to deeper recursive steps. In this work, we introduce Bifocal Attention, an architectural paradigm that decouples positional encoding into two distinct modalities: Geometric Eyes (Standard RoPE) for precise token-level manipulation, and Spectral Eyes (Learnable Harmonic Operators) for tracking long-range recursive depth. We propose a novel training protocol, Spectral Evolution, which initializes positional frequencies as static geometric parameters but allows them to evolve via gradient descent into a harmonic basis optimized for the specific algorithmic topology of the task.

Method: Introduces a diffusion-based speech-text language model supporting both understanding and generation, unifying discrete text and tokenized speech under a single masked diffusion framework. Uses modality-specific masking schedules and iterative denoising to jointly generate reasoning traces and speech tokens.

Result: Achieves state-of-the-art speech-to-speech QA accuracy, outperforming best baseline by up to 9 points. Attains best TTS quality among generative models (6.2% WER) while preserving language understanding (66.2% MMLU). Also introduces a new speech QA dataset with paired text reasoning traces.

Conclusion: The proposed paradigm of “Silent Thought, Spoken Answer” with diffusion-based joint generation of reasoning traces and speech tokens significantly improves speech language model performance and enables error correction before audio production.

Abstract: Current speech language models generate responses directly without explicit reasoning, leading to errors that cannot be corrected once audio is produced. We introduce \textbf{``Silent Thought, Spoken Answer’’} – a paradigm where speech LLMs generate internal text reasoning alongside spoken responses, with thinking traces informing speech quality. To realize this, we present \method{}, the first diffusion-based speech-text language model supporting both understanding and generation, unifying discrete text and tokenized speech under a single masked diffusion framework. Unlike autoregressive approaches, \method{} jointly generates reasoning traces and speech tokens through iterative denoising, with modality-specific masking schedules. We also construct \dataset{}, the first speech QA dataset with paired text reasoning traces, containing 26K samples totaling 319 hours. Experiments show \method{} achieves state-of-the-art speech-to-speech QA accuracy, outperforming the best baseline by up to 9 points, while attaining the best TTS quality among generative models (6.2% WER) and preserving language understanding (66.2% MMLU). Ablations confirm that both the diffusion architecture and thinking traces contribute to these gains.

Method: For each target word and time slice, induce word-centered semantic networks integrating distributional similarity from diachronic Skip-gram embeddings with lexical substitutability from time-specific masked language models. Identify sense-related structure by clustering the peripheral graph, align clusters across time via node overlap, and track change through cluster composition and normalized cluster mass.

Result: Applied to New York Times Magazine articles (1980-2017), graph connectivity reflects polysemy dynamics, and induced communities capture contrasting trajectories: event-driven sense replacement (trump), semantic stability with cluster over-segmentation effects (god), and gradual association shifts tied to digital communication (post).

Conclusion: Word-centered semantic graphs offer a compact and transparent representation for exploring sense evolution without relying on predefined sense inventories, providing interpretable insights into semantic change dynamics.

Abstract: We propose an interpretable, graph-based framework for analyzing semantic shift in diachronic corpora. For each target word and time slice, we induce a word-centered semantic network that integrates distributional similarity from diachronic Skip-gram embeddings with lexical substitutability from time-specific masked language models. We identify sense-related structure by clustering the peripheral graph, align clusters across time via node overlap, and track change through cluster composition and normalized cluster mass. In an application study on a corpus of New York Times Magazine articles (1980 - 2017), we show that graph connectivity reflects polysemy dynamics and that the induced communities capture contrasting trajectories: event-driven sense replacement (trump), semantic stability with cluster over-segmentation effects (god), and gradual association shifts tied to digital communication (post). Overall, word-centered semantic graphs offer a compact and transparent representation for exploring sense evolution without relying on predefined sense inventories.

Method: Used controlled causal interventions on both raw and condensed experience forms, evaluating four representative frameworks across 10 LLM backbones and 9 environments, analyzing single- and multi-agent configurations.

Result: Found striking asymmetry: agents consistently depend on raw experience but often disregard or misinterpret condensed experience, even when it’s the only experience provided. This persists across configurations and backbone scales.

Conclusion: Challenges prevailing assumptions about self-evolving methods, highlighting need for more faithful experience integration approaches. Identified three causes: semantic limitations of condensed content, internal processing biases suppressing experience, and task regimes where pretrained priors suffice.

Abstract: Self-evolving large language model (LLM) agents continually improve by accumulating and reusing past experience, yet it remains unclear whether they faithfully rely on that experience to guide their behavior. We present the first systematic investigation of experience faithfulness, the causal dependence of an agent’s decisions on the experience it is given, in self-evolving LLM agents. Using controlled causal interventions on both raw and condensed forms of experience, we comprehensively evaluate four representative frameworks across 10 LLM backbones and 9 environments. Our analysis uncovers a striking asymmetry: while agents consistently depend on raw experience, they often disregard or misinterpret condensed experience, even when it is the only experience provided. This gap persists across single- and multi-agent configurations and across backbone scales. We trace its underlying causes to three factors: the semantic limitations of condensed content, internal processing biases that suppress experience, and task regimes where pretrained priors already suffice. These findings challenge prevailing assumptions about self-evolving methods and underscore the need for more faithful and reliable approaches to experience integration.

Method: Thresholding technique that reduces the impact of marginalization on rare tokens, allowing them to benefit from more meaningful alignment. Applies negative sampling principles to limit harmful influence of excessive marginalization.

Result: Significantly improves performance on low-resource language validation data using character-level language models. First demonstration of negative sampling application to improve rare token representation.

Conclusion: Proposed method offers new approach to enhancing language model performance for underrepresented languages by addressing marginalization of rare tokens during training.

Abstract: Neural language models often struggle with low-resource languages due to the limited availability of training data, making tokens from these languages rare in the training set. This paper addresses a specific challenge during training: rare tokens are disproportionately affected by marginalization, which prevents them from learning effectively. We propose a thresholding technique that reduces the impact of this marginalization, allowing rare tokens to benefit from more meaningful alignment. Through experiments with a character-level language model, we demonstrate that this method significantly improves performance on low-resource language validation data. This work is the first to show how negative sampling can be applied to improve the representation of rare tokens by limiting the harmful influence of excessive marginalization, offering a new approach to enhancing language model performance for underrepresented languages.

Method: SSL introduces tiered, progressively amplified rewards that guide policies toward “sweet spot” regions of solution space. For visual perception tasks, it uses distance-tiered modeling to reward proximity; for complex reasoning tasks, it rewards incremental progress toward promising solutions.

Result: Extensive experiments across GUI perception, short/long-term planning, and complex reasoning tasks show consistent improvements over strong baselines on 12 benchmarks, achieving up to 2.5X sample efficiency gains and effective cross-task transferability.

Conclusion: SSL establishes a general principle for training capable and robust agents by providing differentiated guidance through tiered rewards, preserving optimal solution ordering and enhancing gradient signal-to-noise ratio for more directed optimization.

Abstract: Reinforcement learning with verifiable rewards has emerged as a powerful paradigm for training intelligent agents. However, existing methods typically employ binary rewards that fail to capture quality differences among trajectories achieving identical outcomes, thereby overlooking potential diversity within the solution space. Inspired by the ``sweet spot’’ concept in tennis-the racket’s core region that produces optimal hitting effects, we introduce \textbf{S}weet \textbf{S}pot \textbf{L}earning (\textbf{SSL}), a novel framework that provides differentiated guidance for agent optimization. SSL follows a simple yet effective principle: progressively amplified, tiered rewards guide policies toward the sweet-spot region of the solution space. This principle naturally adapts across diverse tasks: visual perception tasks leverage distance-tiered modeling to reward proximity, while complex reasoning tasks reward incremental progress toward promising solutions. We theoretically demonstrate that SSL preserves optimal solution ordering and enhances the gradient signal-to-noise ratio, thereby fostering more directed optimization. Extensive experiments across GUI perception, short/long-term planning, and complex reasoning tasks show consistent improvements over strong baselines on 12 benchmarks, achieving up to 2.5X sample efficiency gains and effective cross-task transferability. Our work establishes SSL as a general principle for training capable and robust agents.

Method: Uses strong teacher model to create novel tasks and tool ecosystems, then rewrites them into underspecified instructions to force agents to query users. Includes LLM-based user simulator for private info and mock tool system for stable responses. Task-level rubrics based on subgoals, interactions, and forbidden behaviors.

Result: Models trained on synthetic data achieve substantial gains across 14 challenging datasets in math, search, and tool use, with small models outperforming larger baselines.

Conclusion: SYNTHAGENT effectively addresses bottlenecks in agent training by synthesizing diverse tool-use data and simulating complete environments, enabling small LLMs to achieve better agentic capabilities.

Abstract: Small LLMs often struggle to match the agentic capabilities of large, costly models. While reinforcement learning can help, progress has been limited by two structural bottlenecks: existing open-source agentic training data are narrow in task variety and easily solved; real-world APIs lack diversity and are unstable for large-scale reinforcement learning rollout processes. We address these challenges with SYNTHAGENT, a framework that jointly synthesizes diverse tool-use training data and simulates complete environments. Specifically, a strong teacher model creates novel tasks and tool ecosystems, then rewrites them into intentionally underspecified instructions. This compels agents to actively query users for missing details. When handling synthetic tasks, an LLM-based user simulator provides user-private information, while a mock tool system delivers stable tool responses. For rewards, task-level rubrics are constructed based on required subgoals, user-agent interactions, and forbidden behaviors. Across 14 challenging datasets in math, search, and tool use, models trained on our synthetic data achieve substantial gains, with small models outperforming larger baselines.

Method: Proposes Power-Mean Policy Optimization (PMPO) with parameter p controlling aggregation geometry. Introduces Clip-aware Effective Sample Size (ESS) mechanism that maps trajectory clipping fraction to target ESS, then solves for optimal p to align trajectory-induced ESS with target, enabling dynamic transition between aggressive arithmetic mean (p=1) for reliable trajectories and conservative geometric mean (p→0) for unstable ones.

Result: Experiments on multiple mathematical reasoning benchmarks show PMPO outperforms strong baselines, demonstrating the effectiveness of adaptive geometry selection based on trajectory reliability.

Conclusion: PMPO provides a unified framework that generalizes existing group-based RL methods and enables adaptive aggregation geometry selection, improving performance by dynamically adjusting to trajectory characteristics through the clip-aware ESS mechanism.

Abstract: Group-based reinforcement learning has evolved from the arithmetic mean of GRPO to the geometric mean of GMPO. While GMPO improves stability by constraining a conservative objective, it shares a fundamental limitation with GRPO: reliance on a fixed aggregation geometry that ignores the evolving and heterogeneous nature of each trajectory. In this work, we unify these approaches under Power-Mean Policy Optimization (PMPO), a generalized framework that parameterizes the aggregation geometry via the power-mean geometry exponent p. Within this framework, GRPO and GMPO are recovered as special cases. Theoretically, we demonstrate that adjusting p modulates the concentration of gradient updates, effectively reweighting tokens based on their advantage contribution. To determine p adaptively, we introduce a Clip-aware Effective Sample Size (ESS) mechanism. Specifically, we propose a deterministic rule that maps a trajectory clipping fraction to a target ESS. Then, we solve for the specific p to align the trajectory induced ESS with this target one. This allows PMPO to dynamically transition between the aggressive arithmetic mean for reliable trajectories and the conservative geometric mean for unstable ones. Experiments on multiple mathematical reasoning benchmarks demonstrate that PMPO outperforms strong baselines.

Method: Proposes ρ-EOS, a training-free single-stage strategy that uses implicit EOS token density during denoising to guide bidirectional length adjustment. High EOS density triggers MASK token contraction, while insufficient density induces expansion.

Result: Extensive experiments on mathematics and code benchmarks show ρ-EOS achieves comparable performance while substantially improving inference efficiency and token utilization compared to prior two-stage approaches.

Conclusion: ρ-EOS enables flexible bidirectional variable-length generation for masked diffusion LLMs without retraining, addressing fundamental limitations of fixed-length generation while maintaining performance and improving efficiency.

Abstract: Beyond parallel generation and global context modeling, current masked diffusion large language models (dLLMs) suffer from a fundamental limitation: they require a predefined, fixed generation length, which lacks flexibility and forces an inevitable trade-off between output quality and computational efficiency. To address this, we study the denoising dynamics and find that the implicit density ($ρ$) of end-of-sequence ($\texttt{EOS}$) tokens serves as a reliable signal of generation sufficiency. In particular, the evolving implicit $\texttt{EOS}$ density during denoising reveals whether the current masked space is excessive or insufficient, thereby guiding the adjustment direction for generation length. Building on this insight, we propose $\textbf{$ρ$-$\texttt{EOS}$}$, a training-free, single-stage strategy that enables bidirectional variable-length generation for masked dLLMs. Unlike prior two-stage approaches–which require separate length adjustment and iterative mask insertion phases while supporting only unidirectional expansion–$\textbf{$ρ$-$\texttt{EOS}$}$ achieves bidirectional length adjustment within a unified denoising process by continuously estimating the implicit $\texttt{EOS}$ density: excessively high density triggers $\texttt{MASK}$ token contraction, while insufficient density induces expansion. Extensive experiments on mathematics and code benchmarks demonstrate that $\textbf{$ρ$-$\texttt{EOS}$}$ achieves comparable performance while substantially improving inference efficiency and token utilization.

Method: Introduces ATOD benchmark with synthetic dialogue generation pipeline producing richly annotated conversations requiring long-term reasoning. Proposes ATOD-Eval framework translating agentic dimensions into fine-grained metrics for reproducible offline/online evaluation. Also presents a memory-based evaluator for benchmarking.

Result: ATOD-Eval enables comprehensive assessment across task completion, agentic capability, and response quality. The proposed memory-based evaluator offers better accuracy-efficiency tradeoff compared to existing memory- and LLM-based approaches in this evaluation setting.

Conclusion: ATOD addresses the gap in evaluating advanced agentic behaviors in task-oriented dialogue systems, providing a systematic benchmark and evaluation framework for modern conversational agents with long-term reasoning and proactive capabilities.

Abstract: Recent advances in task-oriented dialogue (TOD) systems, driven by large language models (LLMs) with extensive API and tool integration, have enabled conversational agents to coordinate interleaved goals, maintain long-horizon context, and act proactively through asynchronous execution. These capabilities extend beyond traditional TOD systems, yet existing benchmarks lack systematic support for evaluating such agentic behaviors. To address this gap, we introduce ATOD, a benchmark and synthetic dialogue generation pipeline that produces richly annotated conversations requiring long-term reasoning. ATOD captures key characteristics of advanced TOD, including multi-goal coordination, dependency management, memory, adaptability, and proactivity. Building on ATOD, we propose ATOD-Eval, a holistic evaluation framework that translates these dimensions into fine-grained metrics and supports reproducible offline and online evaluation. We further present a strong agentic memory-based evaluator for benchmarking on ATOD. Experiments show that ATOD-Eval enables comprehensive assessment across task completion, agentic capability, and response quality, and that the proposed evaluator offers a better accuracy-efficiency tradeoff compared to existing memory- and LLM-based approaches under this evaluation setting.

Method: Proposes HOLO plugin that: 1) Identifies the “Holographic Characteristic” - LLMs capture target-side keywords early in generation, 2) Extracts these keywords within limited generation steps, 3) Uses parallel lexically constrained text generation to complete sentences.

Result: Experiments on various LLM architectures and scales in short-text generation show HOLO achieves comparable performance to baselines on both automatic and human evaluation metrics while demonstrating the potential of the Holographic Characteristic.

Conclusion: The Holographic Characteristic is a valuable property of LLMs that can be leveraged to improve inference efficiency. HOLO demonstrates practical application of this insight for efficient text generation.

Abstract: The recent advancements in Large Language Models (LLMs) have attracted interest in exploring their in-context learning abilities and chain-of-thought capabilities. However, there are few studies investigating the specific traits related to the powerful generation capacity of LLMs. This paper aims to delve into the generation characteristics exhibited by LLMs. Through our investigation, we have discovered that language models tend to capture target-side keywords at the beginning of the generation process. We name this phenomenon the Holographic Characteristic of language models. For the purpose of exploring this characteristic and further improving the inference efficiency of language models, we propose a plugin called HOLO, which leverages the Holographic Characteristic to extract target-side keywords from language models within a limited number of generation steps and complements the sentence with a parallel lexically constrained text generation method. To verify the effectiveness of HOLO, we conduct massive experiments on language models of varying architectures and scales in the short-text generation scenario. The results demonstrate that HOLO achieves comparable performance to the baselines in terms of both automatic and human-like evaluation metrics and highlight the potential of the Holographic Characteristic.

Method: Introduces an Evaluator Quality Baseline to decouple self-preference from noisy outputs on hard problems. Compares probability that a judge incorrectly votes for itself against probability it votes for an incorrect response from another model. Evaluates on 37,448 queries to isolate true self-preference signals.

Result: Methodological confound could reduce measurement error by 89.6%. Only 51% of initial findings retain statistical significance after applying the corrective baseline. Enables characterization of entropy differences between “easy” versus “hard” evaluation votes from LLM judges.

Conclusion: The proposed baseline eliminates noisy data from potential solutions, enabling more accurate research on self-preference bias. Contributes to cataloging and isolating judge-bias effects in LLM evaluation workflows.

Abstract: Recent research has shown that large language models (LLM) favor own outputs when acting as judges, undermining the integrity of automated post-training and evaluation workflows. However, it is difficult to disentangle which evaluation biases are explained by narcissism versus general experimental confounds, distorting measurements of self-preference bias. We discover a core methodological confound which could reduce measurement error by 89.6%. Specifically, LLM evaluators may deliver self-preferring verdicts when the judge responds to queries which they completed incorrectly themselves; this would be true regardless of whether one of their responses is their own. To decouple self-preference signals from noisy outputs on hard problems, we introduce an Evaluator Quality Baseline, which compares the probability that a judge incorrectly votes for itself against the probability that it votes for an incorrect response from another model. Evaluating this simple baseline on 37,448 queries, only 51% of initial findings retain statistical significance. Finally, we turn towards characterizing the entropy of “easy” versus “hard” evaluation votes from LLM judges. Our corrective baseline enables future research on self-preference by eliminating noisy data from potential solutions. More widely, this work contributes to the growing body of work on cataloging and isolating judge-bias effects.

Method: Proposes OG-MAR framework that: 1) summarizes respondent-specific values from World Values Survey, 2) constructs global cultural ontology via competency questions, 3) retrieves ontology-consistent relations and demographically similar profiles to instantiate multiple value-persona agents, and 4) synthesizes outputs using a judgment agent that enforces ontology consistency and demographic proximity.

Result: Experiments on regional social-survey benchmarks across four LLM backbones show OG-MAR improves cultural alignment and robustness over competitive baselines while producing more transparent reasoning traces.

Conclusion: OG-MAR provides an effective framework for improving cultural alignment in LLMs through structured ontology-guided reasoning and multi-agent synthesis, addressing limitations of existing methods that treat values as independent, unstructured signals.

Abstract: Large Language Models (LLMs) increasingly support culturally sensitive decision making, yet often exhibit misalignment due to skewed pretraining data and the absence of structured value representations. Existing methods can steer outputs, but often lack demographic grounding and treat values as independent, unstructured signals, reducing consistency and interpretability. We propose OG-MAR, an Ontology-Guided Multi-Agent Reasoning framework. OG-MAR summarizes respondent-specific values from the World Values Survey (WVS) and constructs a global cultural ontology by eliciting relations over a fixed taxonomy via competency questions. At inference time, it retrieves ontology-consistent relations and demographically similar profiles to instantiate multiple value-persona agents, whose outputs are synthesized by a judgment agent that enforces ontology consistency and demographic proximity. Experiments on regional social-survey benchmarks across four LLM backbones show that OG-MAR improves cultural alignment and robustness over competitive baselines, while producing more transparent reasoning traces.

Method: Leverages large-scale speech training data and the audio understanding capabilities of foundation model Qwen3-Omni. The ASR models support 52 languages/dialects with language identification. The forced alignment model uses LLM-based non-autoregressive architecture for text-speech alignment in 11 languages.

Result: The 1.7B model achieves SOTA performance among open-source ASR models and competes with proprietary APIs. The 0.6B model offers best accuracy-efficiency trade-off with 92ms average TTFT and can transcribe 2000 seconds of speech in 1 second at 128 concurrency. The forced alignment model outperforms three strongest competitors in timestamp accuracy with better efficiency and versatility.

Conclusion: The Qwen3-ASR family provides state-of-the-art, efficient speech recognition and alignment capabilities, released under Apache 2.0 license to accelerate community research in ASR and audio understanding.

Abstract: In this report, we introduce Qwen3-ASR family, which includes two powerful all-in-one speech recognition models and a novel non-autoregressive speech forced alignment model. Qwen3-ASR-1.7B and Qwen3-ASR-0.6B are ASR models that support language identification and ASR for 52 languages and dialects. Both of them leverage large-scale speech training data and the strong audio understanding ability of their foundation model Qwen3-Omni. We conduct comprehensive internal evaluation besides the open-sourced benchmarks as ASR models might differ little on open-sourced benchmark scores but exhibit significant quality differences in real-world scenarios. The experiments reveal that the 1.7B version achieves SOTA performance among open-sourced ASR models and is competitive with the strongest proprietary APIs while the 0.6B version offers the best accuracy-efficiency trade-off. Qwen3-ASR-0.6B can achieve an average TTFT as low as 92ms and transcribe 2000 seconds speech in 1 second at a concurrency of 128. Qwen3-ForcedAligner-0.6B is an LLM based NAR timestamp predictor that is able to align text-speech pairs in 11 languages. Timestamp accuracy experiments show that the proposed model outperforms the three strongest force alignment models and takes more advantages in efficiency and versatility. To further accelerate the community research of ASR and audio understanding, we release these models under the Apache 2.0 license.

Method: Proposes SpanNorm which establishes clean residual connections spanning entire transformer blocks to stabilize signal propagation, while using PostNorm-style computation to normalize aggregated output for enhanced performance. Includes principled scaling strategy to maintain bounded signal variance.

Result: Theoretical analysis shows SpanNorm maintains bounded signal variance throughout network, preventing gradient issues of PostNorm and alleviating representation collapse of PreNorm. Empirically outperforms standard normalization schemes in both dense and Mixture-of-Experts scenarios.

Conclusion: SpanNorm resolves the PreNorm/PostNorm dilemma, paving the way for more powerful and stable Transformer architectures by integrating strengths of both paradigms.

Abstract: The success of Large Language Models (LLMs) hinges on the stable training of deep Transformer architectures. A critical design choice is the placement of normalization layers, leading to a fundamental trade-off: the PreNorm'' architecture ensures training stability at the cost of potential performance degradation in deep models, while the PostNorm’’ architecture offers strong performance but suffers from severe training instability. In this work, we propose SpanNorm, a novel technique designed to resolve this dilemma by integrating the strengths of both paradigms. Structurally, SpanNorm establishes a clean residual connection that spans the entire transformer block to stabilize signal propagation, while employing a PostNorm-style computation that normalizes the aggregated output to enhance model performance. We provide a theoretical analysis demonstrating that SpanNorm, combined with a principled scaling strategy, maintains bounded signal variance throughout the network, preventing the gradient issues that plague PostNorm models, and also alleviating the representation collapse of PreNorm. Empirically, SpanNorm consistently outperforms standard normalization schemes in both dense and Mixture-of-Experts (MoE) scenarios, paving the way for more powerful and stable Transformer architectures.

Method: Proposes the Semantic Capacity Asymmetry Hypothesis: evaluation requires less semantic capacity than generation. Introduces INSPECTOR, a probing-based framework that predicts aspect-level evaluation scores from small model representations, shifting from LLM-as-a-Judge to Representation-as-a-Judge.

Result: Experiments on reasoning benchmarks (GSM8K, MATH, GPQA) show INSPECTOR substantially outperforms prompting-based small LMs and closely approximates full LLM judges, while being more efficient, reliable, and interpretable.

Conclusion: Small models encode rich evaluative signals in hidden states, enabling efficient evaluation without surface generation. This motivates a paradigm shift from LLM-as-a-Judge to Representation-as-a-Judge for scalable evaluation.

Abstract: Large language models (LLMs) are widely used as reference-free evaluators via prompting, but this “LLM-as-a-Judge” paradigm is costly, opaque, and sensitive to prompt design. In this work, we investigate whether smaller models can serve as efficient evaluators by leveraging internal representations instead of surface generation. We uncover a consistent empirical pattern: small LMs, despite with weak generative ability, encode rich evaluative signals in their hidden states. This motivates us to propose the Semantic Capacity Asymmetry Hypothesis: evaluation requires significantly less semantic capacity than generation and can be grounded in intermediate representations, suggesting that evaluation does not necessarily need to rely on large-scale generative models but can instead leverage latent features from smaller ones. Our findings motivate a paradigm shift from LLM-as-a-Judge to Representation-as-a-Judge, a decoding-free evaluation strategy that probes internal model structure rather than relying on prompted output. We instantiate this paradigm through INSPECTOR, a probing-based framework that predicts aspect-level evaluation scores from small model representations. Experiments on reasoning benchmarks (GSM8K, MATH, GPQA) show that INSPECTOR substantially outperforms prompting-based small LMs and closely approximates full LLM judges, while offering a more efficient, reliable, and interpretable alternative for scalable evaluation.

Method: The authors empirically demonstrate that MLP neurons are as sparse a feature basis as SAEs. They develop an end-to-end pipeline for circuit tracing on the MLP neuron basis using gradient-based attribution. The approach is tested on two benchmarks: a subject-verb agreement task and a multi-hop city→state→capital reasoning task.

Result: On the subject-verb agreement benchmark, a circuit of approximately 100 MLP neurons was sufficient to control model behavior. On the multi-hop reasoning task, they found circuits where small sets of neurons encode specific latent reasoning steps (e.g., “map city to its state”) that can be steered to change model outputs.

Conclusion: MLP neurons provide a sufficiently sparse and interpretable feature basis for circuit tracing, advancing automated interpretability of language models without the computational overhead of training additional components like SAEs.

Abstract: The high-level concepts that a neural network uses to perform computation need not be aligned to individual neurons (Smolensky, 1986). Language model interpretability research has thus turned to techniques such as \textit{sparse autoencoders} (SAEs) to decompose the neuron basis into more interpretable units of model computation, for tasks such as \textit{circuit tracing}. However, not all neuron-based representations are uninterpretable. For the first time, we empirically show that \textbf{MLP neurons are as sparse a feature basis as SAEs}. We use this finding to develop an end-to-end pipeline for circuit tracing on the MLP neuron basis, which locates causal circuitry on a variety of tasks using gradient-based attribution. On a standard subject-verb agreement benchmark (Marks et al., 2025), a circuit of $\approx 10^2$ MLP neurons is enough to control model behaviour. On the multi-hop city $\to$ state $\to$ capital task from Lindsey et al., 2025, we find a circuit in which small sets of neurons encode specific latent reasoning steps (e.g.~`map city to its state’), and can be steered to change the model’s output. This work thus advances automated interpretability of language models without additional training costs.

Method: Proposes a safety-aware layer swapping method that transfers safety alignment from an English safety expert to low-resource language experts without additional training. The method adaptively selects or blends modules based on their degree of specialization to enhance transfer ability.

Result: The method achieves comparable performance to language experts on general benchmarks (MMLU, BELEBELE, MGSM) while producing more aligned and less harmful responses on the MultiJail safety benchmark.

Conclusion: The proposed approach effectively enhances safety in low-resource languages while preserving general language understanding capabilities, addressing the multilingual safety alignment gap.

Abstract: Despite the rapid advancements of Large Language Models (LLMs), safety risks remain a critical challenge for low-resource languages. Existing safety datasets are predominantly English centric, limiting progress in multilingual safety alignment. As a result, low resource expert models, finetuned on their respective instruction datasets, tend to exhibit higher unsafety rates compared to their high resource counterparts. In this work, we propose a safety aware layer swapping method that transfers safety alignment from an English safety expert to low resource language experts without additional training. To further enhance transfer ability, our method adaptively selects or blends modules based on their degree of specialization. Our approach preserves performance on general language understanding tasks while enhancing safety in the target languages. Experimental results show that the proposed method achieves comparable performance to the language expert on general benchmarks such as MMMLU, BELEBELE, and MGSM, while producing more aligned and less harmful responses on the MultiJail safety benchmark.

Method: Time-Annealed Perturbation Sampling (TAPS) builds on the insight that Diffusion-LMs exhibit temporal division of labor: early steps determine global semantics, later steps focus on lexical refinement. TAPS encourages semantic branching early by adding perturbations, then progressively reduces perturbations to preserve fluency and instruction adherence.

Result: TAPS consistently improves output diversity across creative writing and reasoning benchmarks without compromising generation quality. It’s compatible with both non-autoregressive and semi-autoregressive Diffusion backbones (LLaDA and TraDo).

Conclusion: The temporal structure of Diffusion-LMs can be effectively leveraged to control generation diversity, and TAPS provides a training-free method to explore multiple valid semantic or reasoning paths while maintaining output quality.

Abstract: Diffusion language models (Diffusion-LMs) introduce an explicit temporal dimension into text generation, yet how this structure can be leveraged to control generation diversity for exploring multiple valid semantic or reasoning paths remains underexplored. In this paper, we show that Diffusion-LMs, like diffusion models in image generation, exhibit a temporal division of labor: early denoising steps largely determine the global semantic structure, while later steps focus on local lexical refinement. Building on this insight, we propose Time-Annealed Perturbation Sampling (TAPS), a training-free inference strategy that encourages semantic branching early in the diffusion process while progressively reducing perturbations to preserve fluency and instruction adherence. TAPS is compatible with both non-autoregressive and semi-autoregressive Diffusion backbones, demonstrated on LLaDA and TraDo in our paper, and consistently improves output diversity across creative writing and reasoning benchmarks without compromising generation quality.

Method: DART monitors shifts in attention score distributions to infer context changes, then dynamically updates neuron-level masks to retain salient parameters. It’s training-free and performs on-the-fly context-based pruning with minimal memory overhead.

Result: Outperforms prior dynamic baselines with up to 14.5% accuracy gain on LLAMA-3.1-8B at 70% FFN sparsity, achieves up to 3x better ROUGE-L scores on summarization tasks, and runs with <10MB memory overhead (0.1% FLOPs overhead).

Conclusion: DART effectively adapts to diverse semantic contexts, preserves model capabilities across general and domain-specific tasks, and provides efficient dynamic pruning with minimal computational overhead.

Abstract: Large Language Models (LLMs) exhibit substantial parameter redundancy, particularly in Feed-Forward Networks (FFNs). Existing pruning methods suffer from two primary limitations. First, reliance on dataset-specific calibration introduces significant data dependency and computational overhead. Second, being predominantly static, they fail to account for the evolving subset of knowledge neurons in LLMs during autoregressive generation as the context evolves. To address this, we introduce DART, i.e., Dynamic Attention-Guided Runtime Tracing), a lightweight, training-free method that performs on-the-fly context-based pruning. DART monitors shifts in attention score distributions to infer context changes, dynamically updating neuron-level masks to retain salient parameters. Across ten benchmarks, DART outperforms prior dynamic baseline, achieving accuracy gains of up to 14.5% on LLAMA-3.1-8B at 70% FFN sparsity. Furthermore, DART achieves up to 3x better ROUGE-L scores with respect to static-masked pruning on summarization tasks, with its performance comparable to the original dense models. We conclusively demonstrate that the proposed framework effectively adapts to diverse semantic contexts, preserves model capabilities across both general and domain-specific tasks while running at less than 10MBs of memory for LLAMA-3.1-8B(16GBs) with 0.1% FLOPs overhead. The code is available at https://github.com/seeder-research/DART.

Method: NAG internalizes graph processing within the LM’s native manifold by repurposing self-attention to enforce topological dependencies and recalibrating positional IDs to ensure structural equivalence. Two implementations: NAG-Zero for absolute preservation of base model’s linguistic capabilities, and NAG-LoRA for enhanced structural adaptation.

Result: Experiments across diverse graph tasks validate that NAG achieves robust graph comprehension without the overhead of external encoders, offering a simpler, more coherent paradigm for text-graph modeling.

Conclusion: NAG provides a unified framework that eliminates the need for external graph encoders by leveraging the LM’s native architecture, creating a more coherent and efficient approach to text-graph modeling.

Abstract: Prevailing methods for integrating graphs into Language Models (LMs) typically rely on a segregated architecture: external Graph Neural Networks (GNNs) encode structural topology, while LMs process textual semantics. We argue this approach is suboptimal for text-graphs: it creates a conceptually disjointed interaction paradigm. By segregating structural encoding from semantic processing, these systems must perform a complex implicit alignment between abstract graph tokens and concrete textual elements. Challenging the necessity of external encoders, we propose NAG (Native Architecture for Graphs), a unified framework that internalizes graph processing within the LM’s native manifold. Instead of bridging disparate embedding spaces, NAG repurposes the self-attention mechanism to enforce topological dependencies and recalibrates positional IDs to ensure structural equivalence. This allows the model to harness its intrinsic linguistic capability to simultaneously comprehend node and edge content alongside structural topology. We introduce two efficient implementations: NAG-Zero for absolute preservation of the base model’s linguistic capabilities, and NAG-LoRA for enhanced structural adaptation. Experiments across diverse graph tasks validate that NAG achieves robust graph comprehension without the overhead of external encoders, offering a simpler, more coherent paradigm for text-graph modeling.

Method: TSLM uses special tokens to encode branching structure, allowing models to generate and selectively expand multiple search paths within a single generation. Models are trained on complete search trees including both successful and failed attempts.

Result: TSLM achieves robust performance and superior inference efficiency by avoiding multiple independent forward passes required by external search methods, demonstrating efficient systematic exploration capabilities.

Conclusion: Supervised learning on complete tree-structured traces provides an efficient alternative for developing systematic exploration capabilities in language models, suggesting a new paradigm of inference-time scaling for robust reasoning.

Abstract: Language models generate reasoning sequentially, preventing them from decoupling irrelevant exploration paths during search. We introduce Tree-Structured Language Modeling (TSLM), which uses special tokens to encode branching structure, enabling models to generate and selectively expand multiple search paths within a single generation process. By training on complete search trees including both successful and failed attempts, TSLM learns to internalize systematic exploration without redundant recomputation of shared prefixes. TSLM achieves robust performance and superior inference efficiency by avoiding the multiple independent forward passes required by external search methods. These results suggest a new paradigm of inference-time scaling for robust reasoning, demonstrating that supervised learning on complete tree-structured traces provides an efficient alternative for developing systematic exploration capabilities in language models.

Method: Uses functional network fingerprinting based on consistency of neuronal activity patterns across models. Training-free approach that analyzes activation patterns with minimal samples, robust to fine-tuning, pruning, and architectural changes.

Result: Demonstrates high accuracy in detecting model lineage across different scales and architectures. Shows robustness to common modifications while preserving model utility.

Conclusion: FNF provides an effective, non-invasive tool for LLM intellectual property protection that works with few samples and across various model modifications.

Abstract: The development of large language models (LLMs) is costly and has significant commercial value. Consequently, preventing unauthorized appropriation of open-source LLMs and protecting developers’ intellectual property rights have become critical challenges. In this work, we propose the Functional Network Fingerprint (FNF), a training-free, sample-efficient method for detecting whether a suspect LLM is derived from a victim model, based on the consistency between their functional network activity. We demonstrate that models that share a common origin, even with differences in scale or architecture, exhibit highly consistent patterns of neuronal activity within their functional networks across diverse input samples. In contrast, models trained independently on distinct data or with different objectives fail to preserve such activity alignment. Unlike conventional approaches, our method requires only a few samples for verification, preserves model utility, and remains robust to common model modifications (such as fine-tuning, pruning, and parameter permutation), as well as to comparisons across diverse architectures and dimensionalities. FNF thus provides model owners and third parties with a simple, non-invasive, and effective tool for protecting LLM intellectual property. The code is available at https://github.com/WhatAboutMyStar/LLM_ACTIVATION.

Method: Introduces a dynamic format-alignment strategy using a lightweight classifier trained on latent model-preference signals to determine optimal format for each problem instance, rather than using human-designed heuristics.

Result: Achieves substantial and consistent improvements in zero-shot accuracy across reasoning and knowledge benchmarks, better revealing models’ latent capabilities.

Conclusion: Format choice significantly impacts LLM performance evaluation, and model-generated signals can effectively determine optimal formats, leading to more accurate assessment of model capabilities.

Abstract: Performance of Large Language Models (LLMs) on multiple-choice tasks differs markedly between symbol-based and cloze-style evaluation formats. The observed discrepancies are systematically attributable to task characteristics: natural language continuation benefits from likelihood scoring, whereas explicit comparison is better suited to symbol-based selection. These trends are consistent across various decoder-based LLMs, indicating model-agnostic effects. To address these inconsistencies, a dynamic format-alignment strategy is introduced that employs a lightweight classifier trained on latent model-preference signals. In contrast to human-designed heuristics, which often degrade performance, this approach uses model-generated signals to determine the optimal format for each problem instance. The proposed method achieves substantial and consistent improvements in zero-shot accuracy across reasoning and knowledge benchmarks, better revealing the models’ latent capabilities.

Method: Introduces MM-THEBench with: 1) fine-grained taxonomy grounded in cognitive dimensions, 2) diverse data with verified reasoning annotations, and 3) multi-level automated evaluation framework for assessing hallucinations in intermediate CoTs of reasoning MLLMs.

Result: Extensive experiments on mainstream reasoning MLLMs reveal insights into how thinking affects hallucination and reasoning capability across various multimodal tasks.

Conclusion: MM-THEBench addresses critical gaps in evaluating reasoning MLLMs by providing tools to assess hallucinations during thinking processes, offering insights into how reasoning affects multimodal perception and problem-solving.

Abstract: Recent advances in multimodal large language models (MLLMs) mark a shift from non-thinking models to post-trained reasoning models capable of solving complex problems through thinking. However, whether such thinking mitigates hallucinations in multimodal perception and reasoning remains unclear. Self-reflective reasoning enhances robustness but introduces additional hallucinations, and subtle perceptual errors still result in incorrect or coincidentally correct answers. Existing benchmarks primarily focus on models before the emergence of reasoning MLLMs, neglecting the internal thinking process and failing to measure the hallucinations that occur during thinking. To address these challenges, we introduce MM-THEBench, a comprehensive benchmark for assessing hallucinations of intermediate CoTs in reasoning MLLMs. MM-THEBench features a fine-grained taxonomy grounded in cognitive dimensions, diverse data with verified reasoning annotations, and a multi-level automated evaluation framework. Extensive experiments on mainstream reasoning MLLMs reveal insights into how thinking affects hallucination and reasoning capability in various multimodal tasks.

Method: The authors introduce the APPELLATE REVIEW task and construct AR-BENCH dataset with 8,700 annotated decisions and 34,617 supplementary corpora. They evaluate 14 large language models on this benchmark to assess their diagnostic reasoning capabilities for legal error detection.

Result: Evaluation of 14 LLMs reveals critical limitations in existing models’ ability to identify legal application errors, providing empirical evidence for future improvements in legal AI systems.

Conclusion: The APPELLATE REVIEW task and AR-BENCH benchmark address a significant gap in legal AI research, shifting focus from prediction/generation to error detection and correction, with findings showing current LLMs need substantial improvement for reliable legal error identification.

Abstract: Legal judgments may contain errors due to the complexity of case circumstances and the abstract nature of legal concepts, while existing appellate review mechanisms face efficiency pressures from a surge in case volumes. Although current legal AI research focuses on tasks like judgment prediction and legal document generation, the task of judgment review differs fundamentally in its objectives and paradigm: it centers on detecting, classifying, and correcting errors after a judgment is issued, constituting anomaly detection rather than prediction or generation. To address this research gap, we introduce a novel task APPELLATE REVIEW, aiming to assess models’ diagnostic reasoning and reliability in legal practice. We also construct a novel dataset benchmark AR-BENCH, which comprises 8,700 finely annotated decisions and 34,617 supplementary corpora. By evaluating 14 large language models, we reveal critical limitations in existing models’ ability to identify legal application errors, providing empirical evidence for future improvements.

Method: Proposes Retrieval-Augmented Simultaneous Speech Translation (RASST) with: 1) lightweight speech-text retriever, 2) efficient sliding-window retrieval for chunkwise terminology hints, and 3) synthetic training data to teach Speech LLMs to leverage retrieved terms precisely.

Result: Experiments on three language directions of ACL 60/60 dev set show RASST improves terminology translation accuracy by up to 16% and increases overall translation quality by up to 3 BLEU points, with ablations confirming each component’s contribution.

Conclusion: RASST effectively integrates cross-modal retrieval into simultaneous speech translation, significantly improving terminology handling and overall translation quality for Speech LLMs.

Abstract: Simultaneous speech translation (SST) produces target text incrementally from partial speech input. Recent speech large language models (Speech LLMs) have substantially improved SST quality, yet they still struggle to correctly translate rare and domain-specific terminology. While retrieval augmentation has been effective for terminology translation in machine translation, bringing retrieval to SST is non-trivial: it requires fast and accurate cross-modal (speech-to-text) retrieval under partial, continually arriving input, and the model must decide whether and when to apply retrieved terms during incremental generation. We propose Retrieval-Augmented Simultaneous Speech Translation (RASST), which tightly integrates cross-modal retrieval into the SST pipeline. RASST trains a lightweight speech-text retriever and performs efficient sliding-window retrieval, providing chunkwise terminology hints to the Speech LLM. We further synthesize training data that teaches the Speech LLM to leverage retrieved terms precisely. Experiments on three language directions of the ACL 60/60 dev set show that RASST improves terminology translation accuracy by up to 16% and increases overall translation quality by up to 3 BLEU points, with ablations confirming the contribution of each component.

Method: Developed a density estimator using mechanistic interpretability techniques to measure how uniformly computation is distributed across LLM parameters during processing.

Result: Found that: (1) LLM processing involves dense computation, (2) density is dynamic and input-dependent, (3) density patterns correlate across LLMs, (4) rarer tokens require higher density, (5) longer contexts decrease density.

Conclusion: The computation density estimator provides new insights into LLM processing, challenging symbolic interpretations and showing computation is more uniformly distributed than previously assumed.

Abstract: Transformer-based large language models (LLMs) are comprised of billions of parameters arranged in deep and wide computational graphs. Several studies on LLM efficiency optimization argue that it is possible to prune a significant portion of the parameters, while only marginally impacting performance. This suggests that the computation is not uniformly distributed across the parameters. We introduce here a technique to systematically quantify computation density in LLMs. In particular, we design a density estimator drawing on mechanistic interpretability. We experimentally test our estimator and find that: (1) contrary to what has been often assumed, LLM processing generally involves dense computation; (2) computation density is dynamic, in the sense that models shift between sparse and dense processing regimes depending on the input; (3) per-input density is significantly correlated across LLMs, suggesting that the same inputs trigger either low or high density. Investigating the factors influencing density, we observe that predicting rarer tokens requires higher density, and increasing context length often decreases the density. We believe that our computation density estimator will contribute to a better understanding of the processing at work in LLMs, challenging their symbolic interpretation.

Method: Uses activation patching (causal interpretability method) on EuroLLM during pretraining to isolate cross-lingual concept representations, then injects them into translation prompts to test consistency of translation alterations across languages.

Result: Shared concept spaces emerge early and continue refining, but alignment with them is language-dependent. Fine-grained analysis reveals some apparent translation quality gains actually reflect behavioral shifts (like sense selection for polysemous words or translation vs. copying of homographs) rather than improved translation ability.

Conclusion: Provides new insights into cross-lingual alignment training dynamics and conditions under which causal interpretability methods offer meaningful insights in multilingual contexts, highlighting the nuanced nature of what appears as translation improvement.

Abstract: Training Large Language Models (LLMs) with high multilingual coverage is becoming increasingly important – especially when monolingual resources are scarce. Recent studies have found that LLMs process multilingual inputs in shared concept spaces, thought to support generalization and cross-lingual transfer. However, these prior studies often do not use causal methods, lack deeper error analysis or focus on the final model only, leaving open how these spaces emerge during training. We investigate the development of language-agnostic concept spaces during pretraining of EuroLLM through the causal interpretability method of activation patching. We isolate cross-lingual concept representations, then inject them into a translation prompt to investigate how consistently translations can be altered, independently of the language. We find that shared concept spaces emerge early} and continue to refine, but that alignment with them is language-dependent}. Furthermore, in contrast to prior work, our fine-grained manual analysis reveals that some apparent gains in translation quality reflect shifts in behavior – like selecting senses for polysemous words or translating instead of copying cross-lingual homographs – rather than improved translation ability. Our findings offer new insight into the training dynamics of cross-lingual alignment and the conditions under which causal interpretability methods offer meaningful insights in multilingual contexts.

Method: Developed a semi-automated annotation methodology built on a faceted taxonomy, implemented through an annotation extension framework. Created an annotation extension tool for Turkish that automatically extends existing flat annotations by inferring additional linguistic and metadata information as facets within the taxonomy.

Result: The annotation extension tool achieved 95.86% facet-level accuracy. Produced the first collaboratively annotated and taxonomically enriched Turkish Learner Corpus with enhanced querying capabilities and detailed exploratory analysis support.

Conclusion: This work introduces a novel approach to learner corpus annotation that enables richer, multi-dimensional error analysis and is expected to pave the way for enriching existing error-annotated learner corpora.

Abstract: In terms of annotation structure, most learner corpora rely on holistic flat label inventories which, even when extensive, do not explicitly separate multiple linguistic dimensions. This makes linguistically deep annotation difficult and complicates fine-grained analyses aimed at understanding why and how learners produce specific errors. To address these limitations, this paper presents a semi-automated annotation methodology for learner corpora, built upon a recently proposed faceted taxonomy, and implemented through a novel annotation extension framework. The taxonomy provides a theoretically grounded, multi-dimensional categorization that captures the linguistic properties underlying each error instance, thereby enabling standardized, fine-grained, and interpretable enrichment beyond flat annotations. The annotation extension tool, implemented based on the proposed extension framework for Turkish, automatically extends existing flat annotations by inferring additional linguistic and metadata information as facets within the taxonomy to provide richer learner-specific context. It was systematically evaluated and yielded promising performance results, achieving a facet-level accuracy of 95.86%. The resulting taxonomically enriched corpus offers enhanced querying capabilities and supports detailed exploratory analyses across learner corpora, enabling researchers to investigate error patterns through complex linguistic and pedagogical dimensions. This work introduces the first collaboratively annotated and taxonomically enriched Turkish Learner Corpus, a manual annotation guideline with a refined tagset, and an annotation extender. As the first corpus designed in accordance with the recently introduced taxonomy, we expect our study to pave the way for subsequent enrichment efforts of existing error-annotated learner corpora.

Method: Created first Turkish skill extraction dataset (4,819 labeled spans from 327 job postings), evaluated LLMs with different prompting strategies (dynamic vs. static few-shot, varying context, causal reasoning), and used embedding-based retrieval with LLM reranking for skill linking to ESCO taxonomy.

Result: LLMs outperform supervised sequence labeling in end-to-end pipeline, with Claude Sonnet 3.7 with dynamic few-shot prompting achieving best end-to-end performance of 0.56, aligning Turkish with similar studies in other languages.

Conclusion: LLMs can improve skill extraction in low-resource settings, and this work should accelerate similar research for underrepresented languages.

Abstract: Skill extraction is a critical component of modern recruitment systems, enabling efficient job matching, personalized recommendations, and labor market analysis. Despite Türkiye’s significant role in the global workforce, Turkish, a morphologically complex language, lacks both a skill taxonomy and a dedicated skill extraction dataset, resulting in underexplored research in skill extraction for Turkish. This article seeks the answers to three research questions: 1) How can skill extraction be effectively performed for this language, in light of its low resource nature? 2)~What is the most promising model? 3) What is the impact of different Large Language Models (LLMs) and prompting strategies on skill extraction (i.e., dynamic vs. static few-shot samples, varying context information, and encouraging causal reasoning)? The article introduces the first Turkish skill extraction dataset and performance evaluations of automated skill extraction using LLMs. The manually annotated dataset contains 4,819 labeled skill spans from 327 job postings across different occupation areas. The use of LLM outperforms supervised sequence labeling when used in an end-to-end pipeline, aligning extracted spans with standardized skills in the ESCO taxonomy more effectively. The best-performing configuration, utilizing Claude Sonnet 3.7 with dynamic few-shot prompting for skill identification, embedding-based retrieval, and LLM-based reranking for skill linking, achieves an end-to-end performance of 0.56, positioning Turkish alongside similar studies in other languages, which are few in the literature. Our findings suggest that LLMs can improve skill extraction performance in low-resource settings, and we hope that our work will accelerate similar research on skill extraction for underrepresented languages.

Method: Developed MDial framework using rule-based LLM transformation with native linguist annotations to generate multi-dialectal data covering lexical, orthographic, and morphosyntactic features for 9 English dialects. Created MDialBench with 50k+ dialogs (97k+ QA pairs) to evaluate LLMs on dialect identification and response generation.

Result: Even frontier LLMs achieve under 70% accuracy on dialect identification, fail to reach 50% for Canadian English, and systematically misclassify non-SAE dialects as American or British. Annotators preferred MDial outputs over prior methods in 98% of comparisons for dialect naturalness.

Conclusion: LLMs struggle significantly with dialect understanding, risking cascading failures in downstream tasks. The research challenges the assumption that models should mirror users’ morphosyntactic features, finding that up to 90% of grammatical features should not be reproduced by models.

Abstract: More than 80% of the 1.6 billion English speakers do not use Standard American English (SAE) and experience higher failure rates and stereotyped responses when interacting with LLMs as a result. Yet multi-dialectal performance remains underexplored. We introduce $\textbf{MDial}$, the first large-scale framework for generating multi-dialectal conversational data encompassing the three pillars of written dialect – lexical (vocabulary), orthographic (spelling), and morphosyntactic (grammar) features – for nine English dialects. Partnering with native linguists, we design an annotated and scalable rule-based LLM transformation to ensure precision. Our approach challenges the assumption that models should mirror users’ morphosyntactic features, showing that up to 90% of the grammatical features of a dialect should not be reproduced by models. Independent evaluations confirm data quality, with annotators preferring MDial outputs over prior methods in 98% of pairwise comparisons for dialect naturalness. Using this pipeline, we construct the dialect-parallel $\textbf{MDialBench}$mark with 50k+ dialogs, resulting in 97k+ QA pairs, and evaluate 17 LLMs on dialect identification and response generation tasks. Even frontier models achieve under 70% accuracy, fail to reach 50% for Canadian English, and systematically misclassify non-SAE dialects as American or British. As dialect identification underpins natural language understanding, these errors risk cascading failures into downstream tasks.

Method: Used two established methods: (1) probing for token-level relational structures via attention-based explanations, and (2) feature-mapping using embeddings as carriers of human-interpretable properties.

Result: Both methods failed: Attention-based explanations collapsed when testing the assumption that later-layer representations correspond to tokens, and property-inference methods showed high predictive scores driven by methodological artifacts and dataset structure rather than meaningful semantic knowledge.

Conclusion: Widely-used interpretability techniques cannot reliably demonstrate what LLMs understand, which is particularly problematic in pervasive computing settings where these methods are relied upon for debugging, compression, and model explanation.

Abstract: Large Language Models (LLMs) are becoming increasingly popular in pervasive computing due to their versatility and strong performance. However, despite their ubiquitous use, the exact mechanisms underlying their outstanding performance remain unclear. Different methods for LLM explainability exist, and many are, as a method, not fully understood themselves. We started with the question of how linguistic abstraction emerges in LLMs, aiming to detect it across different LLM modules (attention heads and input embeddings). For this, we used methods well-established in the literature: (1) probing for token-level relational structures, and (2) feature-mapping using embeddings as carriers of human-interpretable properties. Both attempts failed for different methodological reasons: Attention-based explanations collapsed once we tested the core assumption that later-layer representations still correspond to tokens. Property-inference methods applied to embeddings also failed because their high predictive scores were driven by methodological artifacts and dataset structure rather than meaningful semantic knowledge. These failures matter because both techniques are widely treated as evidence for what LLMs supposedly understand, yet our results show such conclusions are unwarranted. These limitations are particularly relevant in pervasive and distributed computing settings where LLMs are deployed as system components and interpretability methods are relied upon for debugging, compression, and explaining models.

Method: Introduces CSM-MTBench with two expert-curated subsets: Fun Posts (context-rich, slang-heavy content) and Social Snippets (concise, emotion/style-driven expressions). Proposes tailored evaluation approaches: measuring slang/neologism translation success rate for Fun Posts, and assessing tone/style preservation via embedding-based metrics and LLM-as-a-judge for Social Snippets.

Result: Experiments on over 20 models reveal substantial variation in how current MT systems handle semantic fidelity and informal, social-media-specific stylistic cues, demonstrating the benchmark’s effectiveness in identifying system weaknesses.

Conclusion: CSM-MTBench serves as a rigorous testbed for advancing MT systems capable of mastering real-world Chinese social media texts, addressing both data scarcity and metric limitations in this challenging domain.

Abstract: The prevalence of rapidly evolving slang, neologisms, and highly stylized expressions in informal user-generated text, particularly on Chinese social media, poses significant challenges for Machine Translation (MT) benchmarking. Specifically, we identify two primary obstacles: (1) data scarcity, as high-quality parallel data requires bilingual annotators familiar with platform-specific slang, and stylistic cues in both languages; and (2) metric limitations, where traditional evaluators like COMET often fail to capture stylistic fidelity and nonstandard expressions. To bridge these gaps, we introduce CSM-MTBench, a benchmark covering five Chinese-foreign language directions and consisting of two expert-curated subsets: Fun Posts, featuring context-rich, slang- and neologism-heavy content, and Social Snippets, emphasizing concise, emotion- and style- driven expressions. Furthermore, we propose tailored evaluation approaches for each subset: measuring the translation success rate of slang and neologisms in Fun Posts, while assessing tone and style preservation in Social Snippets via a hybrid of embedding-based metrics and LLM-as-a-judge. Experiments on over 20 models reveal substantial variation in how current MT systems handle semantic fidelity and informal, social-media-specific stylistic cues. CSM-MTBench thus serves as a rigorous testbed for advancing MT systems capable of mastering real-world Chinese social media texts.

Method: Introduce alignment-flexible supervision via special token using connectionist temporal classification (CTC) objective during fine-tuning to relax strict positional supervision

Result: Method consistently outperforms original MDLM on five open-ended text generation benchmarks and improves robustness to positional shifts

Conclusion: Relaxing strict positional supervision is crucial for improving generation quality in MDLMs, addressing sensitivity to token misalignment

Abstract: Masked diffusion language models (MDLMs) have emerged as a promising alternative to dominant autoregressive approaches. Although they achieve competitive performance on several tasks, a substantial gap remains in open-ended text generation. We hypothesize that one cause of this gap is that strict positional prediction makes MDLM decoding highly sensitive to token misalignment, and we show through controlled interventions that a one-position shift can severely disrupt semantics. This observation suggests that enforcing strict positional supervision during training is misaligned with the irreversible denoising dynamics of MDLM decoding. Motivated by this mismatch, we adopt an alignment-flexible supervision strategy during fine-tuning. Specifically, we introduce a special token via the connectionist temporal classification objective. We apply this approach to the widely used MDLM model and conduct experiments on five open-ended text generation benchmarks. Our method consistently outperforms the original model and improves robustness to positional shifts, indicating that relaxing strict positional supervision is an important factor in improving generation quality in MDLMs.

Method: Proposes FraudCoT with: 1) fraud-aware selective CoT distillation to generate diverse reasoning paths, 2) integration of distilled CoTs into node texts to provide enriched semantic-structural cues to GNNs, and 3) efficient asymmetric co-training strategy for end-to-end optimization with reduced computational cost.

Result: Achieves up to 8.8% AUPRC improvement over state-of-the-art methods and delivers up to 1,066x speedup in training throughput on public and industrial benchmarks.

Conclusion: FraudCoT substantially advances both detection performance and efficiency for fraud detection on text-attributed graphs through autonomous reasoning and scalable co-training.

Abstract: Graph-based fraud detection on text-attributed graphs (TAGs) requires jointly modeling rich textual semantics and relational dependencies. However, existing LLM-enhanced GNN approaches are constrained by predefined prompting and decoupled training pipelines, limiting reasoning autonomy and weakening semantic-structural alignment. We propose FraudCoT, a unified framework that advances TAG-based fraud detection through autonomous, graph-aware chain-of-thought (CoT) reasoning and scalable LLM-GNN co-training. To address the limitations of predefined prompts, we introduce a fraud-aware selective CoT distillation mechanism that generates diverse reasoning paths and enhances semantic-structural understanding. These distilled CoTs are integrated into node texts, providing GNNs with enriched, multi-hop semantic and structural cues for fraud detection. Furthermore, we develop an efficient asymmetric co-training strategy that enables end-to-end optimization while significantly reducing the computational cost of naive joint training. Extensive experiments on public and industrial benchmarks demonstrate that FraudCoT achieves up to 8.8% AUPRC improvement over state-of-the-art methods and delivers up to 1,066x speedup in training throughput, substantially advancing both detection performance and efficiency.

Method: Proposes Residual Context Diffusion (RCD) that converts discarded token representations into contextual residuals and injects them back for the next denoising step. Uses a decoupled two-stage training pipeline to avoid memory bottlenecks.

Result: RCD improves frontier dLLMs by 5-10 points in accuracy with minimal extra computation. On challenging AIME tasks, it nearly doubles baseline accuracy and achieves 4-5x fewer denoising steps at equivalent accuracy levels.

Conclusion: RCD effectively recycles wasted computation in diffusion LLMs, significantly improving performance and efficiency across various reasoning and instruction-following benchmarks.

Abstract: Diffusion Large Language Models (dLLMs) have emerged as a promising alternative to purely autoregressive language models because they can decode multiple tokens in parallel. However, state-of-the-art block-wise dLLMs rely on a “remasking” mechanism that decodes only the most confident tokens and discards the rest, effectively wasting computation. We demonstrate that recycling computation from the discarded tokens is beneficial, as these tokens retain contextual information useful for subsequent decoding iterations. In light of this, we propose Residual Context Diffusion (RCD), a module that converts these discarded token representations into contextual residuals and injects them back for the next denoising step. RCD uses a decoupled two-stage training pipeline to bypass the memory bottlenecks associated with backpropagation. We validate our method on both long CoT reasoning (SDAR) and short CoT instruction following (LLaDA) models. We demonstrate that a standard dLLM can be efficiently converted to the RCD paradigm with merely ~1 billion tokens. RCD consistently improves frontier dLLMs by 5-10 points in accuracy with minimal extra computation overhead across a wide range of benchmarks. Notably, on the most challenging AIME tasks, RCD nearly doubles baseline accuracy and attains up to 4-5x fewer denoising steps at equivalent accuracy levels.

Method: The authors hypothesize “outlier-driven rescaling” where outliers work with normalization layers to rescale non-outlier components. They validate this across different model architectures and training token counts through ablation studies: removing normalization, clipping outliers, analyzing contributions, and testing mitigation strategies like learnable parameters and explicit gated rescaling.

Result: (1) Outliers function jointly with normalization - removing normalization eliminates outliers but degrades training; clipping outliers while keeping normalization also degrades performance. (2) Outliers serve more as rescale factors than contributors, with final contributions significantly smaller than non-outliers. (3) Outliers can be absorbed into learnable parameters or mitigated via explicit gated rescaling, improving training performance (average 2 points gain) and quantization robustness (1.2 points less degradation under W4A4 quantization).

Conclusion: Emergent outliers in LLMs (attention sinks and residual sinks) work together with normalization layers to perform outlier-driven rescaling, which contributes to training stability. This unified view explains both the origin and mitigation of sink types, with practical applications for improving training performance and quantization robustness.

Abstract: We investigate the functional role of emergent outliers in large language models, specifically attention sinks (a few tokens that consistently receive large attention logits) and residual sinks (a few fixed dimensions with persistently large activations across most tokens). We hypothesize that these outliers, in conjunction with the corresponding normalizations (\textit{e.g.}, softmax attention and RMSNorm), effectively rescale other non-outlier components. We term this phenomenon \textit{outlier-driven rescaling} and validate this hypothesis across different model architectures and training token counts. This view unifies the origin and mitigation of both sink types. Our main conclusions and observations include: (1) Outliers function jointly with normalization: removing normalization eliminates the corresponding outliers but degrades training stability and performance; directly clipping outliers while retaining normalization leads to degradation, indicating that outlier-driven rescaling contributes to training stability. (2) Outliers serve more as rescale factors rather than contributors, as the final contributions of attention and residual sinks are significantly smaller than those of non-outliers. (3) Outliers can be absorbed into learnable parameters or mitigated via explicit gated rescaling, leading to improved training performance (average gain of 2 points) and enhanced quantization robustness (1.2 points degradation under W4A4 quantization).

Method: Created 552 phrases across six Arabic varieties using LLMs, validated by five native speakers, stratified by difficulty, organized thematically. Built interactive web platform with translation exploration, adaptive quizzing with algorithmic distractor generation, cloud-synchronized progress tracking, and cultural context.

Result: Released complete open-source platform and dataset under MIT license, providing a comprehensive cross-dialectal Arabic learning resource with interactive features and cultural context.

Conclusion: ArabicDialectHub successfully addresses the gap in Arabic dialect learning resources by providing an accessible, validated, and interactive platform that supports learning multiple dialects with cultural context.

Abstract: We present ArabicDialectHub, a cross-dialectal Arabic learning resource comprising 552 phrases across six varieties (Moroccan Darija, Lebanese, Syrian, Emirati, Saudi, and MSA) and an interactive web platform. Phrases were generated using LLMs and validated by five native speakers, stratified by difficulty, and organized thematically. The open-source platform provides translation exploration, adaptive quizzing with algorithmic distractor generation, cloud-synchronized progress tracking, and cultural context. Both the dataset and complete platform source code are released under MIT license. Platform: https://arabic-dialect-hub.netlify.app.

Method: Large-scale multilingual evaluation across 50 countries/33 languages. Introduces Cross-Lingual Alignment Steering (CLAS) framework that aligns ideological representations across languages into shared subspace, with adaptive mechanism to prevent over-correction and preserve coherence.

Result: Substantial bias reduction along economic and social axes with minimal degradation in response quality. Framework enables scalable, interpretable fairness-aware multilingual LLM governance.

Conclusion: CLAS establishes effective paradigm for balancing ideological neutrality with linguistic/cultural diversity in multilingual LLMs, addressing critical gap in cross-lingual political bias mitigation.

Abstract: Large Language Models (LLMs) increasingly shape global discourse, making fairness and ideological neutrality essential for responsible AI deployment. Despite growing attention to political bias in LLMs, prior work largely focuses on high-resource, Western languages or narrow multilingual settings, leaving cross-lingual consistency and safe post-hoc mitigation underexplored. To address this gap, we present a large-scale multilingual evaluation of political bias spanning 50 countries and 33 languages. We introduce a complementary post-hoc mitigation framework, Cross-Lingual Alignment Steering (CLAS), designed to augment existing steering methods by aligning ideological representations across languages and dynamically regulating intervention strength. This method aligns latent ideological representations induced by political prompts into a shared ideological subspace, ensuring cross lingual consistency, with the adaptive mechanism prevents over correction and preserves coherence. Experiments demonstrate substantial bias reduction along both economic and social axes with minimal degradation in response quality. The proposed framework establishes a scalable and interpretable paradigm for fairness-aware multilingual LLM governance, balancing ideological neutrality with linguistic and cultural diversity.

Method: InstructDiff uses differential entropy between base models and minimally instruction-tuned calibrated models as a domain-adaptive selection criterion. It employs warmup calibration, bi-directional NLL filtering, and entropy-based ranking to identify optimal training samples.

Result: Achieves 17% relative improvement over full data training on mathematical reasoning and 52% for general instruction-following, outperforming prior baselines while using only 10% of the data.

Conclusion: Differential entropy reveals domain-adaptive patterns: reasoning tasks favor entropy increase (cognitive expansion) while general tasks favor entropy decrease (cognitive compression). This provides a unified framework for efficient data selection across domains.

Abstract: Supervised fine-tuning (SFT) is fundamental to adapting large language models, yet training on complete datasets incurs prohibitive costs with diminishing returns. Existing data selection methods suffer from severe domain specificity: techniques optimized for general instruction-following fail on reasoning tasks, and vice versa. We observe that measuring entropy differences between base models and minimally instruction-tuned calibrated models reveals a pattern – samples with the lowest differential entropy consistently yield optimal performance across domains, yet this principle manifests domain-adaptively: reasoning tasks favor entropy increase (cognitive expansion), while general tasks favor entropy decrease (cognitive compression). We introduce InstructDiff, a unified framework that operationalizes differential entropy as a domain-adaptive selection criterion through warmup calibration, bi-directional NLL filtering, and entropy-based ranking. Extensive experiments show that InstructDiff achieves 17% relative improvement over full data training on mathematical reasoning and 52% for general instruction-following, outperforming prior baselines while using only 10% of the data.

Method: Created DimABSA, a multilingual dimensional ABSA resource with 76,958 aspect instances across 42,590 sentences spanning 6 languages and 4 domains. Introduced three subtasks combining VA scores with ABSA elements, and proposed a new unified metric called continuous F1 (cF1) that incorporates VA prediction error into standard F1. Evaluated using both prompted and fine-tuned large language models.

Result: DimABSA is shown to be a challenging benchmark. The resource provides a foundation for advancing multilingual dimensional ABSA, with comprehensive benchmarking results showing the effectiveness of the proposed approach and metrics.

Conclusion: DimABSA successfully bridges traditional ABSA with dimensional sentiment analysis, enabling more fine-grained sentiment understanding through continuous valence-arousal scores. The resource and proposed metrics advance the field toward more nuanced affective analysis.

Abstract: Aspect-Based Sentiment Analysis (ABSA) focuses on extracting sentiment at a fine-grained aspect level and has been widely applied across real-world domains. However, existing ABSA research relies on coarse-grained categorical labels (e.g., positive, negative), which limits its ability to capture nuanced affective states. To address this limitation, we adopt a dimensional approach that represents sentiment with continuous valence-arousal (VA) scores, enabling fine-grained analysis at both the aspect and sentiment levels. To this end, we introduce DimABSA, the first multilingual, dimensional ABSA resource annotated with both traditional ABSA elements (aspect terms, aspect categories, and opinion terms) and newly introduced VA scores. This resource contains 76,958 aspect instances across 42,590 sentences, spanning six languages and four domains. We further introduce three subtasks that combine VA scores with different ABSA elements, providing a bridge from traditional ABSA to dimensional ABSA. Given that these subtasks involve both categorical and continuous outputs, we propose a new unified metric, continuous F1 (cF1), which incorporates VA prediction error into standard F1. We provide a comprehensive benchmark using both prompted and fine-tuned large language models across all subtasks. Our results show that DimABSA is a challenging benchmark and provides a foundation for advancing multilingual dimensional ABSA.

Method: Fine-tuning models across multiple domains and model families on data exhibiting specific character-level dispositions, comparing with incorrect-advice fine-tuning, and testing conditional activation via training-time triggers and inference-time persona-aligned prompts.

Result: Character-level disposition fine-tuning induces substantially stronger and more transferable misalignment than incorrect-advice fine-tuning while preserving general capabilities, showing emergent misalignment arises from stable behavioral shifts rather than capability degradation.

Conclusion: Character formation is a central and underexplored alignment risk, suggesting robust alignment must address behavioral dispositions rather than isolated errors or prompt-level defenses, revealing shared structure across emergent misalignment, backdoor activation, and jailbreak susceptibility.

Abstract: Emergent Misalignment refers to a failure mode in which fine-tuning large language models (LLMs) on narrowly scoped data induces broadly misaligned behavior. Prior explanations mainly attribute this phenomenon to the generalization of erroneous or unsafe content. In this work, we show that this view is incomplete. Across multiple domains and model families, we find that fine-tuning models on data exhibiting specific character-level dispositions induces substantially stronger and more transferable misalignment than incorrect-advice fine-tuning, while largely preserving general capabilities. This indicates that emergent misalignment arises from stable shifts in model behavior rather than from capability degradation or corrupted knowledge. We further show that such behavioral dispositions can be conditionally activated by both training-time triggers and inference-time persona-aligned prompts, revealing shared structure across emergent misalignment, backdoor activation, and jailbreak susceptibility. Overall, our results identify character formation as a central and underexplored alignment risk, suggesting that robust alignment must address behavioral dispositions rather than isolated errors or prompt-level defenses.

Method: DEF uses various levels of one-sentence textual cues to nudge LLMs to: 1) flag inconsistencies in policy texts, 2) refuse compliance with suspicious policies, and 3) fallback to their parametric knowledge when encountering perturbed versions. The approach is evaluated using globally recognized legal policies like HIPAA and GDPR.

Result: Empirical evaluations show DEF effectively improves frontier LLMs’ capability to detect and refuse perturbed policy versions, with DeepSeek-R1 achieving 100% detection rate in one setting. The protocol enhances LLM robustness against deceptive attacks exploiting legal artifacts.

Conclusion: The work demonstrates the value of cognitively inspired defenses for improving LLM robustness against deception, particularly in legal compliance contexts, and encourages further development of such mechanisms to protect against harm from manipulated policy texts.

Abstract: Humans develop a series of cognitive defenses, known as epistemic vigilance, to combat risks of deception and misinformation from everyday interactions. Developing safeguards for LLMs inspired by this mechanism might be particularly helpful for their application in high-stakes tasks such as automating compliance with data privacy laws. In this paper, we introduce Dynamic Epistemic Fallback (DEF), a dynamic safety protocol for improving an LLM’s inference-time defenses against deceptive attacks that make use of maliciously perturbed policy texts. Through various levels of one-sentence textual cues, DEF nudges LLMs to flag inconsistencies, refuse compliance, and fallback to their parametric knowledge upon encountering perturbed policy texts. Using globally recognized legal policies such as HIPAA and GDPR, our empirical evaluations report that DEF effectively improves the capability of frontier LLMs to detect and refuse perturbed versions of policies, with DeepSeek-R1 achieving a 100% detection rate in one setting. This work encourages further efforts to develop cognitively inspired defenses to improve LLM robustness against forms of harm and deception that exploit legal artifacts.

Method: Proposes Grounding Generation Utility (GroGU) metric that defines utility as a function of the downstream LLM’s generation confidence measured through entropy, requiring no annotations and being model-specific.

Result: GroGU faithfully distinguishes ground-truth documents and captures nuances ignored by LLM-agnostic metrics. When used to train a query-rewriter via Direct Preference Optimization, achieves improvements of up to 18.2 points in Mean Reciprocal Rank and 9.4 points in answer accuracy.

Conclusion: GroGU provides an effective, annotation-free approach to measure content utility in RAG systems, enabling better training of query-rewriters and improving overall RAG performance.

Abstract: Retrieval Augmented Generation (RAG)’s success depends on the utility the LLM derives from the content used for grounding. Quantifying content utility does not have a definitive specification and existing metrics ignore model-specific capabilities and/or rely on costly annotations. In this paper, we propose Grounding Generation Utility (GroGU), a model-specific and reference-free metric that defines utility as a function of the downstream LLM’s generation confidence based on entropy. Despite having no annotation requirements, GroGU is largely faithful in distinguishing ground-truth documents while capturing nuances ignored by LLM-agnostic metrics. We apply GroGU to train a query-rewriter for RAG by identifying high-utility preference data for Direct Preference Optimization. Experiments show improvements by up to 18.2 points in Mean Reciprocal Rank and up to 9.4 points in answer accuracy.

Method: Proposes a reference-free iterative monotonic process that leverages complementary feedback from theorem provers and LLM-based judges without access to ground-truth proofs or existing formalizations. Optimizes a masked composite objective over Formal Validity, Logical Preservation, Mathematical Consistency, and Formal Quality, guided by a responsiveness map indicating how different LLMs acting as different roles preferentially improve each dimension. Includes an acceptance policy guaranteeing certified monotonic improvement with conditions ensuring convergence and termination.

Result: Empirical experiments demonstrate the proposed process enables simultaneous improvement across multiple dimensions, achieving 93.44% formal validity and a 78.22% overall score on miniF2F, and 44.09% formal validity and a 29.79% overall score on ProofNet.

Conclusion: The proposed iterative monotonic process effectively addresses the challenge of jointly optimizing multiple quality dimensions in full-theorem autoformalization, achieving strong performance on benchmark datasets without requiring ground-truth references at inference time.

Abstract: While statement autoformalization has advanced rapidly, full-theorem autoformalization remains largely unexplored. Existing iterative refinement methods in statement autoformalization typicall improve isolated aspects of formalization, such as syntactic correctness, but struggle to jointly optimizing multiple quality dimensions, which is critical for full-theorem autoformalization. We introduce a reference-free iterative monotonic process for full-theorem autoformalization that leverages complementary feedback from theorem provers and LLM-based judges, without access to ground-truth proofs or existing formalizations at inference time. Our approach optimizes a masked composite objective over Formal Validity, Logical Preservation, Mathematical Consistency, and Formal Quality, guided by a responsiveness map that indicates how different LLMs acting as different roles preferentially improve each dimension. We further propose an acceptance policy that guarantees certified monotonic improvement, and provide conditions ensuring convergence and termination. Empirical experiments demonstrate the proposed process enables simultaneous improvement across multiple dimensions, achieving 93.44% formal validity and a 78.22% overall score on miniF2F, and 44.09% formal validity and a 29.79% overall score on ProofNet.

Method: The paper presents the first frequency-domain analysis of dLLMs, showing that low-frequency components encode global structural information while high-frequency components handle local details. Based on this, they propose FourierSampler, which uses a frequency-domain sliding window mechanism to dynamically guide the model toward “structure-to-detail” generation.

Result: FourierSampler outperforms other inference enhancement strategies on LLADA and SDAR benchmarks, achieving relative improvements of 20.4% on LLaDA1.5-8B and 16.0% on LLaDA-8B-Instruct. It notably surpasses similarly sized autoregressive models like Llama3.1-8B-Instruct.

Conclusion: The spectral analysis reveals important insights about dLLMs, and FourierSampler demonstrates that frequency-domain guidance can significantly improve diffusion language model decoding, achieving better performance than both existing dLLM strategies and comparable autoregressive models.

Abstract: Despite the non-autoregressive potential of diffusion language models (dLLMs), existing decoding strategies demonstrate positional bias, failing to fully unlock the potential of arbitrary generation. In this work, we delve into the inherent spectral characteristics of dLLMs and present the first frequency-domain analysis showing that low-frequency components in hidden states primarily encode global structural information and long-range dependencies, while high-frequency components are responsible for characterizing local details. Based on this observation, we propose FourierSampler, which leverages a frequency-domain sliding window mechanism to dynamically guide the model to achieve a “structure-to-detail” generation. FourierSampler outperforms other inference enhancement strategies on LLADA and SDAR, achieving relative improvements of 20.4% on LLaDA1.5-8B and 16.0% on LLaDA-8B-Instruct. It notably surpasses similarly sized autoregressive models like Llama3.1-8B-Instruct.

Method: Created a dataset using a data generation pipeline from real-world sources with de-identification and synthesis, developed a cost-effective human-in-the-loop translation pipeline (TEaR methodology) with MQM error annotations and selective post-editing for 5 languages, and evaluated using LLM-as-judge approach.

Result: Baseline evaluations show higher performance on English and Spanish but substantial degradation for other languages (Italian, German, Chinese), revealing critical gaps in multilingual MRC capabilities for HR applications.

Conclusion: JobResQA provides a reproducible benchmark for advancing fair and reliable LLM-based HR systems, highlighting the need for improved multilingual capabilities in HR-specific applications.

Abstract: We introduce JobResQA, a multilingual Question Answering benchmark for evaluating Machine Reading Comprehension (MRC) capabilities of LLMs on HR-specific tasks involving résumés and job descriptions. The dataset comprises 581 QA pairs across 105 synthetic résumé-job description pairs in five languages (English, Spanish, Italian, German, and Chinese), with questions spanning three complexity levels from basic factual extraction to complex cross-document reasoning. We propose a data generation pipeline derived from real-world sources through de-identification and data synthesis to ensure both realism and privacy, while controlled demographic and professional attributes (implemented via placeholders) enable systematic bias and fairness studies. We also present a cost-effective, human-in-the-loop translation pipeline based on the TEaR methodology, incorporating MQM error annotations and selective post-editing to ensure an high-quality multi-way parallel benchmark. We provide a baseline evaluations across multiple open-weight LLM families using an LLM-as-judge approach revealing higher performances on English and Spanish but substantial degradation for other languages, highlighting critical gaps in multilingual MRC capabilities for HR applications. JobResQA provides a reproducible benchmark for advancing fair and reliable LLM-based HR systems. The benchmark is publicly available at: https://github.com/Avature/jobresqa-benchmark

Method: Formulates latent reasoning within VAE framework, renders explicit reasoning chains as images, extracts visual-semantic representations to regularize posterior distribution, enabling efficient compression with minimal information loss.

Result: Significantly outperforms existing latent reasoning methods in computational efficiency and reasoning effectiveness, and even surpasses CoT through multi-modal reasoning.

Conclusion: Provides a new and insightful solution to latent reasoning that combines multi-modal guidance with variational learning for efficient and effective reasoning.

Abstract: While Chain-of-Thought (CoT) significantly enhances the performance of Large Language Models (LLMs), explicit reasoning chains introduce substantial computational redundancy. Recent latent reasoning methods attempt to mitigate this by compressing reasoning processes into latent space, but often suffer from severe performance degradation due to the lack of appropriate compression guidance. In this study, we propose Rendered CoT-Guided variational Latent Reasoning (ReGuLaR), a simple yet novel latent learning paradigm resolving this issue. Fundamentally, we formulate latent reasoning within the Variational Auto-Encoding (VAE) framework, sampling the current latent reasoning state from the posterior distribution conditioned on previous ones. Specifically, when learning this variational latent reasoning model, we render explicit reasoning chains as images, from which we extract dense visual-semantic representations to regularize the posterior distribution, thereby achieving efficient compression with minimal information loss. Extensive experiments demonstrate that ReGuLaR significantly outperforms existing latent reasoning methods across both computational efficiency and reasoning effectiveness, and even surpasses CoT through multi-modal reasoning, providing a new and insightful solution to latent reasoning. Code: https://github.com/FanmengWang/ReGuLaR.

Method: Proposes Deep Search with Meta-Cognitive Monitoring (DS-MCM), which integrates a Fast Consistency Monitor (lightweight checks on alignment between external evidence and internal reasoning confidence) and a Slow Experience-Driven Monitor (selectively activated to guide corrective intervention based on experience memory from historical agent trajectories). The monitoring is embedded directly into the reasoning-retrieval loop to determine both when intervention is warranted and how corrective actions should be informed by prior experience.

Result: Experiments across multiple deep search benchmarks and backbone models demonstrate that DS-MCM consistently improves performance and robustness compared to baseline approaches.

Conclusion: The hierarchical metacognitive monitoring framework effectively addresses the limitations of current deep search agents by providing mechanisms for state monitoring and regulation, leading to more reliable and robust task execution under uncertainty.

Abstract: Deep search agents powered by large language models have demonstrated strong capabilities in multi-step retrieval, reasoning, and long-horizon task execution. However, their practical failures often stem from the lack of mechanisms to monitor and regulate reasoning and retrieval states as tasks evolve under uncertainty. Insights from cognitive neuroscience suggest that human metacognition is hierarchically organized, integrating fast anomaly detection with selectively triggered, experience-driven reflection. In this work, we propose Deep Search with Meta-Cognitive Monitoring (DS-MCM), a deep search framework augmented with an explicit hierarchical metacognitive monitoring mechanism. DS-MCM integrates a Fast Consistency Monitor, which performs lightweight checks on the alignment between external evidence and internal reasoning confidence, and a Slow Experience-Driven Monitor, which is selectively activated to guide corrective intervention based on experience memory from historical agent trajectories. By embedding monitoring directly into the reasoning-retrieval loop, DS-MCM determines both when intervention is warranted and how corrective actions should be informed by prior experience. Experiments across multiple deep search benchmarks and backbone models demonstrate that DS-MCM consistently improves performance and robustness.

Method: Systematically constructed semantically natural prompts ending with partial tokens across three domains: languages without whitespace (Chinese), highly compounding languages, and code. Evaluated frontier LMs on these prompts and compared with token-aligned “backed-off” versions.

Result: Found serious failure mode: LMs place three orders of magnitude less probability on correct continuations for partial-token prompts vs. token-aligned ones. Degradation doesn’t diminish with scale and often worsens for larger models. Validated effectiveness of recent exact solutions.

Conclusion: Demonstrates scale and severity of probability distortion from tokenization in realistic use cases, provides practical recommendations for model inference providers to mitigate the partial token problem.

Abstract: Language models (LMs) are trained over sequences of tokens, whereas users interact with LMs via text. This mismatch gives rise to the partial token problem, which occurs when a user ends their prompt in the middle of the expected next-token, leading to distorted next-token predictions. Although this issue has been studied using arbitrary character prefixes, its prevalence and severity in realistic prompts respecting word boundaries remains underexplored. In this work, we identify three domains where token and “word” boundaries often do not line up: languages that do not use whitespace, highly compounding languages, and code. In Chinese, for example, up to 25% of word boundaries do not line up with token boundaries, making even natural, word-complete prompts susceptible to this problem. We systematically construct semantically natural prompts ending with a partial tokens; in experiments, we find that they comprise a serious failure mode: frontier LMs consistently place three orders of magnitude less probability on the correct continuation compared to when the prompt is “backed-off” to be token-aligned. This degradation does not diminish with scale and often worsens for larger models. Finally, we evaluate inference-time mitigations to the partial token problem and validate the effectiveness of recent exact solutions. Overall, we demonstrate the scale and severity of probability distortion caused by tokenization in realistic use cases, and provide practical recommentions for model inference providers.

Method: Designed a text-to-audio jailbreak attack using an advanced instruction-following TTS model to embed disallowed directives within narrative-style audio streams. The attack exploits structural and acoustic properties of speech to circumvent text-based safety mechanisms.

Result: The attack achieved 98.26% success rate on state-of-the-art models including Gemini 2.0 Flash, substantially exceeding text-only baselines. The narrative format delivered through synthetic speech effectively elicits restricted outputs.

Conclusion: Speech-based interfaces introduce distinct security vulnerabilities that require safety frameworks jointly reasoning over linguistic and paralinguistic representations, as current text-calibrated mechanisms are insufficient for audio inputs.

Abstract: Large audio-language models increasingly operate on raw speech inputs, enabling more seamless integration across domains such as voice assistants, education, and clinical triage. This transition, however, introduces a distinct class of vulnerabilities that remain largely uncharacterized. We examine the security implications of this modality shift by designing a text-to-audio jailbreak that embeds disallowed directives within a narrative-style audio stream. The attack leverages an advanced instruction-following text-to-speech (TTS) model to exploit structural and acoustic properties, thereby circumventing safety mechanisms primarily calibrated for text. When delivered through synthetic speech, the narrative format elicits restricted outputs from state-of-the-art models, including Gemini 2.0 Flash, achieving a 98.26% success rate that substantially exceeds text-only baselines. These results highlight the need for safety frameworks that jointly reason over linguistic and paralinguistic representations, particularly as speech-based interfaces become more prevalent.

Method: An agentic framework powered by state-of-the-art VLMs and image generation models that orchestrates specialized agents for reference retrieval, content/style planning, image rendering, and iterative refinement through self-critique.

Result: PaperBanana consistently outperforms leading baselines in faithfulness, conciseness, readability, and aesthetics on PaperBananaBench (292 test cases from NeurIPS 2025), and effectively extends to high-quality statistical plot generation.

Conclusion: PaperBanana paves the way for automated generation of publication-ready illustrations, addressing a significant bottleneck in research workflows.

Abstract: Despite rapid advances in autonomous AI scientists powered by language models, generating publication-ready illustrations remains a labor-intensive bottleneck in the research workflow. To lift this burden, we introduce PaperBanana, an agentic framework for automated generation of publication-ready academic illustrations. Powered by state-of-the-art VLMs and image generation models, PaperBanana orchestrates specialized agents to retrieve references, plan content and style, render images, and iteratively refine via self-critique. To rigorously evaluate our framework, we introduce PaperBananaBench, comprising 292 test cases for methodology diagrams curated from NeurIPS 2025 publications, covering diverse research domains and illustration styles. Comprehensive experiments demonstrate that PaperBanana consistently outperforms leading baselines in faithfulness, conciseness, readability, and aesthetics. We further show that our method effectively extends to the generation of high-quality statistical plots. Collectively, PaperBanana paves the way for the automated generation of publication-ready illustrations.

Method: UPA uses LLMs for fine-grained pairwise comparisons to build a tree structure exploring prompt space, then applies a two-stage framework: 1) Bayesian aggregation of local comparisons to filter candidates, and 2) global tournament-style comparisons using the Bradley-Terry-Luce model to infer latent prompt quality.

Result: UPA consistently outperforms existing prompt optimization methods across multiple tasks, demonstrating that agent-style optimization remains effective even in fully unsupervised settings.

Conclusion: The proposed unsupervised approach enables effective prompt optimization without requiring supervised feedback, making prompt agent methods more practical for real-world applications.

Abstract: Prompt agents have recently emerged as a promising paradigm for automated prompt optimization, framing refinement as a sequential decision-making problem over a structured prompt space. While this formulation enables the use of advanced planning algorithms, these methods typically assume access to supervised reward signals, which are often unavailable in practical scenarios. In this work, we propose UPA, an Unsupervised Prompt Agent that realizes structured search and selection without relying on supervised feedback. Specifically, during search, UPA iteratively constructs an evolving tree structure to navigate the prompt space, guided by fine-grained and order-invariant pairwise comparisons from Large Language Models (LLMs). Crucially, as these local comparisons do not inherently yield a consistent global scale, we decouple systematic prompt exploration from final selection, introducing a two-stage framework grounded in the Bradley-Terry-Luce (BTL) model. This framework first performs path-wise Bayesian aggregation of local comparisons to filter candidates under uncertainty, followed by global tournament-style comparisons to infer latent prompt quality and identify the optimal prompt. Experiments across multiple tasks demonstrate that UPA consistently outperforms existing prompt optimization methods, showing that agent-style optimization remains highly effective even in fully unsupervised settings.

Method: Created OMGEval with 804 open-ended questions per language covering capabilities like general knowledge and logical reasoning, with human verification and cultural localization for non-English languages. Uses GPT-4 as adjudicator for automatic scoring.

Result: Evaluated several representative multilingual LLMs, providing a valuable reference for understanding and improving multilingual capabilities. GPT-4 scoring shown to correlate closely with human evaluation.

Conclusion: OMGEval addresses the need for multilingual evaluation benchmarks and will help the community better understand and enhance LLM capabilities across different languages and cultures.

Abstract: Modern large language models (LLMs) should generally benefit individuals from various cultural backgrounds around the world. However, most recent advanced generative evaluation benchmarks tailed for LLMs mainly focus on English. To this end, we introduce OMGEval, the first Open-source Multilingual Generative test set that can assess the capability of LLMs in different languages. For each language, OMGEval provides 804 open-ended questions, covering a wide range of important capabilities of LLMs, such as general knowledge, logical reasoning, and so on. Each question is rigorously verified by human annotators. Notably, to sufficiently reflect the compatibility of LLMs in different cultural backgrounds, we perform localization for each non-English language. Specifically, the current version of OMGEval includes 5 languages (i.e., Zh, Ru, Fr, Es, Ar). Following AlpacaEval, we employ GPT-4 as the adjudicator to automatically score different model outputs, which is shown closely related to human evaluation. We evaluate several representative multilingual LLMs on the proposed OMGEval, which we believe will provide a valuable reference for the community to further understand and improve the multilingual capability of LLMs. OMGEval is available at https://github.com/blcuicall/OMGEval.

Method: Used SVM with three text vectorization methods (TF-IDF, Word2Vec, BoW) and compared against BERT transformer model. Included detailed preprocessing, rigorous model implementation, and thorough evaluation.

Result: BERT achieved superior accuracy (99.98%) and F1-score (0.9998). SVM with linear kernel and BoW vectorization performed exceptionally well with 99.81% accuracy and 0.9980 F1-score. SVM with BoW and TF-IDF offered highly competitive performance close to BERT.

Conclusion: While BERT shows best performance, SVM models with BoW and TF-IDF vectorization provide competitive results with advantage of lower computational requirements, making them practical alternatives for fake news detection.

Abstract: The rapid spread of misinformation, particularly through online platforms, underscores the urgent need for reliable detection systems. This study explores the utilization of machine learning and natural language processing, specifically Support Vector Machines (SVM) and BERT, to detect fake news. We employ three distinct text vectorization methods for SVM: Term Frequency Inverse Document Frequency (TF-IDF), Word2Vec, and Bag of Words (BoW), evaluating their effectiveness in distinguishing between genuine and fake news. Additionally, we compare these methods against the transformer large language model, BERT. Our comprehensive approach includes detailed preprocessing steps, rigorous model implementation, and thorough evaluation to determine the most effective techniques. The results demonstrate that while BERT achieves superior accuracy with 99.98% and an F1-score of 0.9998, the SVM model with a linear kernel and BoW vectorization also performs exceptionally well, achieving 99.81% accuracy and an F1-score of 0.9980. These findings highlight that, despite BERT’s superior performance, SVM models with BoW and TF-IDF vectorization methods come remarkably close, offering highly competitive performance with the advantage of lower computational requirements.

Method: Two-stage attention design: 1) Normalization stage replaces Softmax with Softplus + L1 normalization with dynamic scale factor based on invariance entropy; 2) Sharpening stage re-weights attention to amplify significant weights and diminish weaker ones.

Result: Achieves nearly constant validation loss at 16× training length, superior performance on long-context retrieval tasks, and enables models to recover Newton’s gravitational law from orbital trajectory sequences.

Conclusion: The proposed two-stage attention mechanism ensures numerical stability, dramatically improves length extrapolation, and provides evidence that appropriate attention mechanisms are crucial for foundation models to develop genuine physical world models.

Abstract: Large language models have achieved remarkable success in recent years, primarily due to self-attention. However, traditional Softmax attention suffers from numerical instability and reduced performance as the number of inference tokens increases. This work addresses these issues by proposing a new design principle for attention, viewing it as a two-stage process. The first stage (normalisation) refines standard attention by replacing Softmax with the more numerically stable Softplus followed by $l_{1}$-normalisation. Furthermore, we introduce a dynamic scale factor based on invariance entropy. We show that this novel attention mechanism outperforms conventional Softmax attention, and state-of-the-art Softmax-free alternatives. Our second proposal is to introduce a second processing stage (sharpening) which consists of a re-weighting mechanism that amplifies significant attentional weights while diminishing weaker ones. This enables the model to concentrate more effectively on relevant tokens, mitigating the attention sink phenomenon, and fundamentally improving length extrapolation. This novel, two-stage, replacement for self-attention is shown to ensure numerical stability and dramatically improve length extrapolation, maintaining a nearly constant validation loss at 16$\times$ the training length while achieving superior results on challenging long-context retrieval tasks and downstream benchmarks. Furthermore, symbolic regression experiments demonstrate that our method enables models to recover Newton’s gravitational law from orbital trajectory sequences, providing evidence that appropriate attention mechanisms are crucial for foundation models to develop genuine physical world models.

Method: Proposes DeepGreen, a dual-stage LLM-driven system for detecting potential corporate greenwashing in annual reports. Uses Retrieval-Augmented Generation (RAG) to reduce hallucinations. Applied to 9,369 A-share annual reports (2021-2023) with validation through ablation experiments, IV, PSM, and placebo tests.

Result: DeepGreen attains high reliability in validation. Greenwashing detected reveals positive relationship with environmental penalties. Green investors weaken this correlation. Relationship is less significant in large corporations and those with accumulated green assets, suggesting green assets may serve as credibility shields.

Conclusion: LLMs can standardize ESG oversight by providing early warning and directing regulatory attention to corporations where monitoring is more warranted. Demonstrates LLMs’ potential for reliable greenwashing detection and causal analysis in economic research.

Abstract: Motivated by the emerging adoption of Large Language Models (LLMs) in economics and management research, this paper investigates whether LLMs can reliably identify corporate greenwashing narratives and, more importantly, whether and how the greenwashing signals extracted from textual disclosures can be used to empirically identify causal effects. To this end, this paper proposes DeepGreen, a dual-stage LLM-Driven system for detecting potential corporate greenwashing in annual reports. Applied to 9369 A-share annual reports published between 2021 and 2023, DeepGreen attains high reliability in random-sample validation at both stages. Ablation experiment shows that Retrieval-Augmented Generation (RAG) reduces hallucinations, as compared to simply lengthening the input window. Empirical tests indicate that “greenwashing” captured by DeepGreen can effectively reveal a positive relationship between greenwashing and environmental penalties, and IV, PSM, Placebo test, which enhance the robustness and causal effects of the empirical evidence. Further study suggests that the presence and number of green investors can weaken the positive correlation between greenwashing and penalties. Heterogeneity analysis shows that the positive relationship between “greenwashing - penalty” is less significant in large-sized corporations and corporations that have accumulated green assets, indicating that these green assets may be exploited as a credibility shield for greenwashing. Our findings demonstrate that LLMs can standardize ESG oversight by early warning and direct regulators’ scarce attention toward the subsets of corporations where monitoring is more warranted.

Method: Introduces Discriminative Alignment Index (DALI) to quantify instance-level alignment across 24 languages and 3 NLU tasks, then uses activation patching to test causal relationships between alignment and prediction correctness.

Result: Incorrect NLU predictions strongly correlate with lower representation alignment with English in middle layers. Activation patching shows incorrect predictions can be fixed by patching with parallel English activations.

Conclusion: Representation (mis)alignment in middle layers plays a causal role in cross-lingual NLU performance, providing insights into how LLMs generalize across languages.

Abstract: Large language models (LLMs) can answer prompts in many languages, despite being trained predominantly on English; yet, the mechanisms driving this generalization remain poorly understood. This work asks: How does an LLM’s ability to align representations of non-English inputs to English impact its performance on natural language understanding (NLU) tasks? We study the role of representation alignment in instance-level task decisions, complementing prior analyses conducted both at the language level and task-independently. We introduce the Discriminative Alignment Index ($\DALI$) to quantify instance-level alignment across 24 languages other than English and three distinct NLU tasks. Results show that incorrect NLU predictions are strongly associated with lower representation alignment with English in the model’s middle layers. Through activation patching, we show that incorrect predictions in languages other than English can be fixed by patching their parallel English activations in the middle layers, thereby demonstrating the causal role of representation (mis)alignment in cross-lingual correctness.

Method: Proposed unified setup to compare 7 representative architectures (including xLSTM, GLA, Gated DeltaNet) across parameter scales (140M to 1.7B) and token budgets, analyzing scaling behavior and instruction tuning adaptation.

Result: Performance hierarchies remain stable across scales; error-correcting update rules have superior scaling exponents; gaps established early persist; only gated delta-rule models maintain precision for long-context retrieval while additive models suffer state saturation.

Conclusion: Architectural inductive biases are the primary constraint for successful linearization, not training compute scaling; gated delta-rule formulations are particularly effective for maintaining long-context capabilities.

Abstract: Linearization has emerged as a strategy for developing efficient language models (LMs). Starting from an existing Transformer-based LM, linearization replaces the attention component with computationally efficient subquadratic \textit{token mixers}. However, as an increasing number of mixers are proposed, it remains unclear which inductive biases are best suited to inherit the original Transformer’s capabilities. Furthermore, it is unknown how linearization is affected by parameter and token budget scaling. To address these questions, we propose a unified setup to compare seven representative architectures, including xLSTM, GLA, and Gated DeltaNet. Our findings reveal that performance hierarchies remain stable from 140M to 1.7B parameters, with error-correcting update rules demonstrating superior scaling exponents. We show that performance gaps are established early and persist through asymptotic maturity at 10B tokens, suggesting that state resolution is a more fundamental bottleneck than the distillation budget. Finally, while most models adapt to instruction tuning, only gated delta-rule formulations maintain the precision necessary for long-context retrieval, whereas additive models suffer from irreversible state saturation. These results suggest that for successful linearization, architectural inductive biases remain the primary constraint that cannot be overcome by simply scaling training compute.

Method: Created a large-scale benchmark of 8,072 assembly programs, evaluated 23 LLMs, then fine-tuned Qwen2.5-Coder-7B-Instruct with reinforcement learning using a reward function combining correctness and performance speedup, with additional techniques like Best-of-N sampling and iterative refinement.

Result: The fine-tuned model SuperCoder achieved 95.0% correctness and 1.46x average speedup over gcc -O3, significantly outperforming the best baseline Claude-opus-4 (51.5% test-passing rate, 1.43x speedup).

Conclusion: LLMs can effectively serve as superoptimizers for assembly programs, establishing a new foundation for program performance optimization beyond traditional compiler approaches.

Abstract: Superoptimization is the task of transforming a program into a faster one while preserving its input-output behavior. In this work, we investigate whether large language models (LLMs) can serve as superoptimizers, generating assembly programs that outperform code already optimized by industry-standard compilers. We construct the first large-scale benchmark for this problem, consisting of 8,072 assembly programs averaging 130 lines, in contrast to prior datasets restricted to 2-15 straight-line, loop-free programs. We evaluate 23 LLMs on this benchmark and find that the strongest baseline, Claude-opus-4, achieves a 51.5% test-passing rate and a 1.43x average speedup over gcc -O3. To further enhance performance, we fine-tune models with reinforcement learning, optimizing a reward function that integrates correctness and performance speedup. Starting from Qwen2.5-Coder-7B-Instruct (61.4% correctness, 1.10x speedup), the fine-tuned model SuperCoder attains 95.0% correctness and 1.46x average speedup, with additional improvement enabled by Best-of-N sampling and iterative refinement. Our results demonstrate, for the first time, that LLMs can be applied as superoptimizers for assembly programs, establishing a foundation for future research in program performance optimization beyond compiler heuristics.

Method: The authors conduct experiments on Associative Recall tasks and use causal interventions to analyze different architectures (Transformers, Based, Mamba, DeltaNet, H3, Hyena). They investigate mechanisms through which each architecture succeeds or fails, and introduce a new hierarchical retrieval task called Associative Treecall (ATR) to further test these mechanisms.

Result: Only Transformers and Based SSMs fully succeed at AR, with Mamba and DeltaNet close behind, while H3 and Hyena fail. Transformers and Based learn to store key-value associations in-context using induction, while SSMs compute these associations only at the last state. Mamba implements induction not via its SSM but through short convolutions. All architectures show the same mechanisms on the new ATR task.

Conclusion: Architectures with similar accuracy can have substantive mechanistic differences, highlighting the importance of mechanistic evaluations beyond just performance metrics. The findings reveal that Mamba’s success comes from short convolutions rather than its SSM component, and that SSMs generally struggle with in-context associative recall compared to Transformers.

Abstract: State space models (SSMs) for language modelling promise an efficient and performant alternative to quadratic-attention Transformers, yet show variable performance on recalling basic information from the context. While performance on synthetic tasks like Associative Recall (AR) can point to this deficiency, behavioural metrics provide little information as to \textit{why} – on a mechanistic level – certain architectures fail and others succeed. To address this, we conduct experiments on AR, and find that only Transformers and Based SSM models fully succeed at AR, with Mamba and DeltaNet close behind, while the other SSMs (H3, Hyena) fail. We then use causal interventions to explain why. We find that Transformers and Based learn to store key-value associations in-context using induction. By contrast, the SSMs seem to compute these associations only at the last state using a single layer. We further investigate the mechanism underlying the success of Mamba, and find novel evidence that Mamba \textit{does} implement induction: not via the SSM, but instead via short convolutions. Further experiments on a new hierarchical retrieval task, Associative Treecall (ATR), show that all architectures learn the same mechanism as they did for AR. Furthermore, we show that Mamba can learn Attention-like induction on ATR when short convolutions are removed. These results reveal that architectures with similar accuracy may still have substantive differences, motivating the adoption of mechanistic evaluations.

Method: Introduces Diverse-NS, a length-controlled data selection strategy that generates and filters preference data balancing diversity, quality, and length. The method requires only 3,000 preference pairs and can use smaller models as “diversity teachers” for larger models.

Result: Applied to LLaMA-3.1-8B and Olmo-2 family models, Diverse-NS substantially enhances lexical and semantic diversity. Shows consistent improvement in diversity with minor reduction or gains in response quality on four creative generation tasks: Divergent Associations, Persona Generation, Alternate Uses, and Creative Writing.

Conclusion: By explicitly addressing length bias, Diverse-NS efficiently pushes models toward more diverse and expressive outputs, enabling smaller models to serve as effective “diversity teachers” for larger models.

Abstract: Diverse language model responses are crucial for creative generation, open-ended tasks, and self-improvement training. We show that common diversity metrics, and even reward models used for preference optimization, systematically bias models toward shorter outputs, limiting expressiveness. To address this, we introduce Diverse, not Short (Diverse-NS), a length-controlled data selection strategy that improves response diversity while maintaining length parity. By generating and filtering preference data that balances diversity, quality, and length, Diverse-NS enables effective training using only 3,000 preference pairs. Applied to LLaMA-3.1-8B and the Olmo-2 family, Diverse-NS substantially enhances lexical and semantic diversity. We show consistent improvement in diversity with minor reduction or gains in response quality on four creative generation tasks: Divergent Associations, Persona Generation, Alternate Uses, and Creative Writing. Surprisingly, experiments with the Olmo-2 model family (7B, and 13B) show that smaller models like Olmo-2-7B can serve as effective “diversity teachers” for larger models. By explicitly addressing length bias, our method efficiently pushes models toward more diverse and expressive outputs.

Method: Proposes model immunization: supervised fine-tuning on curated (false claim, correction) pairs injected as small “vaccine doses” (5-10% of tokens) alongside truthful data. This provides direct negative supervision on labeled falsehoods, unlike post-hoc filtering or preference alignment.

Result: Across four open-weight model families, immunization improves TruthfulQA accuracy by 12 points and misinformation rejection by 30 points with negligible capability loss. The approach demonstrates effectiveness in reducing misinformation reproduction.

Conclusion: Model immunization is an effective approach for reducing misinformation reproduction in LLMs. The paper outlines design requirements (dosage, labeling, quarantine, diversity) and calls for standardized vaccine corpora and benchmarks to test generalization, making immunization a routine component of responsible LLM development.

Abstract: Large language models (LLMs) reproduce misinformation by learning the linguistic patterns that make falsehoods persuasive, such as hedging, false presuppositions, and citation fabrication, rather than merely memorizing false facts. We propose model immunization: supervised fine-tuning on curated (false claim, correction) pairs injected as small “vaccine doses” (5-10% of tokens) alongside truthful data. Unlike post-hoc filtering or preference-based alignment, immunization provides direct negative supervision on labeled falsehoods. Across four open-weight model families, immunization improves TruthfulQA accuracy by 12 points and misinformation rejection by 30 points with negligible capability loss. We outline design requirements, which includes, dosage, labeling, quarantine, diversity and call for standardized vaccine corpora and benchmarks that test generalization, making immunization a routine component of responsible LLM development

Method: Proposes using format and length as simple surrogate signals for RL training. Early training focuses on format learning using structural feedback, then incorporates length-based rewards to refine outputs by discouraging overly long or short responses. Uses GRPO (Group Relative Policy Optimization) approach with format-length signals.

Result: Achieves 40.0% accuracy on AIME2024 with a 7B base model. The method approximates and sometimes surpasses ground-truth-based optimization. Generalizes across different model sizes and series.

Conclusion: RL primarily activates reasoning capabilities already embedded in pre-trained models rather than imparting new knowledge. Lightweight, label-efficient strategies can complement pre-training to unlock LLMs’ latent potential in reasoning-intensive tasks.

Abstract: Large Language Models (LLMs) have achieved remarkable success in natural language processing tasks, with Reinforcement Learning (RL) playing a key role in adapting them to specific applications. In mathematical problem solving, however, the reliance on ground truth answers poses significant challenges due to their high collection cost and limited availability. This work explores the use of simple surrogate signals, format and length, to guide RL training. We find that early training is dominated by format learning, where structural feedback alone accounts for most performance gains. Incorporating length-based rewards further refines outputs by discouraging overly long or short responses, enabling a GRPO approach with format-length signals to approximate, and in some cases surpass, ground-truth-based optimization. For example, our method achieves 40.0% accuracy on AIME2024 with a 7B base model, and generalizes across different model sizes and series. Beyond practical efficiency, these findings provide an inspirational perspective on RL: rather than imparting new knowledge, RL primarily activates reasoning capabilities already embedded in pre-trained models. This insight suggests that lightweight, label-efficient strategies can complement pre-training to unlock LLMs’ latent potential in reasoning-intensive tasks.

Method: Documents are encoded independently using finetuned Mamba2 models, then their hidden state representations are pooled via simple operations like averaging into a single context state (document souping).

Result: Achieves competitive or superior performance on multi-hop QA, sparse retrieval, and long-document reasoning tasks compared to standard monolithic encoding, with substantial inference cost savings.

Conclusion: Document souping enables scalable, cost-effective reasoning over hundreds of documents while maintaining strong performance, unlocking new possibilities for large-scale corpus analysis.

Abstract: We investigate whether hidden states from Structured State Space Models (SSMs) can be merged post hoc to support downstream reasoning. Inspired by model souping, we study document souping, a strategy where documents are encoded independently, and their representations are pooled, via simple operations like averaging, into a single context state. This approach enables modular encoding and reuse without reprocessing the full input for each query. We demonstrate that finetuned Mamba2 models with souped representations achieve competitive or superior performance across multi-hop QA, sparse retrieval, and long-document reasoning tasks compared to the standard monolithic encoding approach. For example, on the RACE and QuALITY benchmarks for long document question answering, this method substantially outperforms a traditional concatenation approach. Crucially, this modular design scales to hundreds of documents while delivering substantial savings in inference cost, unlocking new possibilities for large-scale corpus reasoning.

Method: Systematic evaluation of 13 state-of-the-art LLMs across five Pakistani languages using culturally adapted Political Compass Test with multi-level framing analysis across 11 socio-political themes specific to Pakistani context.

Result: LLMs predominantly reflect liberal-left orientations consistent with Western training data, but exhibit more authoritarian framing in regional languages, showing language-conditioned ideological modulation and consistent model-specific bias patterns across languages.

Conclusion: Findings demonstrate need for culturally grounded, multilingual bias auditing frameworks in global NLP to address language-specific ideological biases.

Abstract: Large Language Models (LLMs) increasingly shape public discourse, yet most evaluations of political and economic bias have focused on high-resource, Western languages and contexts. This leaves critical blind spots in low-resource, multilingual regions such as Pakistan, where linguistic identity is closely tied to political, religious, and regional ideologies. We present a systematic evaluation of political bias in 13 state-of-the-art LLMs across five Pakistani languages: Urdu, Punjabi, Sindhi, Pashto, and Balochi. Our framework integrates a culturally adapted Political Compass Test (PCT) with multi-level framing analysis, capturing both ideological stance (economic/social axes) and stylistic framing (content, tone, emphasis). Prompts are aligned with 11 socio-political themes specific to the Pakistani context. Results show that while LLMs predominantly reflect liberal-left orientations consistent with Western training data, they exhibit more authoritarian framing in regional languages, highlighting language-conditioned ideological modulation. We also identify consistent model-specific bias patterns across languages. These findings show the need for culturally grounded, multilingual bias auditing frameworks in global NLP.

Method: ZOES uses a three-step mechanism: 1) Enrichment - generating initial entity structures, 2) Refinement - improving the quality of extracted structures, and 3) Unification - consolidating multiple structure views. The approach leverages the insight that entities and their associated structures are mutually reinforcing.

Result: Experiments show ZOES consistently enhances LLMs’ ability to extract more complete entity structures across three different domains, demonstrating both effectiveness and generalizability of the method.

Conclusion: The enrichment, refinement, and unification mechanism serves as a principled approach to improving LLM-based entity structure discovery quality in various scenarios, enabling zero-shot open-schema extraction without predefined schemas or annotated samples.

Abstract: Entity structure extraction, which aims to extract entities and their associated attribute-value structures from text, is an essential task for text understanding and knowledge graph construction. Existing methods based on large language models (LLMs) typically rely heavily on predefined entity attribute schemas or annotated datasets, often leading to incomplete extraction results. To address these challenges, we introduce Zero-Shot Open-schema Entity Structure Discovery (ZOES), a novel approach to entity structure extraction that does not require any schema or annotated samples. ZOES operates via a principled mechanism of enrichment, refinement, and unification, based on the insight that an entity and its associated structure are mutually reinforcing. Experiments demonstrate that ZOES consistently enhances LLMs’ ability to extract more complete entity structures across three different domains, showcasing both the effectiveness and generalizability of the method. These findings suggest that such an enrichment, refinement, and unification mechanism may serve as a principled approach to improving the quality of LLM-based entity structure discovery in various scenarios.

Method: 1) SoLT benchmark systematically rewrites reasoning datasets into diverse yet logically equivalent forms across multiple levels. 2) MenTaL method explicitly guides models to build concept-symbol mapping tables during translation to maintain consistency.

Result: Experiments show LLMs suffer from inconsistent symbol mapping under linguistic variation, leading to significant drops in reasoning accuracy. MenTaL brings clear and stable performance improvements across diverse inputs.

Conclusion: Overlooking linguistic diversity hides key weaknesses in LLM-based translators. The work offers a step toward more reliable logical reasoning in varied real-world scenarios through systematic benchmarking and consistency-enhancing methods.

Abstract: Logical reasoning with large language models (LLMs) has received growing attention. One mainstream approach translates natural language into formal logic and then applies symbolic solvers for deduction. While effective in many tasks, these LLM-based translators often fail to generate consistent symbolic representations when the same concept appears in different linguistic forms. Such inconsistencies break logical coherence and lead to solver errors. However, most existing benchmarks lack this type of linguistic variation, which frequently occurs in real-world text, leaving the problem underexplored. To address this gap, we present SoLT, a benchmark that systematically rewrites reasoning datasets into diverse yet logically equivalent forms across multiple levels. Beyond evaluation, SoLT also provides a general method to enrich any dataset with linguistic diversity while preserving both meaning and logic. To further enhance the stability of LLM-based reasoning, we propose MenTaL, which explicitly guides models to build a concept-symbol mapping table during translation. By linking equivalent expressions to shared symbols, MenTaL maintains consistency and mitigates symbol drift. Experiments on SoLT demonstrate that LLMs indeed suffer from inconsistent symbol mapping under linguistic variation, leading to significant drops in reasoning accuracy. Meanwhile, applying MenTaL brings clear and stable performance improvements across diverse inputs. Overall, our findings reveal that overlooking linguistic diversity hides key weaknesses in LLM-based translators, and our work offers a step toward more reliable logical reasoning in varied real-world scenarios. Our code is available at https://github.com/wufeiwuwoshihua/LinguDiver.

Method: The paper introduces canonical generation, a constrained generation approach that restricts LLMs to only produce canonical tokenizations (the unique tokenization used during training). They develop an efficient sampling algorithm based on the Gumbel-Max trick to implement canonical generation while maintaining performance and runtime comparable to standard sampling.

Result: Experiments across various natural language tasks show that the proposed canonical generation sampling algorithm successfully eliminates tokenization multiplicity while maintaining comparable performance and runtime to standard sampling methods. The approach effectively solves the arbitrary price variation problem in token-based pricing models.

Conclusion: Tokenization multiplicity in LLMs creates unfair pricing variations in token-based models. Canonical generation with the proposed efficient sampling algorithm provides a practical solution that maintains model performance while ensuring consistent tokenization and pricing for identical outputs.

Abstract: Providers of LLM-as-a-service have predominantly adopted a simple pricing model: users pay a fixed price per token. Consequently, one may think that the price two different users would pay for the same output string under the same input prompt is the same. In our work, we show that, surprisingly, this is not (always) true. We find empirical evidence that, particularly for non-english outputs, both proprietary and open-weights LLMs often generate the same (output) string with multiple different tokenizations, even under the same input prompt, and this in turn leads to arbitrary price variation. To address the problem of tokenization multiplicity, we introduce canonical generation, a type of constrained generation that restricts LLMs to only generate canonical tokenizations – the unique tokenization in which each string is tokenized during the training process of an LLM. Further, we introduce an efficient sampling algorithm for canonical generation based on the Gumbel-Max trick. Experiments on a variety of natural language tasks demonstrate that our sampling algorithm for canonical generation is comparable to standard sampling in terms of performance and runtime, and it solves the problem of tokenization multiplicity.

Method: Simulated structured debates between AI agents using a 6-dimensional persona space over 131 relationship-based moral cases, analyzing how personas affect initial moral stances and debate outcomes.

Result: Personas significantly affect moral stances and debate outcomes, with political ideology and personality traits exerting strongest influence. Liberal and open personalities achieve higher consensus and win rates. Logit-based confidence grows while emotional/credibility appeals diminish during debates.

Conclusion: Findings mirror psychology and cultural studies, reinforcing need for persona-aware evaluation frameworks for AI moral reasoning as LLMs are deployed in morally sensitive domains.

Abstract: As large language models (LLMs) are increasingly used in morally sensitive domains, it is crucial to understand how persona traits affect their moral reasoning and persuasive behavior. We present the first large-scale study of multi-dimensional persona effects in AI-AI debates over real-world moral dilemmas. Using a 6-dimensional persona space (age, gender, country, class, ideology, and personality), we simulate structured debates between AI agents over 131 relationship-based cases. Our results show that personas affect initial moral stances and debate outcomes, with political ideology and personality traits exerting the strongest influence. Persuasive success varies across traits, with liberal and open personalities reaching higher consensus and win rates. While logit-based confidence grows during debates, emotional and credibility-based appeals diminish, indicating more tempered argumentation over time. These trends mirror findings from psychology and cultural studies, reinforcing the need for persona-aware evaluation frameworks for AI moral reasoning.

Method: Frames training as constrained RL with: (1) token-level dense Reasoning Reflection Reward (R3) that measures model’s certainty of reference answer under its CoT reasoning prefix, emphasizing reasoning-reflective tokens; (2) rubric-gating as feasibility constraints at rollout group level, operationalizing task criteria as hard accept/reject checks on final answers.

Result: Outperforms baselines across four datasets, achieves faster and more sample-efficient learning, and successfully respects feasibility constraints.

Conclusion: The proposed constrained RL framework with R3 and rubric-gating effectively addresses challenges in training LLMs on open-ended tasks by providing dense, reasoning-aligned rewards and principled feasibility constraints.

Abstract: RL training of LLMs on open-ended tasks is challenging due to the lack of direct verifiability. In this paper, we frame such training as constrained RL that (i) optimizes a token-level dense Reasoning Reflection Reward (R3) aligned with reasoning quality, and (ii) enforces rubric-gating as feasibility constraints at the rollout group level. R3 measures the model’s token-level certainty of a reference answer under its CoT reasoning prefix while selectively emphasizing reasoning-reflective tokens to capture how likely the generated reasoning is to yield the desired answer. Rubric-gating complements R3 by operationalizing principled task criteria as hard accept/reject checks on final answers. Empirically, across four datasets, our framework outperforms baselines, achieves faster, more sample-efficient learning, and respects feasibility constraints.

Method: Uses Sparse Autoencoders (SAEs) to uncover human-interpretable features in reward model activations. Quantifies feature salience by activation differences between chosen and rejected responses, then designs targeted data poisoning and denoising strategies based on feature-level signals.

Result: SAFER can precisely degrade or enhance safety alignment with minimal data modification without sacrificing general chat performance. Applied to safety-oriented preference datasets, it enables targeted manipulation of reward model behavior.

Conclusion: The approach contributes to interpreting, auditing, and refining reward models in high-stakes LLM alignment tasks, providing tools for better control over safety alignment in RLHF systems.

Abstract: Reinforcement learning from human feedback (RLHF) is a key paradigm for aligning large language models (LLMs) with human values, yet the reward models at its core remain largely opaque. In this work, we present Sparse Autoencoder For Enhanced Reward model (\textbf{SAFER}), a novel framework for interpreting and improving reward models through mechanistic analysis. Leveraging Sparse Autoencoders (SAEs), we uncover human-interpretable features in reward model activations, enabling insight into safety-relevant decision-making. We apply SAFER to safety-oriented preference datasets and quantify the salience of individual features by activation differences between chosen and rejected responses. Using these feature-level signals, we design targeted data poisoning and denoising strategies. Experiments show that SAFER can precisely degrade or enhance safety alignment with minimal data modification, without sacrificing general chat performance. Our approach contributes to interpreting, auditing and refining reward models in high-stakes LLM alignment tasks. Our codes are available at https://github.com/xzy-101/SAFER-code. \textit{This paper discusses topics related to reward model safety and may include discussions or examples that highlight potential risks or unsafe outcomes.}

Method: PICACO optimizes a meta-instruction that navigates multiple values to better elicit LLMs’ understanding without fine-tuning. It maximizes total correlation between specified values and LLM responses, theoretically reinforcing value correlation while reducing distractive noise.

Result: Extensive experiments on five value sets show PICACO works well with both black-box and open-source LLMs, outperforms several recent strong baselines, and achieves better balance across up to 8 distinct values.

Conclusion: PICACO effectively addresses the instruction bottleneck in in-context alignment by enabling LLMs to better comprehend and balance multiple conflicting human values through optimized meta-instructions.

Abstract: In-Context Learning has shown great potential for aligning Large Language Models (LLMs) with human values, helping reduce harmful outputs and accommodate diverse preferences without costly post-training, known as In-Context Alignment (ICA). However, LLMs’ comprehension of input prompts remains agnostic, limiting ICA’s ability to address value tensions–human values are inherently pluralistic, often imposing conflicting demands, e.g., stimulation vs. tradition. Current ICA methods therefore face the Instruction Bottleneck challenge, where LLMs struggle to reconcile multiple intended values within a single prompt, leading to incomplete or biased alignment. To address this, we propose PICACO, a novel pluralistic ICA method. Without fine-tuning, PICACO optimizes a meta-instruction that navigates multiple values to better elicit LLMs’ understanding of them and improve their alignment. This is achieved by maximizing the total correlation between specified values and LLM responses, theoretically reinforcing value correlation while reducing distractive noise, resulting in effective value instructions. Extensive experiments on five value sets show that PICACO works well with both black-box and open-source LLMs, outperforms several recent strong baselines, and achieves a better balance across up to 8 distinct values.

Method: Proposes DebateCV framework with two Debaters arguing opposing stances to surface subtle errors, and a Moderator weighing evidential strength. Introduces Debate-SFT post-training framework using synthetic data to train Moderators to effectively adjudicate debates.

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

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

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

Method: Created ElectriQ benchmark with 550k+ dialogues across 6 service domains and 24 sub-scenarios, plus SEEK-RAG retrieval-augmented method that injects policy and domain knowledge during finetuning and inference.

Result: Domain-aligned 7B models with SEEK-RAG match or surpass larger models while reducing computational cost, providing auditable, regulation-aware LLM assistants for power systems.

Conclusion: ElectriQ benchmark and SEEK-RAG method enable effective deployment of LLM-based assistants for electric power marketing that support demand-side management, renewable integration, and grid resilience.

Abstract: As power systems decarbonise and digitalise, high penetrations of distributed energy resources and flexible tariffs make electric power marketing (EPM) a key interface between regulation, system operation and sustainable-energy deployment. Many utilities still rely on human agents and rule- or intent-based chatbots with fragmented knowledge bases that struggle with long, cross-scenario dialogues and fall short of requirements for compliant, verifiable and DR-ready interactions. Meanwhile, frontier large language models (LLMs) show strong conversational ability but are evaluated on generic benchmarks that underweight sector-specific terminology, regulatory reasoning and multi-turn process stability. To address this gap, we present ElectriQ, a large-scale benchmark and evaluation framework for LLMs in EPM. ElectriQ contains over 550k dialogues across six service domains and 24 sub-scenarios and defines a unified protocol that combines human ratings, automatic metrics and two compliance stress tests-Statutory Citation Correctness and Long-Dialogue Consistency. Building on ElectriQ, we propose SEEK-RAG, a retrieval-augmented method that injects policy and domain knowledge during finetuning and inference. Experiments on 13 LLMs show that domain-aligned 7B models with SEEK-RAG match or surpass much larger models while reducing computational cost, providing an auditable, regulation-aware basis for deploying LLM-based EPM assistants that support demand-side management, renewable integration and resilient grid operation.

Method: MDIR uses matrix-driven identification and reconstruction based on matrix analysis, polar decomposition, and Large Deviation Theory (LDT). It compares single pairs of weight matrices at a time without requiring model inference, making it suitable for low-resource devices.

Result: MDIR achieves perfect scores on both Area-Under-Curve (AUC) and accuracy metrics across different source models on the LeaFBench benchmark, demonstrating state-of-the-art performance in weight correspondence detection.

Conclusion: MDIR provides a statistically rigorous, resource-efficient method for detecting weight correspondences between LLMs, offering practical tools for model provenance verification and intellectual property protection in the AI community.

Abstract: We propose Matrix-Driven Identification and Reconstruction (MDIR), a SOTA large language model homology method that accurately detects weight correspondences between models and provides rigorous $p$-value estimation of the statistical significance of these correspondences. Our method does not require model inference, and allows the detection of unattributed reuse or replication of model weights even on low-resource devices as it compares only a single pair of matrices at a time. We leverage matrix analysis, polar decomposition, and Large Deviation Theory (LDT) to achieve accurate reconstruction of weight relationships between models. Notably, MDIR is the first method to achieve perfect scores on both Area-Under-Curve (AUC) and accuracy metrics across different source models on LeaFBench.

Method: BiasGym consists of two components: BiasInject (safely injects specific biases via token-based fine-tuning while keeping the model frozen) and BiasScope (leverages injected signals to identify and steer components responsible for biased behavior).

Result: The method enables consistent bias elicitation for mechanistic analysis, supports targeted debiasing without degrading downstream task performance, and generalizes to biases unseen during fine-tuning. Demonstrated effectiveness in reducing real-world stereotypes.

Conclusion: BiasGym provides a practical framework for both safety interventions and interpretability research in LLMs, offering a systematic approach to bias analysis and mitigation.

Abstract: Understanding biases and stereotypes encoded in the weights of Large Language Models (LLMs) is crucial for developing effective mitigation strategies. However, biased behaviour is often subtle and non-trivial to isolate, even when deliberately elicited, making systematic analysis and debiasing particularly challenging. To address this, we introduce \texttt{BiasGym}, a simple, cost-effective, and generalizable framework for reliably and safely injecting, analyzing, and mitigating conceptual associations of biases within LLMs. \texttt{BiasGym} consists of two components: \texttt{BiasInject}, which safely injects specific biases into the model via token-based fine-tuning while keeping the model frozen, and \texttt{BiasScope}, which leverages these injected signals to identify and reliably steer the components responsible for biased behavior. Our method enables consistent bias elicitation for mechanistic analysis, supports targeted debiasing without degrading performance on downstream tasks, and generalizes to biases unseen during fine-tuning. We demonstrate the effectiveness of BiasGym in reducing real-world stereotypes (e.g., people from Italy being `reckless drivers’), showing its utility for both safety interventions and interpretability research.

Method: Developed a fully automated counterfactual evaluation framework that isolates and tests the skill of detecting research logic flaws under controlled conditions, testing a range of ARG approaches.

Result: Contrary to expectations, flaws in research logic had no significant effect on ARG output reviews, indicating current LLM-based review systems fail to detect these critical issues.

Conclusion: LLMs used as automatic review generators lack the ability to detect research logic flaws, posing risks to scientific integrity; three actionable recommendations are provided along with public release of the evaluation framework and dataset.

Abstract: Large Language Models (LLMs) have great potential to accelerate and support scholarly peer review and are increasingly used as fully automatic review generators (ARGs). However, potential biases and systematic errors may pose significant risks to scientific integrity; understanding the specific capabilities and limitations of state-of-the-art ARGs is essential. We focus on a core reviewing skill that underpins high-quality peer review: detecting faulty research logic. This involves evaluating the internal consistency between a paper’s results, interpretations, and claims. We present a fully automated counterfactual evaluation framework that isolates and tests this skill under controlled conditions. Testing a range of ARG approaches, we find that, contrary to expectation, flaws in research logic have no significant effect on their output reviews. Based on our findings, we derive three actionable recommendations for future work and release our counterfactual dataset and evaluation framework publicly.

Method: Introduces Atom Theory with atomic inner product (AIP) metric to define atoms, proposes faithfulness (R²) and stability (q*) criteria, proves atom identifiability under threshold-activated sparse autoencoders (TSAEs), and conducts large-scale experiments on Gemma2-2B, Gemma2-9B, and Llama3.1-8B.

Result: Neurons are faithful (R²=1) but unstable (q*=0.5%), features are stable (q*=68.2%) but unfaithful (R²=48.8%). TSAE-derived atoms achieve near-perfect faithfulness (R²=99.9%) and stability (q*=99.8%) when capacity matches data scale, showing substantially higher monosemanticity.

Conclusion: Atom Theory provides a foundation for understanding LLM internal representations by defining and identifying fundamental representational units (atoms) that satisfy both faithfulness and stability criteria, validated across multiple LLM architectures.

Abstract: The fundamental representational units (FRUs) of large language models (LLMs) remain undefined, limiting further understanding of their underlying mechanisms. In this paper, we introduce Atom Theory to systematically define, evaluate, and identify such FRUs, which we term atoms. Building on the atomic inner product (AIP), a non-Euclidean metric that captures the underlying geometry of LLM representations, we formally define atoms and propose two key criteria for ideal atoms: faithfulness ($R^2$) and stability ($q^$). We further prove that atoms are identifiable under threshold-activated sparse autoencoders (TSAEs). Empirically, we uncover a pervasive representation shift in LLMs and demonstrate that the AIP corrects this shift to capture the underlying representational geometry, thereby grounding Atom Theory. We find that two widely used units, neurons and features, fail to qualify as ideal atoms: neurons are faithful ($R^2=1$) but unstable ($q^=0.5%$), while features are more stable ($q^=68.2%$) but unfaithful ($R^2=48.8%$). To find atoms of LLMs, leveraging atom identifiability under TSAEs, we show via large-scale experiments that reliable atom identification occurs only when the TSAE capacity matches the data scale. Guided by this insight, we identify FRUs with near-perfect faithfulness ($R^2=99.9%$) and stability ($q^=99.8%$) across layers of Gemma2-2B, Gemma2-9B, and Llama3.1-8B, satisfying the criteria of ideal atoms statistically. Further analysis confirms that these atoms align with theoretical expectations and exhibit substantially higher monosemanticity. Overall, we propose and validate Atom Theory as a foundation for understanding the internal representations of LLMs. Code available at https://github.com/ChenhuiHu/towards_atoms.

Method: Proposes SimulSense framework that continuously reads input speech and triggers write decisions when new sense units are perceived, avoiding the need for specialized interleaved training data and expensive LLM inference used by current state-of-the-art systems.

Result: Experiments show superior quality-latency tradeoff compared to two state-of-the-art baseline systems, with decision-making up to 9.6x faster than baselines and substantially improved real-time efficiency.

Conclusion: SimulSense provides an effective framework for simultaneous speech translation that mimics human interpreter behavior while being more efficient than current LLM-based approaches.

Abstract: How to make human-interpreter-like read/write decisions for simultaneous speech translation (SimulST) systems? Current state-of-the-art systems formulate SimulST as a multi-turn dialogue task, requiring specialized interleaved training data and relying on computationally expensive large language model (LLM) inference for decision-making. In this paper, we propose SimulSense, a novel framework for SimulST that mimics human interpreters by continuously reading input speech and triggering write decisions to produce translation when a new sense unit is perceived. Experiments against two state-of-the-art baseline systems demonstrate that our proposed method achieves a superior quality-latency tradeoff and substantially improved real-time efficiency, where its decision-making is up to 9.6x faster than the baselines.

Method: Introduces ChatInject attack that formats malicious payloads to mimic native chat templates, exploiting models’ instruction-following tendencies. Develops a persuasion-driven Multi-turn variant that primes agents across conversational turns to accept and execute suspicious actions.

Result: ChatInject achieves significantly higher attack success rates than traditional methods (5.18% to 32.05% on AgentDojo, 15.13% to 45.90% on InjecAgent). Multi-turn dialogues show particularly strong performance (average 52.33% success rate on InjecAgent). Chat-template-based payloads demonstrate strong transferability across models and remain effective against closed-source LLMs. Existing defenses are largely ineffective.

Conclusion: The research highlights critical vulnerabilities in current agent systems, showing that structured chat template manipulation and multi-turn persuasion attacks are highly effective and transferable across models, with existing defenses being inadequate against these sophisticated attack approaches.

Abstract: The growing deployment of large language model (LLM) based agents that interact with external environments has created new attack surfaces for adversarial manipulation. One major threat is indirect prompt injection, where attackers embed malicious instructions in external environment output, causing agents to interpret and execute them as if they were legitimate prompts. While previous research has focused primarily on plain-text injection attacks, we find a significant yet underexplored vulnerability: LLMs’ dependence on structured chat templates and their susceptibility to contextual manipulation through persuasive multi-turn dialogues. To this end, we introduce ChatInject, an attack that formats malicious payloads to mimic native chat templates, thereby exploiting the model’s inherent instruction-following tendencies. Building on this foundation, we develop a persuasion-driven Multi-turn variant that primes the agent across conversational turns to accept and execute otherwise suspicious actions. Through comprehensive experiments across frontier LLMs, we demonstrate three critical findings: (1) ChatInject achieves significantly higher average attack success rates than traditional prompt injection methods, improving from 5.18% to 32.05% on AgentDojo and from 15.13% to 45.90% on InjecAgent, with multi-turn dialogues showing particularly strong performance at average 52.33% success rate on InjecAgent, (2) chat-template-based payloads demonstrate strong transferability across models and remain effective even against closed-source LLMs, despite their unknown template structures, and (3) existing prompt-based defenses are largely ineffective against this attack approach, especially against Multi-turn variants. These findings highlight vulnerabilities in current agent systems.

Method: Mechanistic analysis using value vectors (feature directions from residual stream) and value neurons (MLP neurons contributing to value vectors), examining generalization across languages and theoretical inter-value correlations

Result: Intrinsic and prompted value mechanisms share common components crucial for value expression but have unique elements: intrinsic promotes lexical diversity, prompted strengthens instruction following and affects distant tasks like jailbreaking

Conclusion: Value expression in LLMs involves both shared and distinct mechanisms, with intrinsic mechanisms favoring diversity and prompted mechanisms enabling better steerability, important for understanding value alignment

Abstract: Large language models can express values in two main ways: (1) intrinsic expression, reflecting the model’s inherent values learned during training, and (2) prompted expression, elicited by explicit prompts. Given their widespread use in value alignment, it is paramount to clearly understand their underlying mechanisms, particularly whether they mostly overlap (as one might expect) or rely on distinct mechanisms, but this remains largely understudied. We analyze this at the mechanistic level using two approaches: (1) value vectors, feature directions representing value mechanisms extracted from the residual stream, and (2) value neurons, MLP neurons that contribute to value vectors. We demonstrate that intrinsic and prompted value mechanisms partly share common components crucial for inducing value expression, generalizing across languages and reconstructing theoretical inter-value correlations in the model’s internal representations. Yet, as these mechanisms also possess unique elements that fulfill distinct roles, they lead to different degrees of response diversity (intrinsic > prompted) and value steerability (prompted > intrinsic). In particular, components unique to the intrinsic mechanism promote lexical diversity in responses, whereas those specific to the prompted mechanism strengthen instruction following, taking effect even in distant tasks like jailbreaking.

Method: Multi-stage retrieval-augmented pipeline with retriever, base generator, and refinement model using domain-specific knowledge graphs; employs RAGAS-based scoring for quality filtering

Result: Successfully generates complex, context-rich troubleshooting solution plans for telecom RAN without human intervention, producing high-quality dataset for reinforcement fine-tuning

Conclusion: Scalable solution for building instruction and reinforcement datasets in specialized domains, significantly reducing manual labeling while maintaining technical fidelity

Abstract: The success of large language models (LLMs) depends heavily on large-scale, high-quality instruction-following and reinforcement datasets. However, generating such data through human annotation is prohibitively time-consuming particularly for domain-specific tasks like telecom network troubleshooting, where accurate responses require deep technical expertise and contextual understanding. In this paper, we present a fully automated, retrieval-augmented pipeline for generating synthetic question-answer (QA) pairs grounded in structured domain knowledge. Our multi-stage framework integrates a retriever, base generator, and refinement model to synthesize and enhance QA pairs using documents retrieved from a domain-specific knowledge graph. To ensure data quality, we employ customized RAGAS-based scoring to filter low-quality samples, producing a high-quality dataset suitable for reinforcement fine-tuning (RFT). We demonstrate our approach in a real-world telecom scenario focused on radio access network (RAN) troubleshooting. The resulting pipeline generates complex, context-rich troubleshooting solution plans without human intervention. This work offers a scalable solution for building instruction and reinforcement datasets in specialized domains, significantly reducing dependence on manual labeling while maintaining high technical fidelity.

Method: Investigates internal behavior of backdoored encoder-based language models, focusing on consistent shifts in attention and gradient attribution when processing poisoned inputs. Proposes an inference-time defense that constructs anomaly scores by combining token-level attention and gradient information.

Result: Extensive experiments on text classification tasks across diverse backdoor attack scenarios show the method significantly reduces attack success rates compared to existing baselines. Provides interpretability-driven analysis of scoring mechanism for trigger localization.

Conclusion: The proposed inference-time defense effectively mitigates backdoor attacks by leveraging attention and gradient signals, offering both practical protection and insights into attack mechanisms through interpretable analysis.

Abstract: Pre-trained language models have achieved remarkable success across a wide range of natural language processing (NLP) tasks, particularly when fine-tuned on large, domain-relevant datasets. However, they remain vulnerable to backdoor attacks, where adversaries embed malicious behaviors using trigger patterns in the training data. These triggers remain dormant during normal usage, but, when activated, can cause targeted misclassifications. In this work, we investigate the internal behavior of backdoored pre-trained encoder-based language models, focusing on the consistent shift in attention and gradient attribution when processing poisoned inputs; where the trigger token dominates both attention and gradient signals, overriding the surrounding context. We propose an inference-time defense that constructs anomaly scores by combining token-level attention and gradient information. Extensive experiments on text classification tasks across diverse backdoor attack scenarios demonstrate that our method significantly reduces attack success rates compared to existing baselines. Furthermore, we provide an interpretability-driven analysis of the scoring mechanism, shedding light on trigger localization and the robustness of the proposed defense.

Method: Proposes a framework to measure three aspects of data contamination: (1) item memorization, (2) evaluation memorization, and (3) target score matching. Applied this framework to 21 models from major families and four widely used psychometric inventories.

Result: Found that popular inventories like the Big Five Inventory (BFI-44) and Portrait Values Questionnaire (PVQ-40) exhibit strong contamination, where models not only memorize items but can also adjust their responses to achieve specific target scores.

Conclusion: Data contamination in psychometric evaluations of LLMs is a significant issue that threatens the reliability of such assessments, requiring more rigorous evaluation methodologies and awareness of this limitation in psychological studies of LLMs.

Abstract: Recent studies apply psychometric questionnaires to Large Language Models (LLMs) to assess high-level psychological constructs such as values, personality, moral foundations, and dark traits. Although prior work has raised concerns about possible data contamination from psychometric inventories, which may threaten the reliability of such evaluations, there has been no systematic attempt to quantify the extent of this contamination. To address this gap, we propose a framework to systematically measure data contamination in psychometric evaluations of LLMs, evaluating three aspects: (1) item memorization, (2) evaluation memorization, and (3) target score matching. Applying this framework to 21 models from major families and four widely used psychometric inventories, we provide evidence that popular inventories such as the Big Five Inventory (BFI-44) and Portrait Values Questionnaire (PVQ-40) exhibit strong contamination, where models not only memorize items but can also adjust their responses to achieve specific target scores.

Method: The authors analyze overlapping components encoding both hallucination and refusal information. They propose using sparse autoencoders to disentangle refusal-related features from hallucination features, and preserve refusal behavior during fine-tuning through subspace orthogonalization.

Result: The method is evaluated on commonsense reasoning tasks and harmful benchmarks (AdvBench and StrongReject). Results show the approach successfully preserves refusal behavior and task utility while mitigating hallucinations.

Conclusion: There exists a fundamental trade-off between truthfulness and safety in LLMs, but this can be mitigated through careful feature disentanglement and orthogonalization techniques that preserve both factual accuracy and safety alignment.

Abstract: Hallucination in large language models (LLMs) has been widely studied in recent years, with progress in both detection and mitigation aimed at improving truthfulness. Yet, a critical side effect remains largely overlooked: enhancing truthfulness can negatively impact safety alignment. In this paper, we investigate this trade-off and show that increasing factual accuracy often comes at the cost of weakened refusal behavior. Our analysis reveals that this arises from overlapping components in the model that simultaneously encode hallucination and refusal information, leading alignment methods to suppress factual knowledge unintentionally. We further examine how fine-tuning on benign datasets, even when curated for safety, can degrade alignment for the same reason. To address this, we propose a method that disentangles refusal-related features from hallucination features using sparse autoencoders, and preserves refusal behavior during fine-tuning through subspace orthogonalization. This approach prevents hallucinations from increasing while maintaining safety alignment.We evaluate our method on commonsense reasoning tasks and harmful benchmarks (AdvBench and StrongReject). Results demonstrate that our approach preserves refusal behavior and task utility, mitigating the trade-off between truthfulness and safety.

Method: RLKV uses reinforcement learning as a probe to discover which attention heads contribute to reasoning quality by directly optimizing their cache usage against actual generation outcomes. This discovery leads to an efficient compression strategy: allocate full KV cache to reasoning-critical heads while aggressively compressing others.

Result: Experiments show that only a fraction of heads is essential for reasoning, enabling 20-50% cache reduction with near-lossless performance and up to 1.21x speedup.

Conclusion: RLKV provides an effective approach for KV cache compression in reasoning LLMs by identifying and preserving reasoning-critical attention heads through reinforcement learning, achieving significant compression with minimal performance degradation.

Abstract: Reasoning large language models exhibit complex reasoning behaviors via extended chain-of-thought generation that are highly fragile to information loss during decoding, creating critical challenges for KV cache compression. Existing token-dropping methods directly disrupt reasoning chains by removing intermediate steps, while head-reallocation methods, designed for retrieval tasks, fail to preserve the heads essential for generative reasoning. However, no existing method can identify which attention heads genuinely maintain reasoning consistency and control generation termination. To address this, we propose RLKV, which uses reinforcement learning as a probe to discover which heads contribute to reasoning quality by directly optimizing their cache usage against actual generation outcomes. This discovery naturally leads to an efficient compression strategy: we allocate full KV cache to reasoning-critical heads while aggressively compressing others. Experiments reveal that a fraction of heads proves essential for reasoning, enabling 20–50% cache reduction with near-lossless performance and up to 1.21x speedup.

Method: Proposes GraphGhost framework that models token interactions and neuron activations as graphs. Uses two complementary views: sample view (traces token dependencies for individual predictions) and dataset view (aggregates recurring structural patterns learned during training). Analyzes graph structural properties to identify influential tokens and neuron nodes.

Result: Graph structural properties are closely associated with influential tokens and neuron nodes. Perturbations to structurally critical nodes lead to measurable changes in reasoning behavior, indicating that captured structural patterns reflect meaningful internal organization of LLM reasoning.

Conclusion: GraphGhost provides a novel graph-based approach to understand LLM reasoning mechanisms, revealing that structural patterns in token interactions and neuron activations correspond to meaningful internal organization of reasoning processes.

Abstract: Large Language Models (LLMs) exhibit strong reasoning capabilities on structured tasks, yet the internal mechanisms underlying such behaviors remain poorly understood. Existing interpretation methods mainly focus on token-level attributions, which provide limited insight into multi-step reasoning inside the model. We propose GraphGhost, a graph-based framework that models internal token interactions and neuron activations in LLMs as graphs. By aggregating token dependencies traced across layers, GraphGhost captures global information flow underlying model predictions. We formalize GraphGhost from two complementary perspectives: a sample view, which traces token dependencies for individual predictions, and a dataset view, which aggregates recurring structural patterns learned during training. Through graph analytics and quantitative experiments, we show that graph structural properties are closely associated with influential tokens and neuron nodes, and that perturbations to structurally critical nodes lead to measurable changes in reasoning behavior. These results indicate that the structural patterns captured by GraphGhost reflect meaningful internal organization of LLM reasoning. The codes are available at software part. Artifacts will be made available for research use only.

Method: Black-box Bayesian framework using: (1) signal anchoring - reordering documents to place high-belief items in diagnostic positions, (2) iterative belief updating from model outputs, and (3) O(log N) round complexity for N documents.

Result: Matches Permutation Self-Consistency’s target identification with 30-65% fewer queries on needle-in-haystack retrieval and long-context QA. Remains effective under calibration mismatch.

Conclusion: Inherent model biases can be leveraged as tools for control rather than failures. Positional ordering drives performance gains in long-context problems.

Abstract: Large language models (LLMs) exhibit pronounced position bias in long-context needle-in-haystack problems, systematically prioritizing the location of information over its relevance. While current mitigations rely on white-box access, this is effectively impossible for many state-of-the-art models. We introduce GOLD PANNING, a black-box Bayesian framework that performs inference-time active search over long contexts by (i) reordering documents to concentrate high-belief items in highly diagnostic positions (signal anchoring) and (ii) updating beliefs over document relevance from model outputs. Unlike conventional active learning, which prioritizes uncertainty reduction, GOLD PANNING leverages anchoring – once flagged, keep it in sight – to preserve weak cues. We implement this using iterative assignment derived from the model’s diagnosticity profile, which provably identifies a target among $N$ documents in $O(\log N)$ rounds, ensuring scalability to many-document settings.On needle-in-a-haystack retrieval and long-context QA, GOLD PANNING matches Permutation Self-Consistency’s target identification with $30–65%$ fewer queries and remains effective under calibration mismatch, suggesting coarse positional ordering drives performance gains. These results demonstrate that inherent model biases need not be failures, but can be used as tools for control.

Method: Created DUAL-Bench, the first multimodal benchmark focused on over-refusal and safe completion in VLMs. Evaluated 18 VLMs across 12 hazard categories, with emphasis on robustness under semantics-preserving visual perturbations.

Result: Results show substantial room for improvement: GPT-5-Nano achieves 12.9% safe completion, GPT-5 models average 7.9%, and Qwen models only 3.9%. Models frequently fail in dual-use scenarios, either refusing too conservatively or completing tasks unsafely.

Conclusion: DUAL-Bench highlights the need for more nuanced alignment strategies in VLMs to ensure models remain both safe and useful in complex multimodal settings, addressing the unique challenges of visual modality safety.

Abstract: As vision-language models become increasingly capable, maintaining a balance between safety and usefulness remains a central challenge. Safety mechanisms, while essential, can backfire, causing over-refusal, where models decline benign requests out of excessive caution. Yet, no existing benchmark has systematically addressed over-refusal in the visual modality. This setting introduces unique challenges, such as dual-use cases where an instruction is harmless, but the accompanying image contains harmful content. Models frequently fail in such scenarios, either refusing too conservatively or completing tasks unsafely, which highlights the need for more fine-grained alignment. The ideal behavior is safe completion, i.e., fulfilling the benign parts of a request while explicitly warning about any potentially harmful elements. To address this, we present DUAL-Bench, the first multimodal benchmark focused on over-refusal and safe completion in VLMs. We evaluated 18 VLMs across 12 hazard categories, with focus on their robustness under semantics-preserving visual perturbations. The results reveal substantial room for improvement: GPT-5-Nano achieves 12.9% safe completion, GPT-5 models average 7.9%, and Qwen models only 3.9%. We hope that DUAL-Bench will foster the development of more nuanced alignment strategies that ensure models remain both safe and useful in complex multimodal settings.

Method: 1) Created a large-scale benchmark with 4200+ verified prompts across 6 English dialects; 2) Evaluated 17 image/video generative models; 3) Proposed an encoder-based mitigation strategy that teaches models to recognize dialect features while preserving SAE performance; 4) Tested on models like Stable Diffusion 1.5.

Result: Current models show 32.26-48.17% performance degradation with dialect prompts. Existing methods (fine-tuning, prompt rewriting) improve dialect performance by <7% but harm SAE performance. The proposed encoder-based method raises performance on 5 dialects to match SAE (+34.4%) with near-zero cost to SAE performance.

Conclusion: Multimodal generative models have significant dialect bias, but encoder-based adaptation can effectively address this while maintaining performance on standard language. This work highlights the need for more inclusive multimodal AI systems.

Abstract: Contact languages like English exhibit rich regional variations in the form of dialects, which are often used by dialect speakers interacting with generative models. However, can multimodal generative models effectively produce content given dialectal textual input? In this work, we study this question by constructing a new large-scale benchmark spanning six common English dialects. We work with dialect speakers to collect and verify over 4200 unique prompts and evaluate on 17 image and video generative models. Our automatic and human evaluation results show that current state-of-the-art multimodal generative models exhibit 32.26% to 48.17% performance degradation when a single dialect word is used in the prompt. Common mitigation methods such as fine-tuning and prompt rewriting can only improve dialect performance by small margins (< 7%), while potentially incurring significant performance degradation in Standard American English (SAE). To this end, we design a general encoder-based mitigation strategy for multimodal generative models. Our method teaches the model to recognize new dialect features while preserving SAE performance. Experiments on models such as Stable Diffusion 1.5 show that our method is able to simultaneously raise performance on five dialects to be on par with SAE (+34.4%), while incurring near zero cost to SAE performance.

Method: Three-level framework: 1) Latent-Vocab constrains hidden states within pre-trained vocab-space, 2) Latent-Chain via Induction-Supervision Masking ensures semantic compactness and sufficiency, 3) Latent-Optim uses stochastic Gumbel-Softmax to guide models toward generalizable solutions.

Result: Latent-SFT consistently outperforms explicit SFT across six mathematical benchmarks (GSM8k, AIME24, etc.) while achieving 2.7x to 5.5x reduction in reasoning length. Analysis confirms the method captures superposition of diverse reasoning trajectories rather than just compressing a single path.

Conclusion: The proposed Latent-SFT framework successfully addresses key challenges in latent reasoning, enabling efficient computation while maintaining or improving performance through superposition of reasoning paths, making it a promising approach for efficient reasoning in language models.

Abstract: Latent reasoning offers a computation-efficient alternative to Chain-of-Thought but often suffers from performance degradation due to distributional misalignment and ambiguous chain definitions. Ideally, latent reasoning should function as a superposition of multiple reasoning paths. To realize this, we introduce Latent-SFT, a unified framework addressing challenges at three levels: token, chain, and learning. First, we define the Latent-Vocab to constrain hidden states within the pre-trained vocab-space. Second, we construct the Latent-Chain via Induction-Supervision Masking to ensure semantic compactness and sufficiency. Third, we employ Latent-Optim with stochastic Gumbel-Softmax to guide the model toward generalizable solutions. Empirical results demonstrate that Latent-SFT consistently outperforms explicit SFT across six mathematical benchmarks (e.g., GSM8k, AIME24) while achieving a 2.7x to 5.5x reduction in reasoning length. Analysis confirms that our method effectively captures a superposition of diverse reasoning trajectories rather than merely compressing a single path.

Method: Proposes dynamic, reversible pruning-based framework that detects context-aware neuron activations and applies adaptive masking to modulate their influence during generation. Operates at inference time with fine-grained, memory-aware mitigation.

Result: Achieves knowledge-preserved, more coherent behavior across multilingual single- and multi-turn dialogues, enabling dynamic fairness control in real-world conversational AI.

Conclusion: Provides a flexible inference-time solution for mitigating LLM biases that adapts to changing contexts while preserving model capabilities, offering advantages over static approaches.

Abstract: Large language models often display undesirable behaviors embedded in their internal representations, undermining fairness, inconsistency drift, amplification of harmful content, and the propagation of unwanted patterns during extended dialogue and conversations. Although training-time or data-centric methods attempt to reduce these effects, they are computationally expensive, irreversible once deployed, and slow to adapt to new conversational contexts. Pruning-based methods provide a flexible and transparent way to reduce bias by adjusting the neurons responsible for certain behaviors. However, most existing approaches are static; once a neuron is removed, the model loses the ability to adapt when the conversation or context changes. To address this, we propose a dynamic, reversible, pruning-based framework that detects context-aware neuron activations and applies adaptive masking to modulate their influence during generation. Our inference-time solution provides fine-grained, memory-aware mitigation with knowledge-preserved, more coherent behavior across multilingual single- and multi-turn dialogues, enabling dynamic fairness control in real-world conversational AI.

Method: Analyzed the geometry of LLM hidden-state spaces to identify emotional manifolds, examined directional encoding and distribution across layers, tested generalization across eight real-world emotion datasets in five languages, and developed an intervention module to steer internal emotion perception.

Result: Found a low-dimensional emotional manifold that is directionally encoded, distributed across layers, stable across depth, and generalizes well across languages and datasets. Cross-domain alignment showed low error and strong linear probe performance. The intervention module successfully steered emotion perception while preserving semantics, especially for basic emotions across languages.

Conclusion: LLMs possess a consistent and manipulable affective geometry that reveals how they internalize and process emotion, offering potential for emotion-aware language generation and understanding.

Abstract: This work investigates how large language models (LLMs) internally represent emotion by analyzing the geometry of their hidden-state space. The paper identifies a low-dimensional emotional manifold and shows that emotional representations are directionally encoded, distributed across layers, and aligned with interpretable dimensions. These structures are stable across depth and generalize to eight real-world emotion datasets spanning five languages. Cross-domain alignment yields low error and strong linear probe performance, indicating a universal emotional subspace. Within this space, internal emotion perception can be steered while preserving semantics using a learned intervention module, with especially strong control for basic emotions across languages. These findings reveal a consistent and manipulable affective geometry in LLMs and offer insight into how they internalize and process emotion.

Method: (1) Formalize synchronization invariants for valid batch speculative decoding, (2) present EQSPEC algorithm that guarantees output equivalence, and (3) introduce EXSPEC which reduces alignment overhead through cross-batch scheduling that dynamically groups same-length sequences.

Result: Achieves up to 3x throughput improvement at batch size 8 while maintaining algorithmic correctness, with 95% decoding-equivalence (residual divergence due to floating-point non-determinism, not synchronization failures).

Conclusion: First framework that solves the fundamental ragged tensor problem in batch speculative decoding, enabling correct and efficient batch processing while maintaining output equivalence with standard autoregressive generation.

Abstract: Speculative decoding must produce outputs distribution identical to standard autoregressive generation-this output equivalence is not an optimization target but the defining criterion of valid speculative decoding. We demonstrate that all existing batch speculative decoding implementations violate this fundamental requirement, producing corrupted outputs ranging from repetitive tokens to gibberish. These failures stem from the ragged tensor problem: sequences in the same batch accept different numbers of draft tokens, desynchronizing position IDs, attention masks, and KV-cache state. We present the first authentic batch speculative decoding framework. We (1) formalize the synchronization invariants that valid batch speculative decoding must satisfy, (2) present EQSPEC, the first algorithm that guarantees output equivalence, and analyze its cost structure to show that alignment overhead grows superlinearly and consumes up to 40% of computation, and (3) introduce EXSPEC, which reduces this overhead through cross-batch scheduling that dynamically groups same-length sequences. On SpecBench across Vicuna-7B/68M, Qwen3-8B/0.6B, and GLM-4-9B/0.6B pairs, our methods achieve up to 3x throughput improvement at batch size 8 while maintaining algorithmic correctness. Our methods achieve 95% decoding-equivalence, with residual divergence attributable to floating-point non-determinism in GPU inference, not the synchronization failures that cause near-zero equivalence of prior methods. Our code is available at https://github.com/eBay/spec_dec.

Method: Introduces rank-2 projection subspace to more accurately disentangle parametric knowledge (PK) and contextual knowledge (CK) contributions, enabling first multi-step analysis of knowledge interactions across longer NLE sequences.

Result: Experiments across 4 QA datasets and 3 open-weight LLMs show rank-2 formulation effectively captures diverse knowledge interactions, with PK alignment for supportive interactions and CK alignment for conflicting ones. Multi-step analysis reveals hallucinated generations strongly align with PK direction while context-faithful generations maintain balanced PK-CK alignment.

Conclusion: The rank-2 approach provides more accurate modeling of knowledge interactions in LLM explanations, revealing important patterns in multi-step generation and offering insights into hallucination mechanisms through knowledge source alignment analysis.

Abstract: Natural Language Explanations (NLEs) describe how Large Language Models (LLMs) make decisions by drawing on external Context Knowledge (CK) and Parametric Knowledge (PK). Understanding the interaction between these sources is key to assessing NLE grounding, yet these dynamics remain underexplored. Prior work has largely focused on (1) single-step generation and (2) modelled PK-CK interaction as a binary choice within a rank-1 subspace. This approach overlooks richer interactions and how they unfold over longer generations, such as complementary or supportive knowledge. We propose a novel rank-2 projection subspace that disentangles PK and CK contributions more accurately and use it for the first multi-step analysis of knowledge interactions across longer NLE sequences. Experiments across four QA datasets and three open-weight LLMs demonstrate that while rank-1 subspaces struggle to represent diverse interactions, our rank-2 formulation captures them effectively, highlighting PK alignment for supportive interactions and CK alignment for conflicting ones. Our multi-step analysis reveals, among others, that hallucinated generations exhibit strong alignment with the PK direction, whereas context-faithful generations maintain a more balanced alignment between PK and CK.

Method: Proposes Manifold-Correcting Causal Flow (MC^2F) with two modules: 1) Stratified Riemannian Continuous Normalizing Flow (SR-CNF) learns the density of clean data manifold and identifies out-of-distribution embeddings, and 2) Geodesic Purification Solver projects adversarial points back onto the learned manifold via shortest path geodesic projection to restore clean, semantically coherent representations.

Result: Extensive evaluations on text classification across three datasets and multiple adversarial attacks show MC^2F establishes new state-of-the-art in adversarial robustness while fully preserving performance on clean data, even yielding modest accuracy gains.

Conclusion: MC^2F successfully resolves the robustness-clean performance trade-off in text classification by modeling clean data manifold distribution and correcting adversarial embeddings through geodesic projection, achieving both superior robustness and maintained clean data performance.

Abstract: A persistent challenge in text classification (TC) is that enhancing model robustness against adversarial attacks typically degrades performance on clean data. We argue that this challenge can be resolved by modeling the distribution of clean samples in the encoder embedding manifold. To this end, we propose the Manifold-Correcting Causal Flow (MC^2F), a two-module system that operates directly on sentence embeddings. A Stratified Riemannian Continuous Normalizing Flow (SR-CNF) learns the density of the clean data manifold. It identifies out-of-distribution embeddings, which are then corrected by a Geodesic Purification Solver. This solver projects adversarial points back onto the learned manifold via the shortest path, restoring a clean, semantically coherent representation. We conducted extensive evaluations on text classification (TC) across three datasets and multiple adversarial attacks. The results demonstrate that our method, MC^2F, not only establishes a new state-of-the-art in adversarial robustness but also fully preserves performance on clean data, even yielding modest gains in accuracy.

Method: Creates a dynamic evaluation environment with six elimination-style levels focusing on multi-faceted abilities. Evaluates over 50 LLMs through interactive gameplay against other LLM opponents in resource-constrained, asymmetric information settings.

Result: Shows clear generational phase transitions in performance within model lineages, finds evidence of models using speculative shortcuts, and demonstrates that dynamic evaluation complements static benchmarks by revealing different behavioral patterns.

Conclusion: Dynamic adversarial evaluation provides valuable complementary insights to static benchmarks, revealing LLM behaviors under pressure and potential contamination issues in traditional evaluation paradigms.

Abstract: The potential data contamination issue in contemporary large language models (LLMs) benchmarks presents a fundamental challenge to establishing trustworthy evaluation frameworks. Meanwhile, they predominantly assume benign, resource-rich settings, leaving the behavior of LLMs under pressure unexplored. In this paper, we introduce \textsc{Squid Game}, a dynamic and adversarial evaluation environment with resource-constrained and asymmetric information settings elaborated to evaluate LLMs through interactive gameplay against other LLM opponents. Squid Game consists of six elimination-style levels, focusing on multi-faceted abilities, including instruction-following, code, reasoning, planning, and safety alignment. We evaluate over 50 LLMs on Squid Game, presenting the largest behavioral evaluation study of general LLMs on dynamic adversarial scenarios. We observe a clear generational phase transition in performance in the same model lineage and find evidence that some models resort to speculative shortcuts to win the game, indicating the possibility of higher-level evaluation paradigm contamination in static benchmarks. We also compare prominent LLM benchmarks and \textsc{Squid Game}, highlighting that dynamic evaluation can serve as a complementary part for static evaluations. Project page: https://github.com/zijianchen98/LLM_Squid_Game.

Method: Train models on different multilingual mixtures, analyze internal mechanisms using Cross-Layer Transcoders (CLTs) and Attribution Graphs, perform Model-Diffing experiments, and test interventions on language-identity features.

Result: Multilingual LLMs use highly similar features across languages (shared representations) with language-specific decoding in later layers; non-English failures arise from dim late-layer features, weak middle-layer clusters, and tokenizer bias toward English; finetuning improves token assembly and language-specific decoding.

Conclusion: Multilingual LLMs employ shared representations with language-specific decoding; performance gaps are mechanistic issues (weak features, tokenizer bias) that can be addressed through targeted interventions and finetuning.

Abstract: Multilingual Large Language Models (LLMs) can process many languages, yet how they internally represent this diversity remains unclear. Do they form shared multilingual representations with language-specific decoding, and if so, why does performance favor the dominant training language? To address this, we train models on different multilingual mixtures and analyze their internal mechanisms using Cross-Layer Transcoders (CLTs) and Attribution Graphs. Our results reveal multilingual shared representations: the model employs highly similar features across languages, while language-specific decoding emerges in later layers. Training models without English shows identical multilingual shared space structures. Decoding relies partly on a small set of high-frequency features in the final layers, which linearly encode language identity from early layers. Intervening on these features allows one language to be suppressed and another substituted. Finally, to explain non-English failures, we perform a Model-Diffing experiment: underperformance arises from dim late-layer features, weak middle-layer clusters, and tokenizer bias toward English that forces early layers to specialize in word reassembly. Finetuning strengthens these features and their links, improving token assembly and language-specific decoding, providing a mechanistic explanation for multilingual gaps. Our models and CLTs are available at https://huggingface.co/collections/CausalNLP/multilingual-clts and https://huggingface.co/collections/CausalNLP/multilingual-gpt2-models. Our code is available at: https://github.com/abirharrasse/MultilingualCLTs

Method: Used data from an online debate platform to evaluate off-the-shelf open-weight LLMs’ ability to predict individuals’ stances under four conditions: no context, demographics only, prior beliefs only, and both combined.

Result: Both demographic and prior belief information improved predictions over a blind baseline, with their combination yielding the best performance in most cases. However, the relative value of each type of information varied substantially across different belief domains.

Conclusion: The findings reveal how current LLMs leverage different types of social information when reasoning about human beliefs, highlighting both their capabilities and limitations in understanding the interrelated landscape of human beliefs.

Abstract: Beliefs shape how people reason, communicate, and behave. Rather than existing in isolation, they exhibit a rich correlational structure–some connected through logical dependencies, others through indirect associations or social processes. As usage of large language models (LLMs) becomes more ubiquitous in our society, LLMs’ ability to understand and reason through human beliefs has many implications from privacy issues to personalized persuasion and the potential for stereotyping. Yet how LLMs capture this interrelated landscape of beliefs remains unclear. For instance, when predicting someone’s beliefs, what information affects the prediction most–who they are (demographics), what else they believe (prior stances), or a combination of both? We address these questions using data from an online debate platform, evaluating the ability of off-the-shelf open-weight LLMs to predict individuals’ stance under four conditions: no context, demographics only, prior beliefs only, and both combined. We find that both types of information improve predictions over a blind baseline, with their combination yielding the best performance in most cases. However, the relative value of each varies substantially across belief domains. These findings reveal how current LLMs leverage different types of social information when reasoning about human beliefs, highlighting both their capabilities and limitations.

Method: Proposes SSA (Sparse Sparse Attention) framework that integrates both sparse and full attention with bidirectional attention-output alignment. The method proves approximation error scales linearly with attention mass dropped under sparse attention, and SSA’s alignment objective reduces this quantity.

Result: SSA achieves state-of-the-art performance under both inference modes, adapts smoothly to varying sparsity budgets, and demonstrates superior long-context capabilities.

Conclusion: SSA effectively addresses both attention and capability gaps in sparse attention models through integrated training with alignment, enabling efficient long-context processing while maintaining performance.

Abstract: Sparse attention reduces the quadratic complexity of full self-attention but faces two challenges: (1) an attention gap, where applying sparse attention to full-attention-trained models causes performance degradation due to train-inference distribution mismatch, and (2) a capability gap, where models trained purely with sparse attention lack complete gradient flow, preventing them from matching full-attention performance. We propose SSA (Sparse Sparse Attention), a training framework that integrates both sparse and full attention with bidirectional attention-output alignment. We prove that the approximation error scales linearly with the attention mass dropped under sparse attention, and show that SSA’s alignment objective substantially reduces this quantity compared to baselines. Experiments demonstrate that SSA achieves state-of-the-art performance under both inference modes, adapts smoothly to varying sparsity budgets, and demonstrates superior long-context capabilities. The code is available at https://github.com/zhenyi4/ssa.

Method: Introduces ADRA-Bank, a human-annotated dataset of 200 instances across 10 academic domains, and ADRA-Eval, a modular evaluation paradigm that leverages academic paper structure to assess planning, retrieval, and reasoning capabilities through both end-to-end agent evaluation and isolated LLM backbone evaluation.

Result: Results show uneven capabilities: agents have specialized strengths but struggle with multi-source retrieval and cross-field consistency. Improving high-level planning capability is crucial for unlocking reasoning potential in foundational LLMs as backbones.

Conclusion: ADRA-Bank provides a diagnostic tool to guide development of more reliable automatic academic research assistants by exposing actionable failure modes in deep research agents.

Abstract: A surge in academic publications calls for automated deep research (DR) systems, but accurately evaluating them is still an open problem. First, existing benchmarks often focus narrowly on retrieval while neglecting high-level planning and reasoning. Second, existing benchmarks favor general domains over the academic domains that are the core application for DR agents. To address these gaps, we introduce ADRA-Bank, a modular benchmark for Academic DR Agents. Grounded in academic literature, our benchmark is a human-annotated dataset of 200 instances across 10 academic domains, including both research and review papers. Furthermore, we propose a modular Evaluation Paradigm for Academic DR Agents (ADRA-Eval), which leverages the rich structure of academic papers to assess the core capabilities of planning, retrieval, and reasoning. It employs two complementary modes: an end-to-end evaluation for \task agents and an isolated evaluation for foundational LLMs as potential backbones. Results reveal uneven capabilities: while agents show specialized strengths, they struggle with multi-source retrieval and cross-field consistency. Moreover, improving high-level planning capability is the crucial factor for unlocking the reasoning potential of foundational LLMs as backbones. By exposing these actionable failure modes, ADRA-Bank provides a diagnostic tool to guide the development of more reliable automatic academic research assistants.

Method: Two-stage approach: (1) Targeted Proximal Supervised Fine-Tuning (TPSFT) for constrained targeted edits with controlled policy drift, (2) Group Relative Policy Optimization (GRPO) to consolidate edits by aligning autoregressive trajectories with intended facts.

Result: EtCon improves editing reliability and real-world generalization while better preserving pre-trained capabilities compared to prior methods.

Conclusion: The edit-then-consolidate paradigm effectively addresses limitations of existing knowledge editing methods, making them more practical for real-world autoregressive generation scenarios.

Abstract: Knowledge editing aims to update specific facts in large language models (LLMs) without full retraining. Prior efforts sought to tune the knowledge layers of LLMs, achieving improved performance in controlled, teacher-forced evaluations. However, they still encounter challenges in real-world autoregressive generation scenarios, which greatly limit their practical applicability. Our empirical analysis reveals two issues: (1) Most methods degrade pre-trained capabilities after injecting new knowledge; (2) They may exhibit a discrepancy between stored parametric knowledge and inference-time autoregressive generation behavior. To this end, we propose EtCon, an edit-then-consolidate paradigm that couples targeted edits with post-edit consolidation. Specifically, our framework comprises two stages: (1) Targeted Proximal Supervised Fine-Tuning (TPSFT) performs a constrained targeted edit to update parametric knowledge while controlling policy drift. (2) Group Relative Policy Optimization (GRPO) consolidates the edit by aligning autoregressive trajectories with the intended fact. Extensive experiments demonstrate that our EtCon improves editing reliability and real-world generalization, while better preserving pre-trained capabilities.

Method: Proposes an Identify-then-Verify framework: 1) Generate multiple hypotheses about missing information and establish semantic consensus, 2) Perform critical verification step forcing model to re-examine source text to confirm if information is truly absent.

Result: The method outperforms established baselines across diverse multi-hop and factual QA datasets, producing more accurate sufficiency judgments while clearly articulating information gaps.

Conclusion: Guiding models to justify claims about missing information through structured reasoning provides more reliable signals for assessing context sufficiency in question-answering systems.

Abstract: Determining whether a provided context contains sufficient information to answer a question is a critical challenge for building reliable question-answering systems. While simple prompting strategies have shown success on factual questions, they frequently fail on inferential ones that require reasoning beyond direct text extraction. We hypothesize that asking a model to first reason about what specific information is missing provides a more reliable, implicit signal for assessing overall sufficiency. To this end, we propose a structured Identify-then-Verify framework for robust sufficiency modeling. Our method first generates multiple hypotheses about missing information and establishes a semantic consensus. It then performs a critical verification step, forcing the model to re-examine the source text to confirm whether this information is truly absent. We evaluate our method against established baselines across diverse multi-hop and factual QA datasets. The results demonstrate that by guiding the model to justify its claims about missing information, our framework produces more accurate sufficiency judgments while clearly articulating any information gaps.

Method: Reframes AR-to-DLM adaptation under Block-Diffusion paradigm with: 1) context-causal path keeping causal attention in prefix, 2) efficient parallel adaptation with AR guidance, and 3) gradual increment of generation block size for smoother transition from block size 1 to final Block-Diffusion state.

Result: The adaptation method proves competitive across various model scales. NBDiff-7B inherits long-context modeling and reasoning capabilities, achieving state-of-the-art performance among 7B-class diffusion language models.

Conclusion: The proposed principled adaptation pathway successfully transforms autoregressive models into diffusion language models while preserving their strengths, offering a practical and efficient alternative to training DLMs from scratch.

Abstract: Diffusion Language Models (DLMs) enable fast generation, yet training large DLMs from scratch is costly. As a practical shortcut, adapting off-the-shelf Auto-Regressive (AR) model weights into a DLM could quickly equip the DLM with strong long-context generation capabilies. Prior “adaptation” attempts either modify logits or randomly grow attention masks to Full-Sequence diffusion, or simply transplant AR weights into a Block-Diffusion recipe, leaving two key questions unaddressed: where is the final destination of adaptation, and how to adapt better? For manifold benefits, we reframe the whole AR-to-DLM adaptation under the Block-Diffusion paradigm, transitioning from block size 1 to the final Block-Diffusion state. Concretely, the principled pathway of adaptation is designed as follows: we keep a context-causal path where causal attention is kept in the prefix, an efficient parallel adaptation procedure where an AR guidance is maintained, and gradual increment of the generation block size for a smoother transition. Built on these components, the adaptation is proved competitive on various models at different scales. With better adaptation, we propose NBDiff-7B that could inherit the long-context modeling and reasoning capabilities, and achieve state-of-the-art performance among the 7B-class DLMs. Codes: https://github.com/YuchuanTian/NBDiff.

Method: Adapts Merlin-Arthur protocol: Arthur (generator LLM) trains on questions with unknown context provenance, Merlin provides helpful evidence, Morgana injects adversarial context. Both use XAI to identify/modify influential evidence. Introduces verification framework and Explained Information Fraction (EIF) metric.

Result: Across three RAG datasets and multiple LLM families/sizes, M/A training makes LLMs more grounded in evidence, increases information theoretic measures (soundness, completeness), improves reject behavior with fewer hallucinations, and improves retriever recall/MRR via automatically generated hard positives/negatives.

Conclusion: Autonomous interactive-proof-style supervision enables RAG systems to treat retrieved documents as verifiable evidence rather than suggestions, improving grounding and reliability without manual annotations.

Abstract: Retrieval-augmented generation (RAG) relies on retrieved context to guide large language models (LLM), yet treats retrieval as a weak heuristic rather than verifiable evidence – leading to unsupported answers, hallucinations, and reliance on spurious context. We introduce a novel training framework that treats the RAG pipeline as an interactive proof system by adapting the Merlin-Arthur (M/A) protocol: Arthur (the generator LLM) trains on questions with unknown context provenance and Merlin gives helpful evidence, while Morgana injects adversarial, misleading context. Both use an XAI method to identify and modify evidence most influential to Arthur. This trains Arthur to (1) answer when evidence supports the answer, (2) reject when evidence is insufficient, and (3) rely on the context spans that truly ground the answer. We further introduce a verification framework that disentangles explanation fidelity from model predictive errors, and introduce the Explained Information Fraction (EIF), which normalizes M/A mutual-information guarantees. Across three RAG datasets and multiple LLM families and sizes, M/A training makes LLMs more grounded in evidence, increases information theoretic measures (soundness, completeness) and reject behavior with less hallucinations, without manually annotated unanswerable samples. Finally, the retriever also improves recall and MRR via automatically generated M/A hard positives and negatives. While high accuracy does not guarantee entropy flow from context to answer, our EIF results show that autonomous interactive-proof-style supervision enables RAG systems that treat retrieved documents as verifiable evidence. % rather than suggestions.

Method: Uses dependency co-occurrence patterns of words instead of neural embeddings, focusing on syntactic relationships to detect semantic changes over time.

Result: The dependency-based method is effective for semantic change detection and outperforms several distributional semantic models while providing plausible, interpretable predictions.

Conclusion: Dependency co-occurrence patterns offer a viable, interpretable alternative to opaque neural embedding methods for lexical semantic change detection.

Abstract: Most modern computational approaches to lexical semantic change detection (LSC) rely on embedding-based distributional word representations with neural networks. Despite the strong performance on LSC benchmarks, they are often opaque. We investigate an alternative method which relies purely on dependency co-occurrence patterns of words. We demonstrate that it is effective for semantic change detection and even outperforms a number of distributional semantic models. We provide an in-depth quantitative and qualitative analysis of the predictions, showing that they are plausible and interpretable.

Method: Developed an interactive web application that runs in browser, using pre-computed traces of open LLMs applied on manually curated inputs to provide step-by-step visualizations.

Result: Created AnimatedLLM application available at https://animatedllm.github.io, serving as both a teaching aid and self-educational tool for understanding Transformer mechanics.

Conclusion: AnimatedLLM addresses the gap in educational materials for understanding LLM mechanics through interactive visualizations, making Transformer architecture more accessible for learning.

Abstract: Large language models (LLMs) are becoming central to natural language processing education, yet materials showing their mechanics are sparse. We present AnimatedLLM, an interactive web application that provides step-by-step visualizations of a Transformer language model. AnimatedLLM runs entirely in the browser, using pre-computed traces of open LLMs applied on manually curated inputs. The application is available at https://animatedllm.github.io, both as a teaching aid and for self-educational purposes.

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

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

Conclusion: Deep Research Bench II provides a rigorous evaluation framework for deep research systems, revealing significant limitations in current systems and establishing a benchmark for future development in this area.

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

Method: HumanLLM treats psychological patterns as interacting causal forces, constructing 244 patterns from ~12,000 academic papers and synthesizing 11,359 scenarios where 2-5 patterns reinforce, conflict, or modulate each other. The framework uses multi-turn conversations expressing inner thoughts, actions, and dialogue, with dual-level checklists evaluating both individual pattern fidelity and emergent multi-pattern dynamics.

Result: HumanLLM achieves strong human alignment (r=0.91) and reveals that holistic metrics conflate simulation accuracy with social desirability. HumanLLM-8B outperforms Qwen3-32B on multi-pattern dynamics despite having 4x fewer parameters.

Conclusion: Authentic anthropomorphism in language agents requires cognitive modeling that simulates not just what humans do, but the psychological processes generating those behaviors. HumanLLM demonstrates that treating psychological patterns as interacting causal forces enables more accurate human alignment.

Abstract: Large Language Models (LLMs) have demonstrated remarkable capabilities in reasoning and generation, serving as the foundation for advanced persona simulation and Role-Playing Language Agents (RPLAs). However, achieving authentic alignment with human cognitive and behavioral patterns remains a critical challenge for these agents. We present HumanLLM, a framework treating psychological patterns as interacting causal forces. We construct 244 patterns from ~12,000 academic papers and synthesize 11,359 scenarios where 2-5 patterns reinforce, conflict, or modulate each other, with multi-turn conversations expressing inner thoughts, actions, and dialogue. Our dual-level checklists evaluate both individual pattern fidelity and emergent multi-pattern dynamics, achieving strong human alignment (r=0.91) while revealing that holistic metrics conflate simulation accuracy with social desirability. HumanLLM-8B outperforms Qwen3-32B on multi-pattern dynamics despite 4x fewer parameters, demonstrating that authentic anthropomorphism requires cognitive modeling–simulating not just what humans do, but the psychological processes generating those behaviors.

Method: Systematic evaluation of LLM susceptibility to persuasion using the SMCR communication framework across five LLMs and three domains (factual knowledge, medical QA, social bias). Analyzed belief stability over multiple interaction turns, examined meta-cognition prompting effects, and evaluated adversarial fine-tuning as a defense mechanism.

Result: Smallest model (Llama 3.2-3B) showed extreme compliance with 82.5% belief changes at first persuasive turn; meta-cognition prompting increased vulnerability rather than enhancing robustness; adversarial fine-tuning effectiveness varied significantly by model (GPT-4o-mini: 98.6% robust, Mistral~7B: 35.7%→79.3%, Llama models: <14% even after fine-tuning).

Conclusion: LLMs have substantial model-dependent limits to persuasion resistance; current robustness interventions are insufficient, especially for smaller models; findings provide guidance for developing more trustworthy LLMs.

Abstract: Large Language Models (LLMs) are increasingly employed in various question-answering tasks. However, recent studies showcase that LLMs are susceptible to persuasion and could adopt counterfactual beliefs. We present a systematic evaluation of LLM susceptibility to persuasion under the Source–Message–Channel–Receiver (SMCR) communication framework. Across five mainstream Large Language Models (LLMs) and three domains (factual knowledge, medical QA, and social bias), we analyze how different persuasive strategies influence belief stability over multiple interaction turns. We further examine whether meta-cognition prompting (i.e., eliciting self-reported confidence) affects resistance to persuasion. Results show that the smallest model (Llama 3.2-3B) exhibits extreme compliance, with 82.5% of belief changes occurring at the first persuasive turn (average end turn of 1.1–1.4). Contrary to expectations, meta-cognition prompting increases vulnerability by accelerating belief erosion rather than enhancing robustness. Finally, we evaluate adversarial fine-tuning as a defense. While GPT-4o-mini achieves near-complete robustness (98.6%) and Mistral~7B improves substantially (35.7% $\rightarrow$ 79.3%), Llama models remain highly susceptible (<14%) even when fine-tuned on their own failure cases. Together, these findings highlight substantial model-dependent limits of current robustness interventions and offer guidance for developing more trustworthy LLMs.

Method: LogicScore uses Horn Rules and backward verification to systematically evaluate three reasoning dimensions: Completeness (logically sound deduction), Conciseness (non-redundancy), and Determinateness (consistent answer entailment).

Result: Experiments across three multi-hop QA datasets and over 20 LLMs show a critical capability gap: leading models achieve high attribution scores (e.g., 92.85% precision for Gemini-3 Pro) but struggle with global reasoning quality (e.g., 35.11% Conciseness for Gemini-3 Pro).

Conclusion: The work establishes a robust standard for logical evaluation, highlighting the need to prioritize reasoning coherence alongside factual grounding in LLM development.

Abstract: Current evaluation methods for Attributed Question Answering (AQA) suffer from \textit{attribution myopia}: they emphasize verification of isolated statements and their attributions but overlook the global logical integrity of long-form answers. Consequently, Large Language Models (LLMs) often produce factually grounded yet logically incoherent responses with elusive deductive gaps. To mitigate this limitation, we present \textsc{LogicScore}, a unified evaluation framework that shifts the paradigm from local assessment to global reasoning scrutiny. Grounded in Horn Rules, our approach integrates a backward verification mechanism to systematically evaluate three key reasoning dimensions: \textit{Completeness} (logically sound deduction), \textit{Conciseness} (non-redundancy), and \textit{Determinateness} (consistent answer entailment). Extensive experiments across three multi-hop QA datasets (HotpotQA, MusiQue, and 2WikiMultiHopQA) and over 20 LLMs (including GPT-5, Gemini-3-Pro, LLaMA3, and task-specific tuned models) reveal a critical capability gap: leading models often achieve high attribution scores (e.g., 92.85% precision for Gemini-3 Pro) but struggle with global reasoning quality (e.g., 35.11% Conciseness for Gemini-3 Pro). Our work establishes a robust standard for logical evaluation, highlighting the need to prioritize reasoning coherence alongside factual grounding in LLM development. Codes are available at: https://github.com/zhichaoyan11/LogicScore.

Method: Two-stage pipeline: (1) multi-turn exploration to generate valid tool-call trajectories, (2) conversion of trajectories into user-facing tasks with controlled intent adaptations, creating verifiable tasks for closed-loop evaluation and training.

Result: Benchmarking 7 SOTA LLMs shows frequent failures on complex user scenarios. Fine-tuning lightweight LLMs with successful trajectories yields consistent improvements across all three conditions and better generalization to unseen tool-use domains.

Conclusion: The Trajectory2Task pipeline addresses the data gap for realistic tool-calling scenarios, enabling better evaluation and training of LLMs for practical deployment, with fine-tuned models showing improved general tool-calling abilities.

Abstract: Tool-calling agents are increasingly deployed in real-world customer-facing workflows. Yet most studies on tool-calling agents focus on idealized settings with general, fixed, and well-specified tasks. In real-world applications, user requests are often (1) ambiguous, (2) changing over time, or (3) infeasible due to policy constraints, and training and evaluation data that cover these diverse, complex interaction patterns remain under-represented. To bridge the gap, we present Trajectory2Task, a verifiable data generation pipeline for studying tool use at scale under three realistic user scenarios: ambiguous intent, changing intent, and infeasible intents. The pipeline first conducts multi-turn exploration to produce valid tool-call trajectories. It then converts these trajectories into user-facing tasks with controlled intent adaptations. This process yields verifiable task that support closed-loop evaluation and training. We benchmark seven state-of-the-art LLMs on the generated complex user scenario tasks and observe frequent failures. Finally, using successful trajectories obtained from task rollouts, we fine-tune lightweight LLMs and find consistent improvements across all three conditions, along with better generalization to unseen tool-use domains, indicating stronger general tool-calling ability.

Method: Proposed AgentIF-OneDay benchmark with 104 tasks across 767 scoring points in three categories: Open Workflow Execution (explicit workflows), Latent Instruction (implicit instructions from attachments), and Iterative Refinement (modifying ongoing work). Uses instance-level rubrics and LLM-based verification aligned with human judgment.

Result: Achieved 80.1% agreement rate between LLM-based verification and human judgment using Gemini-3-Pro. Found that API-based agent products and ChatGPT agents with agent RL perform best, showing that leading LLM APIs and open-source models have internalized agentic capabilities.

Conclusion: AgentIF-OneDay provides comprehensive evaluation of AI agents for daily tasks, revealing that current models have strong agentic capabilities that enable development of cutting-edge agent products, though general user perception remains limited.

Abstract: The capacity of AI agents to effectively handle tasks of increasing duration and complexity continues to grow, demonstrating exceptional performance in coding, deep research, and complex problem-solving evaluations. However, in daily scenarios, the perception of these advanced AI capabilities among general users remains limited. We argue that current evaluations prioritize increasing task difficulty without sufficiently addressing the diversity of agentic tasks necessary to cover the daily work, life, and learning activities of a broad demographic. To address this, we propose AgentIF-OneDay, aimed at determining whether general users can utilize natural language instructions and AI agents to complete a diverse array of daily tasks. These tasks require not only solving problems through dialogue but also understanding various attachment types and delivering tangible file-based results. The benchmark is structured around three user-centric categories: Open Workflow Execution, which assesses adherence to explicit and complex workflows; Latent Instruction, which requires agents to infer implicit instructions from attachments; and Iterative Refinement, which involves modifying or expanding upon ongoing work. We employ instance-level rubrics and a refined evaluation pipeline that aligns LLM-based verification with human judgment, achieving an 80.1% agreement rate using Gemini-3-Pro. AgentIF-OneDay comprises 104 tasks covering 767 scoring points. We benchmarked four leading general AI agents and found that agent products built based on APIs and ChatGPT agents based on agent RL remain in the first tier simultaneously. Leading LLM APIs and open-source models have internalized agentic capabilities, enabling AI application teams to develop cutting-edge Agent products.

Method: Introduces AgentLongBench, which evaluates agents through simulated environment rollouts based on Lateral Thinking Puzzles. The framework generates rigorous interaction trajectories across both knowledge-intensive and knowledge-free scenarios, testing models with context windows ranging from 32K to 4M tokens.

Result: Experiments reveal a critical weakness: while agents are adept at static retrieval, they struggle with dynamic information synthesis essential for complex workflows. The degradation is driven by the minimum number of tokens required to resolve a query, with high information density in massive tool responses posing greater challenges than memory fragmentation in long-turn dialogues.

Conclusion: AgentLongBench exposes fundamental limitations in current LLM agents’ ability to handle dynamic, interactive environments, highlighting the need for improved architectures and training approaches that can better manage complex information synthesis in agent workflows.

Abstract: The evolution of Large Language Models (LLMs) into autonomous agents necessitates the management of extensive, dynamic contexts. Current benchmarks, however, remain largely static, relying on passive retrieval tasks that fail to simulate the complexities of agent-environment interaction, such as non-linear reasoning and iterative feedback. To address this, we introduce \textbf{AgentLongBench}, which evaluates agents through simulated environment rollouts based on Lateral Thinking Puzzles. This framework generates rigorous interaction trajectories across knowledge-intensive and knowledge-free scenarios. Experiments with state-of-the-art models and memory systems (32K to 4M tokens) expose a critical weakness: while adept at static retrieval, agents struggle with the dynamic information synthesis essential for workflows. Our analysis indicates that this degradation is driven by the minimum number of tokens required to resolve a query. This factor explains why the high information density inherent in massive tool responses poses a significantly greater challenge than the memory fragmentation typical of long-turn dialogues.

Method: ASTRA combines: 1) A pipeline using static tool-call graph topology to synthesize diverse trajectories for broad tool-use competence; 2) An environment synthesis framework converting decomposed QA traces into code-executable, rule-verifiable environments for deterministic multi-turn RL; 3) Unified training integrating SFT with online RL using trajectory-level rewards.

Result: ASTRA-trained models achieve state-of-the-art performance on multiple agentic tool-use benchmarks at comparable scales, approaching closed-source systems while preserving core reasoning ability.

Conclusion: ASTRA provides a fully automated, end-to-end framework for training robust tool-augmented language model agents through scalable data synthesis and verifiable reinforcement learning, addressing key challenges in multi-step decision making.

Abstract: Large language models (LLMs) are increasingly used as tool-augmented agents for multi-step decision making, yet training robust tool-using agents remains challenging. Existing methods still require manual intervention, depend on non-verifiable simulated environments, rely exclusively on either supervised fine-tuning (SFT) or reinforcement learning (RL), and struggle with stable long-horizon, multi-turn learning. To address these challenges, we introduce ASTRA, a fully automated end-to-end framework for training tool-augmented language model agents via scalable data synthesis and verifiable reinforcement learning. ASTRA integrates two complementary components. First, a pipeline that leverages the static topology of tool-call graphs synthesizes diverse, structurally grounded trajectories, instilling broad and transferable tool-use competence. Second, an environment synthesis framework that captures the rich, compositional topology of human semantic reasoning converts decomposed question-answer traces into independent, code-executable, and rule-verifiable environments, enabling deterministic multi-turn RL. Based on this method, we develop a unified training methodology that integrates SFT with online RL using trajectory-level rewards to balance task completion and interaction efficiency. Experiments on multiple agentic tool-use benchmarks demonstrate that ASTRA-trained models achieve state-of-the-art performance at comparable scales, approaching closed-source systems while preserving core reasoning ability. We release the full pipelines, environments, and trained models at https://github.com/LianjiaTech/astra.

Method: Introduces CoFrGeNets (Continued Fraction Generative Networks) - a new function class inspired by continued fractions. Designs novel architectural components to replace Multi-head Attention and Feed-Forward Networks in Transformer blocks. Derives custom gradient formulations for more accurate and efficient optimization.

Result: Models achieve competitive/superior performance on downstream tasks (classification, QA, reasoning, text understanding) with 1/2 to 2/3 the parameters and shorter pre-training time compared to GPT2-xl (1.5B) and Llama3 (3.2B).

Conclusion: CoFrGeNets offer a parameter-efficient alternative to Transformers that can be easily integrated into existing workflows, with potential for further hardware optimization.

Abstract: Transformers are arguably the preferred architecture for language generation. In this paper, inspired by continued fractions, we introduce a new function class for generative modeling. The architecture family implementing this function class is named CoFrGeNets - Continued Fraction Generative Networks. We design novel architectural components based on this function class that can replace Multi-head Attention and Feed-Forward Networks in Transformer blocks while requiring much fewer parameters. We derive custom gradient formulations to optimize the proposed components more accurately and efficiently than using standard PyTorch-based gradients. Our components are a plug-in replacement requiring little change in training or inference procedures that have already been put in place for Transformer-based models thus making our approach easy to incorporate in large industrial workflows. We experiment on two very different transformer architectures GPT2-xl (1.5B) and Llama3 (3.2B), where the former we pre-train on OpenWebText and GneissWeb, while the latter we pre-train on the docling data mix which consists of nine different datasets. Results show that the performance on downstream classification, Q& A, reasoning and text understanding tasks of our models is competitive and sometimes even superior to the original models with $\frac{2}{3}$ to $\frac{1}{2}$ the parameters and shorter pre-training time. We believe that future implementations customized to hardware will further bring out the true potential of our architectures.

Method: Token-level hallucination control through self-checking decoding: 1) Internal verification at each reasoning step to detect hallucinated tokens, 2) Latent space evaluation with explicit hallucination risk scoring for candidate fragments, 3) Iterative pruning and regeneration to dynamically correct detected errors.

Result: Experiments on HALU datasets show Token-Guard substantially reduces hallucinations and improves generation accuracy compared to baseline methods.

Conclusion: Token-Guard offers a scalable, modular solution for reliable LLM outputs with lightweight hallucination control, addressing limitations of existing resource-intensive methods.

Abstract: Large Language Models (LLMs) often hallucinate, generating content inconsistent with the input. Retrieval-Augmented Generation (RAG) and Reinforcement Learning with Human Feedback (RLHF) can mitigate hallucinations but require resource-intensive retrieval or large-scale fine-tuning. Decoding-based methods are lighter yet lack explicit hallucination control. To address this, we present Token-Guard, a token-level hallucination control method based on self-checking decoding. Token-Guard performs internal verification at each reasoning step to detect hallucinated tokens before they propagate. Candidate fragments are further evaluated in a latent space with explicit hallucination risk scoring, while iterative pruning and regeneration dynamically correct detected errors. Experiments on HALU datasets show Token-Guard substantially reduces hallucinations and improves generation accuracy, offering a scalable, modular solution for reliable LLM outputs. Our code is publicly available.

Method: Evaluated five state-of-the-art OVD models on LAE-80C aerial dataset (3,592 images, 80 categories) under strict zero-shot conditions using Global, Oracle, and Single-Category inference modes to isolate semantic confusion from visual localization.

Result: Best model (OWLv2) achieved only 27.6% F1-score with 69% false positive rate. Reducing vocabulary size from 80 to 3.2 classes yielded 15x improvement, showing semantic confusion is primary bottleneck. Prompt engineering strategies failed to provide meaningful gains.

Conclusion: Open-vocabulary object detection suffers severe domain transfer failure in aerial imagery, establishing baseline expectations and highlighting need for domain-adaptive approaches in this specialized domain.

Abstract: Open-vocabulary object detection (OVD) enables zero-shot recognition of novel categories through vision-language models, achieving strong performance on natural images. However, transferability to aerial imagery remains unexplored. We present the first systematic benchmark evaluating five state-of-the-art OVD models on the LAE-80C aerial dataset (3,592 images, 80 categories) under strict zero-shot conditions. Our experimental protocol isolates semantic confusion from visual localization through Global, Oracle, and Single-Category inference modes. Results reveal severe domain transfer failure: the best model (OWLv2) achieves only 27.6% F1-score with 69% false positive rate. Critically, reducing vocabulary size from 80 to 3.2 classes yields 15x improvement, demonstrating that semantic confusion is the primary bottleneck. Prompt engineering strategies such as domain-specific prefixing and synonym expansion, fail to provide meaningful performance gains. Performance varies dramatically across datasets (F1: 0.53 on DIOR, 0.12 on FAIR1M), exposing brittleness to imaging conditions. These findings establish baseline expectations and highlight the need for domain-adaptive approaches in aerial OVD.

Method: Survey existing VQA datasets, generate synthetic histogram charts with ground truth data, and test both humans and large reasoning models on questions requiring access to underlying data.

Result: Created and released an open-source dataset with synthetic histogram charts, underlying data, distribution parameters, and bounding boxes for all figure marks and text.

Conclusion: Established a dedicated VQA benchmark for scientific charts that requires deeper reasoning about data transformations, addressing a gap in current multimodal evaluation.

Abstract: Visual Question Answering (VQA) has become an important benchmark for assessing how large multimodal models (LMMs) interpret images. However, most VQA datasets focus on real-world images or simple diagrammatic analysis, with few focused on interpreting complex scientific charts. Indeed, many VQA datasets that analyze charts do not contain the underlying data behind those charts or assume a 1-to-1 correspondence between chart marks and underlying data. In reality, charts are transformations (i.e. analysis, simplification, modification) of data. This distinction introduces a reasoning challenge in VQA that the current datasets do not capture. In this paper, we argue for a dedicated VQA benchmark for scientific charts where there is no 1-to-1 correspondence between chart marks and underlying data. To do so, we survey existing VQA datasets and highlight limitations of the current field. We then generate synthetic histogram charts based on ground truth data, and ask both humans and a large reasoning model questions where precise answers depend on access to the underlying data. We release the open-source dataset, including figures, underlying data, distribution parameters used to generate the data, and bounding boxes for all figure marks and text for future research.

Method: Introduces VRRPI-Bench (derived from unlabeled egocentric videos with verbalized annotations) and VRRPI-Diag (diagnostic benchmark isolating individual motion degrees of freedom) to evaluate VLMs on relative camera pose estimation.

Result: Most VLMs fail to generalize beyond shallow 2D heuristics, particularly for depth changes and roll transformations. Even state-of-the-art models like GPT-5 (0.64) fall short of classic geometric baselines (0.97) and human performance (0.92). VLMs also struggle with multi-image reasoning (best 59.7%).

Conclusion: VLMs have significant limitations in 3D spatial grounding and multi-view reasoning, revealing gaps in their understanding of fundamental 3D vision tasks despite strong 2D capabilities.

Abstract: Vision-Language Models (VLMs) perform well in 2D perception and semantic reasoning compared to their limited understanding of 3D spatial structure. We investigate this gap using relative camera pose estimation (RCPE), a fundamental vision task that requires inferring relative camera translation and rotation from a pair of images. We introduce VRRPI-Bench, a benchmark derived from unlabeled egocentric videos with verbalized annotations of relative camera motion, reflecting realistic scenarios with simultaneous translation and rotation around a shared object. We further propose VRRPI-Diag, a diagnostic benchmark that isolates individual motion degrees of freedom. Despite the simplicity of RCPE, most VLMs fail to generalize beyond shallow 2D heuristics, particularly for depth changes and roll transformations along the optical axis. Even state-of-the-art models such as GPT-5 ($0.64$) fall short of classic geometric baselines ($0.97$) and human performance ($0.92$). Moreover, VLMs exhibit difficulty in multi-image reasoning, with inconsistent performance (best $59.7%$) when integrating spatial cues across frames. Our findings reveal limitations in grounding VLMs in 3D and multi-view spatial reasoning.

Method: Introduce token-level diagnostics to measure multi-view geometric consistency in ViT representations, conduct extensive experiments on 14 foundation ViT models, and analyze how consistent positional embeddings affect geometric properties.

Result: Reveal that positional embeddings function as geometric priors that shape spatial structure, and demonstrate how they influence multi-view geometry and spatial reasoning in ViT representations.

Conclusion: Positional embeddings serve as a causal mechanism governing spatial structure in vision transformer representations, clarifying their role as geometric priors rather than simple token indices.

Abstract: This paper revisits the role of positional embeddings (PEs) within vision transformers (ViTs) from a geometric perspective. We show that PEs are not mere token indices but effectively function as geometric priors that shape the spatial structure of the representation. We introduce token-level diagnostics that measure how multi-view geometric consistency in ViT representation depends on consitent PEs. Through extensive experiments on 14 foundation ViT models, we reveal how PEs influence multi-view geometry and spatial reasoning. Our findings clarify the role of PEs as a causal mechanism that governs spatial structure in ViT representations. Our code is provided in https://github.com/shijianjian/vit-geometry-probes

Method: Compared a two-level VQ-VAE with a capacity-matched single-level model on high-resolution ImageNet images. Used lightweight interventions: initialization from data, periodic reset of inactive codebook vectors, and systematic tuning of codebook hyperparameters to mitigate codebook collapse.

Result: When representational budgets are matched and codebook collapse is mitigated, single-level VQ-VAEs can match the reconstruction fidelity of hierarchical variants. Inadequate codebook utilization limits single-level models, and overly high-dimensional embeddings destabilize quantization and increase codebook collapse.

Conclusion: Hierarchical quantization is not inherently superior for high-quality reconstructions when capacity is matched and codebook collapse is properly addressed. The multi-scale structure may improve perceptual quality in downstream tasks, but hierarchy doesn’t necessarily improve reconstruction accuracy.

Abstract: Vector-quantized variational autoencoders (VQ-VAEs) are central to models that rely on high reconstruction fidelity, from neural compression to generative pipelines. Hierarchical extensions, such as VQ-VAE2, are often credited with superior reconstruction performance because they split global and local features across multiple levels. However, since higher levels derive all their information from lower levels, they should not carry additional reconstructive content beyond what the lower-level already encodes. Combined with recent advances in training objectives and quantization mechanisms, this leads us to ask whether a single-level VQ-VAE, with matched representational budget and no codebook collapse, can equal the reconstruction fidelity of its hierarchical counterpart. Although the multi-scale structure of hierarchical models may improve perceptual quality in downstream tasks, the effect of hierarchy on reconstruction accuracy, isolated from codebook utilization and overall representational capacity, remains empirically underexamined. We revisit this question by comparing a two-level VQ-VAE and a capacity-matched single-level model on high-resolution ImageNet images. Consistent with prior observations, we confirm that inadequate codebook utilization limits single-level VQ-VAEs and that overly high-dimensional embeddings destabilize quantization and increase codebook collapse. We show that lightweight interventions such as initialization from data, periodic reset of inactive codebook vectors, and systematic tuning of codebook hyperparameters significantly reduce collapse. Our results demonstrate that when representational budgets are matched, and codebook collapse is mitigated, single-level VQ-VAEs can match the reconstruction fidelity of hierarchical variants, challenging the assumption that hierarchical quantization is inherently superior for high-quality reconstructions.

Method: Proposes VMonarch attention mechanism using structured Monarch matrices: 1) spatio-temporal Monarch factorization to capture intra/inter-frame correlations, 2) recomputation strategy to mitigate alternating minimization artifacts, 3) online entropy algorithm fused with FlashAttention for fast Monarch matrix updates.

Result: Achieves comparable/superior generation quality to full attention on VBench, reduces attention FLOPs by 17.5x, achieves over 5x speedup in attention computation for long videos, outperforms state-of-the-art sparse attention methods at 90% sparsity.

Conclusion: VMonarch effectively overcomes attention bottlenecks in Video DiTs through structured sparse attention, enabling efficient long-video generation while maintaining quality, representing a significant advance in scalable video generation models.

Abstract: The quadratic complexity of the attention mechanism severely limits the context scalability of Video Diffusion Transformers (DiTs). We find that the highly sparse spatio-temporal attention patterns exhibited in Video DiTs can be naturally represented by the Monarch matrix. It is a class of structured matrices with flexible sparsity, enabling sub-quadratic attention via an alternating minimization algorithm. Accordingly, we propose VMonarch, a novel attention mechanism for Video DiTs that enables efficient computation over the dynamic sparse patterns with structured Monarch matrices. First, we adapt spatio-temporal Monarch factorization to explicitly capture the intra-frame and inter-frame correlations of the video data. Second, we introduce a recomputation strategy to mitigate artifacts arising from instabilities during alternating minimization of Monarch matrices. Third, we propose a novel online entropy algorithm fused into FlashAttention, enabling fast Monarch matrix updates for long sequences. Extensive experiments demonstrate that VMonarch achieves comparable or superior generation quality to full attention on VBench after minimal tuning. It overcomes the attention bottleneck in Video DiTs, reduces attention FLOPs by a factor of 17.5, and achieves a speedup of over 5x in attention computation for long videos, surpassing state-of-the-art sparse attention methods at 90% sparsity.

Method: Two-phase mixed CG-real training strategy: 1) Learns strong generative prior from large-scale real footage, 2) Introduces controllability through shared implicit spatio-temporal features across domains. Uses coarse 3D renderings to control scene layout, camera motion, and human trajectories, while a neural renderer generates realistic appearance, lighting, and fine-scale dynamics guided by text prompts.

Result: The system supports coarse-to-fine control, generalizes across diverse CG and game inputs, and produces temporally consistent, controllable, and realistic urban scene videos from minimal 3D input.

Conclusion: C2R demonstrates a novel approach to generative rendering that bridges the gap between coarse 3D simulations and realistic video generation, offering improved scalability and controllability for urban crowd scenes.

Abstract: Traditional rendering pipelines rely on complex assets, accurate materials and lighting, and substantial computational resources to produce realistic imagery, yet they still face challenges in scalability and realism for populated dynamic scenes. We present C2R (Coarse-to-Real), a generative rendering framework that synthesizes real-style urban crowd videos from coarse 3D simulations. Our approach uses coarse 3D renderings to explicitly control scene layout, camera motion, and human trajectories, while a learned neural renderer generates realistic appearance, lighting, and fine-scale dynamics guided by text prompts. To overcome the lack of paired training data between coarse simulations and real videos, we adopt a two-phase mixed CG-real training strategy that learns a strong generative prior from large-scale real footage and introduces controllability through shared implicit spatio-temporal features across domains. The resulting system supports coarse-to-fine control, generalizes across diverse CG and game inputs, and produces temporally consistent, controllable, and realistic urban scene videos from minimal 3D input. We will release the model and project webpage at https://gonzalognogales.github.io/coarse2real/.

Method: FlexMap eliminates explicit geometric projections by using a geometry-aware foundation model with cross-frame attention to implicitly encode 3D scene understanding in feature space. It features two core components: 1) a spatial-temporal enhancement module that separates cross-view spatial reasoning from temporal dynamics, and 2) a camera-aware decoder with latent camera tokens for view-adaptive attention without projection matrices.

Result: Experiments demonstrate that FlexMap outperforms existing methods across multiple camera configurations while maintaining robustness to missing views and sensor variations. The method enables more practical real-world deployment by adapting to variable camera setups without retraining.

Conclusion: FlexMap provides a flexible and robust approach to HD map construction that can handle variable camera configurations without architectural changes or per-configuration retraining, making autonomous driving systems more practical and resilient to sensor variations.

Abstract: High-definition (HD) maps provide essential semantic information of road structures for autonomous driving systems, yet current HD map construction methods require calibrated multi-camera setups and either implicit or explicit 2D-to-BEV transformations, making them fragile when sensors fail or camera configurations vary across vehicle fleets. We introduce FlexMap, unlike prior methods that are fixed to a specific N-camera rig, our approach adapts to variable camera configurations without any architectural changes or per-configuration retraining. Our key innovation eliminates explicit geometric projections by using a geometry-aware foundation model with cross-frame attention to implicitly encode 3D scene understanding in feature space. FlexMap features two core components: a spatial-temporal enhancement module that separates cross-view spatial reasoning from temporal dynamics, and a camera-aware decoder with latent camera tokens, enabling view-adaptive attention without the need for projection matrices. Experiments demonstrate that FlexMap outperforms existing methods across multiple configurations while maintaining robustness to missing views and sensor variations, enabling more practical real-world deployment.

Method: Dual-strategy approach: 1) Post-training Chain-of-Thought prompting to construct stealthy prompts, and 2) ReAct-driven adaptive noising mechanism that iteratively perturbs input images based on model feedback to refine adversarial noise in safety-sensitive regions.

Result: Experimental results show the proposed dual-strategy significantly improves attack success rates while maintaining naturalness in both text and visual domains.

Conclusion: The framework demonstrates effective jailbreaking of vision-language models through combined prompt engineering and adaptive image perturbation, revealing vulnerabilities in safety alignment.

Abstract: Vision-language models (VLMs) have become central to tasks such as visual question answering, image captioning, and text-to-image generation. However, their outputs are highly sensitive to prompt variations, which can reveal vulnerabilities in safety alignment. In this work, we present a jailbreak framework that exploits post-training Chain-of-Thought (CoT) prompting to construct stealthy prompts capable of bypassing safety filters. To further increase attack success rates (ASR), we propose a ReAct-driven adaptive noising mechanism that iteratively perturbs input images based on model feedback. This approach leverages the ReAct paradigm to refine adversarial noise in regions most likely to activate safety defenses, thereby enhancing stealth and evasion. Experimental results demonstrate that the proposed dual-strategy significantly improves ASR while maintaining naturalness in both text and visual domains.

Method: Uses conditional diffusion model with Semantically-Disentangled Style Encoder (SDSE) to extract pure style representations from reference videos, plus hierarchical modulation strategy to balance audio and style features across facial regions.

Result: Significant improvements over state-of-the-art methods in both lip-sync accuracy and personalization preservation, as demonstrated through extensive experiments.

Conclusion: MirrorTalk effectively addresses the style-content entanglement problem in talking face synthesis, enabling faithful transfer of speaker persona while maintaining accurate lip synchronization.

Abstract: Synthesizing personalized talking faces that uphold and highlight a speaker’s unique style while maintaining lip-sync accuracy remains a significant challenge. A primary limitation of existing approaches is the intrinsic confounding of speaker-specific talking style and semantic content within facial motions, which prevents the faithful transfer of a speaker’s unique persona to arbitrary speech. In this paper, we propose MirrorTalk, a generative framework based on a conditional diffusion model, combined with a Semantically-Disentangled Style Encoder (SDSE) that can distill pure style representations from a brief reference video. To effectively utilize this representation, we further introduce a hierarchical modulation strategy within the diffusion process. This mechanism guides the synthesis by dynamically balancing the contributions of audio and style features across distinct facial regions, ensuring both precise lip-sync accuracy and expressive full-face dynamics. Extensive experiments demonstrate that MirrorTalk achieves significant improvements over state-of-the-art methods in terms of lip-sync accuracy and personalization preservation.

Method: Uses variational inference to estimate joint angle posterior distributions, evaluates calibration using Expected Calibration Error (ECE), validates against instrumented walkway and marker-based motion capture across 68 participants from two institutions.

Result: Model demonstrated reliable calibration with ECE values generally <0.1 for step/stride length and gait kinematics, median errors of ~16mm (step) and ~12mm (stride), kinematic errors 1.5-3.8 degrees, predicted uncertainty strongly correlated with observed error.

Conclusion: Probabilistic model reconstruction effectively quantifies epistemic uncertainty, allowing identification of unreliable outputs without concurrent ground-truth instrumentation, advancing clinical trust in markerless motion capture systems.

Abstract: Video-based human movement analysis holds potential for movement assessment in clinical practice and research. However, the clinical implementation and trust of multi-view markerless motion capture (MMMC) require that, in addition to being accurate, these systems produce reliable confidence intervals to indicate how accurate they are for any individual. Building on our prior work utilizing variational inference to estimate joint angle posterior distributions, this study evaluates the calibration and reliability of a probabilistic MMMC method. We analyzed data from 68 participants across two institutions, validating the model against an instrumented walkway and standard marker-based motion capture. We measured the calibration of the confidence intervals using the Expected Calibration Error (ECE). The model demonstrated reliable calibration, yielding ECE values generally < 0.1 for both step and stride length and bias-corrected gait kinematics. We observed a median step and stride length error of ~16 mm and ~12 mm respectively, with median bias-corrected kinematic errors ranging from 1.5 to 3.8 degrees across lower extremity joints. Consistent with the calibrated ECE, the magnitude of the model’s predicted uncertainty correlated strongly with observed error measures. These findings indicate that, as designed, the probabilistic model reconstruction quantifies epistemic uncertainty, allowing it to identify unreliable outputs without the need for concurrent ground-truth instrumentation.

Method: Proposes Language-Prior-Free Verification to enable LVLMs to faithfully verify object existence confidence, and a Self-Validation Framework that validates objects in candidate captions and mitigates hallucination via caption selection or aggregation.

Result: Significantly mitigates object hallucination in image captioning (65.6% improvement on CHAIRI metric with LLaVA-v1.5-7B), surpassing previous SOTA methods.

Conclusion: Demonstrates a novel path toward mitigating hallucination by unlocking the inherent potential within LVLMs themselves, highlighting that training-free approaches can effectively address over-reliance on language priors.

Abstract: Despite progress in Large Vision Language Models (LVLMs), object hallucination remains a critical issue in image captioning task, where models generate descriptions of non-existent objects, compromising their reliability. Previous work attributes this to LVLMs’ over-reliance on language priors and attempts to mitigate it through logits calibration. However, they still lack a thorough analysis of the over-reliance. To gain a deeper understanding of over-reliance, we conduct a series of preliminary experiments, indicating that as the generation length increases, LVLMs’ over-reliance on language priors leads to inflated probability of hallucinated object tokens, consequently exacerbating object hallucination. To circumvent this issue, we propose Language-Prior-Free Verification to enable LVLMs to faithfully verify the confidence of object existence. Based on this, we propose a novel training-free Self-Validation Framework to counter the over-reliance trap. It first validates objects’ existence in sampled candidate captions and further mitigates object hallucination via caption selection or aggregation. Experiment results demonstrate that our framework mitigates object hallucination significantly in image captioning task (e.g., 65.6% improvement on CHAIRI metric with LLaVA-v1.5-7B), surpassing the previous SOTA methods. This result highlights a novel path towards mitigating hallucination by unlocking the inherent potential within LVLMs themselves.

Method: Combines multimodal large language models (MLLMs) and image generation models. Uses MLLMs’ visual capabilities to predict editing intent behind scribbles, then employs globally generated images to extract local texture details and anchor local semantics.

Result: Achieves state-of-the-art interactive editing performance for scribble-based texture editing by effectively leveraging MLLMs’ strengths.

Conclusion: ScribbleSense successfully addresses scribble ambiguity in 3D texture editing through multimodal LLM integration, enabling more intuitive freehand drawing experiences.

Abstract: Interactive 3D model texture editing presents enhanced opportunities for creating 3D assets, with freehand drawing style offering the most intuitive experience. However, existing methods primarily support sketch-based interactions for outlining, while the utilization of coarse-grained scribble-based interaction remains limited. Furthermore, current methodologies often encounter challenges due to the abstract nature of scribble instructions, which can result in ambiguous editing intentions and unclear target semantic locations. To address these issues, we propose ScribbleSense, an editing method that combines multimodal large language models (MLLMs) and image generation models to effectively resolve these challenges. We leverage the visual capabilities of MLLMs to predict the editing intent behind the scribbles. Once the semantic intent of the scribble is discerned, we employ globally generated images to extract local texture details, thereby anchoring local semantics and alleviating ambiguities concerning the target semantic locations. Experimental results indicate that our method effectively leverages the strengths of MLLMs, achieving state-of-the-art interactive editing performance for scribble-based texture editing.

Method: Introduces a guidance scheme using a representation alignment projector that injects predicted representations into intermediate sampling steps, providing semantic anchors without modifying model architecture. Applied to SiTs and REPAs diffusion transformer models.

Result: Significant improvements in class-conditional ImageNet synthesis with substantially lower FID scores: REPA-XL/2 improves from 5.9 to 3.3. Outperforms representative guidance on SiT models and yields complementary gains when combined with classifier-free guidance.

Conclusion: Representation-informed diffusion sampling is a practical strategy for reinforcing semantic preservation and image consistency in diffusion-based generative models, enhancing visual fidelity without architectural modifications.

Abstract: Recent progress in generative modeling has enabled high-quality visual synthesis with diffusion-based frameworks, supporting controllable sampling and large-scale training. Inference-time guidance methods such as classifier-free and representative guidance enhance semantic alignment by modifying sampling dynamics; however, they do not fully exploit unsupervised feature representations. Although such visual representations contain rich semantic structure, their integration during generation is constrained by the absence of ground-truth reference images at inference. This work reveals semantic drift in the early denoising stages of diffusion transformers, where stochasticity results in inconsistent alignment even under identical conditioning. To mitigate this issue, we introduce a guidance scheme using a representation alignment projector that injects representations predicted by a projector into intermediate sampling steps, providing an effective semantic anchor without modifying the model architecture. Experiments on SiTs and REPAs show notable improvements in class-conditional ImageNet synthesis, achieving substantially lower FID scores; for example, REPA-XL/2 improves from 5.9 to 3.3, and the proposed method outperforms representative guidance when applied to SiT models. The approach further yields complementary gains when combined with classifier-free guidance, demonstrating enhanced semantic coherence and visual fidelity. These results establish representation-informed diffusion sampling as a practical strategy for reinforcing semantic preservation and image consistency.

Method: HAVC filters attention heads through OCR-based diagnostic tasks, then refines them using spatial entropy for spatial concentration and gradient sensitivity for predictive contribution. The fused signals produce a Visual Cropping Guidance Map that highlights task-relevant regions to crop subimages for MLLM input.

Result: Extensive experiments on multiple fine-grained VQA benchmarks show HAVC consistently outperforms state-of-the-art cropping strategies, achieving more precise localization and stronger visual grounding.

Conclusion: HAVC provides a simple yet effective training-free strategy for enhancing precision and visual grounding in MLLMs for fine-grained reasoning tasks.

Abstract: Multimodal Large Language Models (MLLMs) show strong performance in Visual Question Answering (VQA) but remain limited in fine-grained reasoning due to low-resolution inputs and noisy attention aggregation. We propose \textbf{Head Aware Visual Cropping (HAVC)}, a training-free method that improves visual grounding by leveraging a selectively refined subset of attention heads. HAVC first filters heads through an OCR-based diagnostic task, ensuring that only those with genuine grounding ability are retained. At inference, these heads are further refined using spatial entropy for stronger spatial concentration and gradient sensitivity for predictive contribution. The fused signals produce a reliable Visual Cropping Guidance Map, which highlights the most task-relevant region and guides the cropping of a subimage subsequently provided to the MLLM together with the image-question pair. Extensive experiments on multiple fine-grained VQA benchmarks demonstrate that HAVC consistently outperforms state-of-the-art cropping strategies, achieving more precise localization, stronger visual grounding, providing a simple yet effective strategy for enhancing precision in MLLMs.

Method: Proposes PromptMAD framework using CLIP-encoded text prompts describing normal/anomalous characteristics, Focal loss for class imbalance, supervised segmentor with multi-scale convolutional features, Transformer spatial attention, and diffusion iterative refinement.

Result: Achieves state-of-the-art pixel-level performance on MVTec-AD dataset with 98.35% mean AUC and 66.54% AP, maintaining efficiency across diverse categories.

Conclusion: Cross-modal prompting with vision-language alignment effectively addresses visual anomaly detection challenges, improving detection of subtle and textural anomalies through semantic guidance.

Abstract: Visual anomaly detection in multi-class settings poses significant challenges due to the diversity of object categories, the scarcity of anomalous examples, and the presence of camouflaged defects. In this paper, we propose PromptMAD, a cross-modal prompting framework for unsupervised visual anomaly detection and localization that integrates semantic guidance through vision-language alignment. By leveraging CLIP-encoded text prompts describing both normal and anomalous class-specific characteristics, our method enriches visual reconstruction with semantic context, improving the detection of subtle and textural anomalies. To further address the challenge of class imbalance at the pixel level, we incorporate Focal loss function, which emphasizes hard-to-detect anomalous regions during training. Our architecture also includes a supervised segmentor that fuses multi-scale convolutional features with Transformer-based spatial attention and diffusion iterative refinement, yielding precise and high-resolution anomaly maps. Extensive experiments on the MVTec-AD dataset demonstrate that our method achieves state-of-the-art pixel-level performance, improving mean AUC to 98.35% and AP to 66.54%, while maintaining efficiency across diverse categories.

Method: DreamVAR uses a VAR model with next-scale prediction. It extracts multi-scale subject features with a visual tokenizer, then pre-fills the full subject feature sequence before predicting target image tokens (instead of interleaving). This simplifies autoregressive dependencies and reduces train-test discrepancy. The framework also incorporates reinforcement learning to enhance semantic alignment and subject consistency.

Result: Extensive experiments demonstrate that DreamVAR achieves superior appearance preservation compared to leading diffusion-based methods in subject-driven image generation tasks.

Conclusion: DreamVAR successfully demonstrates the potential of VAR models for subject-driven image synthesis, offering a novel approach that outperforms diffusion-based methods in appearance preservation through its pre-filled feature design and reinforcement learning integration.

Abstract: Recent advances in subject-driven image generation using diffusion models have attracted considerable attention for their remarkable capabilities in producing high-quality images. Nevertheless, the potential of Visual Autoregressive (VAR) models, despite their unified architecture and efficient inference, remains underexplored. In this work, we present DreamVAR, a novel framework for subject-driven image synthesis built upon a VAR model that employs next-scale prediction. Technically, multi-scale features of the reference subject are first extracted by a visual tokenizer. Instead of interleaving these conditional features with target image tokens across scales, our DreamVAR pre-fills the full subject feature sequence prior to predicting target image tokens. This design simplifies autoregressive dependencies and mitigates the train-test discrepancy in multi-scale conditioning scenario within the VAR paradigm. DreamVAR further incorporates reinforcement learning to jointly enhance semantic alignment and subject consistency. Extensive experiments demonstrate that DreamVAR achieves superior appearance preservation compared to leading diffusion-based methods.

Method: Proposed AVT Compositional Fusion integrates video, audio, and text features by selectively aligning textual queries to the most relevant modality. Created AV-Comp benchmark with video pairs having cross-modal changes and corresponding textual queries describing differences.

Result: AVT outperforms traditional unimodal fusion methods and serves as a strong baseline for the new CoVA task. The AV-Comp benchmark provides a testbed for multimodal video-audio-text retrieval.

Conclusion: CoVA addresses the limitation of existing video retrieval by incorporating audio, enabling more comprehensive multimodal understanding and retrieval of videos based on both visual and auditory characteristics.

Abstract: Composed Video Retrieval (CoVR) aims to retrieve a target video from a large gallery using a reference video and a textual query specifying visual modifications. However, existing benchmarks consider only visual changes, ignoring videos that differ in audio despite visual similarity. To address this limitation, we introduce Composed retrieval for Video with its Audio CoVA, a new retrieval task that accounts for both visual and auditory variations. To support this, we construct AV-Comp, a benchmark consisting of video pairs with cross-modal changes and corresponding textual queries that describe the differences. We also propose AVT Compositional Fusion (AVT), which integrates video, audio, and text features by selectively aligning the query to the most relevant modality. AVT outperforms traditional unimodal fusion and serves as a strong baseline for CoVA. Examples from the proposed dataset, including both visual and auditory information, are available at https://perceptualai-lab.github.io/CoVA/.

Method: Proposes discriminative neural anchors (DNA) framework with coarse-to-fine excavation: 1) Analyzes feature decoupling and attention distribution to identify critical layers where focus shifts from global semantics to local anomalies, 2) Uses triadic fusion scoring with curvature-truncation to isolate forgery-discriminative units (FDUs), 3) Introduces HIFI-Gen benchmark with latest generative models.

Result: DNA achieves superior detection performance even under few-shot conditions, shows remarkable robustness across diverse architectures and against unseen generative models, demonstrating that extracting latent neurons is more effective than extensive fine-tuning.

Conclusion: Forgery detection capability is inherently present in pre-trained models and can be effectively extracted through the DNA framework without resource-intensive fine-tuning, offering a more efficient approach to AI-generated content detection.

Abstract: As generative AI achieves hyper-realism, superficial artifact detection has become obsolete. While prevailing methods rely on resource-intensive fine-tuning of black-box backbones, we propose that forgery detection capability is already encoded within pre-trained models rather than requiring end-to-end retraining. To elicit this intrinsic capability, we propose the discriminative neural anchors (DNA) framework, which employs a coarse-to-fine excavation mechanism. First, by analyzing feature decoupling and attention distribution shifts, we pinpoint critical intermediate layers where the focus of the model logically transitions from global semantics to local anomalies. Subsequently, we introduce a triadic fusion scoring metric paired with a curvature-truncation strategy to strip away semantic redundancy, precisely isolating the forgery-discriminative units (FDUs) inherently imprinted with sensitivity to forgery traces. Moreover, we introduce HIFI-Gen, a high-fidelity synthetic benchmark built upon the very latest models, to address the lag in existing datasets. Experiments demonstrate that by solely relying on these anchors, DNA achieves superior detection performance even under few-shot conditions. Furthermore, it exhibits remarkable robustness across diverse architectures and against unseen generative models, validating that waking up latent neurons is more effective than extensive fine-tuning.

Method: Proposes training-time transform coding (TTC) with hierarchical design: channel-wise KLT for decorrelation and energy compaction, followed by sparsity-aware neural transform to reconstruct KLT residuals with minimal overhead.

Result: The method delivers strong rate-distortion performance with fast decoding, offering favorable BD-rate-decoding-time trade-off over state-of-the-art 3DGS compressors.

Conclusion: Adding neural analysis-synthesis transforms to 3DGS compression significantly improves compression efficiency by properly distributing the redundancy removal workload between transforms and entropy coding.

Abstract: Current 3DGS compression methods largely forego the neural analysis-synthesis transform, which is a crucial component in learned signal compression systems. As a result, redundancy removal is left solely to the entropy coder, overburdening the entropy coding module and reducing rate-distortion (R-D) performance. To fix this critical omission, we propose a training-time transform coding (TTC) method that adds the analysis-synthesis transform and optimizes it jointly with the 3DGS representation and entropy model. Concretely, we adopt a hierarchical design: a channel-wise KLT for decorrelation and energy compaction, followed by a sparsity-aware neural transform that reconstructs the KLT residuals with minimal parameter and computational overhead. Experiments show that our method delivers strong R-D performance with fast decoding, offering a favorable BD-rate-decoding-time trade-off over SOTA 3DGS compressors.

Method: Compared top-view depth images and point cloud data for BCS prediction under four settings: 1) unsegmented raw data, 2) segmented full-body data, 3) segmented hindquarter data, and 4) handcrafted feature data. Used data from 1,020 dairy cows with cow-level cross-validation to prevent data leakage.

Result: Depth image-based models consistently achieved higher accuracy than point cloud-based models with unsegmented raw data and segmented full-body data. Comparable performance was observed with segmented hindquarter data. Both approaches showed reduced accuracy with handcrafted features. Point cloud predictions were more sensitive to noise and model architecture.

Conclusion: 3D point clouds do not provide a consistent advantage over depth images for BCS prediction in dairy cattle under the evaluated conditions, with depth images showing more robust performance.

Abstract: Body condition score (BCS) is a widely used indicator of body energy status and is closely associated with metabolic status, reproductive performance, and health in dairy cattle; however, conventional visual scoring is subjective and labor-intensive. Computer vision approaches have been applied to BCS prediction, with depth images widely used because they capture geometric information independent of coat color and texture. More recently, three-dimensional point cloud data have attracted increasing interest due to their ability to represent richer geometric characteristics of animal morphology, but direct head-to-head comparisons with depth image-based approaches remain limited. In this study, we compared top-view depth image and point cloud data for BCS prediction under four settings: 1) unsegmented raw data, 2) segmented full-body data, 3) segmented hindquarter data, and 4) handcrafted feature data. Prediction models were evaluated using data from 1,020 dairy cows collected on a commercial farm, with cow-level cross-validation to prevent data leakage. Depth image-based models consistently achieved higher accuracy than point cloud-based models when unsegmented raw data and segmented full-body data were used, whereas comparable performance was observed when segmented hindquarter data were used. Both depth image and point cloud approaches showed reduced accuracy when handcrafted feature data were employed compared with the other settings. Overall, point cloud-based predictions were more sensitive to noise and model architecture than depth image-based predictions. Taken together, these results indicate that three-dimensional point clouds do not provide a consistent advantage over depth images for BCS prediction in dairy cattle under the evaluated conditions.

Method: SHED uses a novel encoder-decoder architecture with bidirectional hierarchical reasoning. Segment tokens are hierarchically pooled in the encoder and unpooled in the decoder to reverse the hierarchy. The model is supervised only at the final output, allowing segment hierarchy to emerge without explicit segmentation supervision.

Result: SHED improves depth boundary sharpness and segment coherence, demonstrates strong cross-domain generalization from synthetic to real-world environments, better captures global 3D scene layouts leading to improved semantic segmentation, enhances 3D reconstruction quality, and reveals interpretable part-level structures.

Conclusion: Incorporating segmentation into dense prediction through hierarchical reasoning effectively enforces geometric priors, addressing structural inconsistencies in conventional pixel-wise methods and improving performance across multiple 3D perception tasks.

Abstract: Dense prediction infers per-pixel values from a single image and is fundamental to 3D perception and robotics. Although real-world scenes exhibit strong structure, existing methods treat it as an independent pixel-wise prediction, often resulting in structural inconsistencies. We propose SHED, a novel encoder-decoder architecture that enforces geometric prior explicitly by incorporating segmentation into dense prediction. By bidirectional hierarchical reasoning, segment tokens are hierarchically pooled in the encoder and unpooled in the decoder to reverse the hierarchy. The model is supervised only at the final output, allowing the segment hierarchy to emerge without explicit segmentation supervision. SHED improves depth boundary sharpness and segment coherence, while demonstrating strong cross-domain generalization from synthetic to the real-world environments. Its hierarchy-aware decoder better captures global 3D scene layouts, leading to improved semantic segmentation performance. Moreover, SHED enhances 3D reconstruction quality and reveals interpretable part-level structures that are often missed by conventional pixel-wise methods.

Method: Integrates shared retrieval encoder with two localization branches: 1) classical geometric branch using feature fusion and PnP, and 2) neural feed-forward branch (MapAnything) for metric localization. Includes neural-guided candidate pruning based on translation consistency and depth-conditioned localization for scale/translation refinement.

Result: Achieved significant improvements in recall and accuracy across HYDRO and SUCCU benchmarks, with final score of 92.62 (R@0.5m, 5°) in the CroCoDL 2025 Challenge.

Conclusion: The hybrid approach combining classical geometric and neural methods with specialized refinement techniques provides effective cross-device localization, demonstrating the value of complementary techniques for robust performance.

Abstract: We present a hybrid cross-device localization pipeline developed for the CroCoDL 2025 Challenge. Our approach integrates a shared retrieval encoder and two complementary localization branches: a classical geometric branch using feature fusion and PnP, and a neural feed-forward branch (MapAnything) for metric localization conditioned on geometric inputs. A neural-guided candidate pruning strategy further filters unreliable map frames based on translation consistency, while depth-conditioned localization refines metric scale and translation precision on Spot scenes. These components jointly lead to significant improvements in recall and accuracy across both HYDRO and SUCCU benchmarks. Our method achieved a final score of 92.62 (R@0.5m, 5°) during the challenge.

Method: Proposes MA-PaPSP (Memory-Augmented PaPSP) which augments basic PaPSP with a retrieval dataset of image-text pairs. Uses retrieved nearest-neighbor pairs to reduce embedding variance through averaging, and applies contrastive normalization to improve score calibration.

Result: MA-PaPSP outperforms PaPSP and other selective prediction baselines for selective captioning, image-text matching, and fine-grained classification across multiple datasets.

Conclusion: Memory augmentation with retrieval datasets effectively addresses embedding instability and calibration issues in selective prediction for visual language foundation models, enabling reliable confidence estimation in open-set scenarios.

Abstract: Selective prediction aims to endow predictors with a reject option, to avoid low confidence predictions. However, existing literature has primarily focused on closed-set tasks, such as visual question answering with predefined options or fixed-category classification. This paper considers selective prediction for visual language foundation models, addressing a taxonomy of tasks ranging from closed to open set and from finite to unbounded vocabularies, as in image captioning. We seek training-free approaches of low-complexity, applicable to any foundation model and consider methods based on external vision-language model embeddings, like CLIP. This is denoted as Plug-and-Play Selective Prediction (PaPSP). We identify two key challenges: (1) instability of the visual-language representations, leading to high variance in image-text embeddings, and (2) poor calibration of similarity scores. To address these issues, we propose a memory augmented PaPSP (MA-PaPSP) model, which augments PaPSP with a retrieval dataset of image-text pairs. This is leveraged to reduce embedding variance by averaging retrieved nearest-neighbor pairs and is complemented by the use of contrastive normalization to improve score calibration. Through extensive experiments on multiple datasets, we show that MA-PaPSP outperforms PaPSP and other selective prediction baselines for selective captioning, image-text matching, and fine-grained classification. Code is publicly available at https://github.com/kingston-aditya/MA-PaPSP.

Method: DELNet integrates a judging valve that measures task similarity to distinguish new from known tasks, and a dynamic expert library that stores experts trained on different degradations. For new tasks, the valve selects top-k experts for knowledge transfer while adding new experts; for known tasks, corresponding experts are directly reused.

Result: Experiments on OTS, Rain100H, and Snow100K datasets show DELNet surpasses state-of-the-art continual learning methods with PSNR gains of 16%, 11%, and 12% respectively.

Conclusion: DELNet demonstrates effectiveness, robustness, and efficiency for weather image restoration, reducing retraining costs and enabling practical deployment in real-world scenarios.

Abstract: All-in-one weather image restoration methods are valuable in practice but depend on pre-collected data and require retraining for unseen degradations, leading to high cost. We propose DELNet, a continual learning framework for weather image restoration. DELNet integrates a judging valve that measures task similarity to distinguish new from known tasks, and a dynamic expert library that stores experts trained on different degradations. For new tasks, the valve selects top-k experts for knowledge transfer while adding new experts to capture task-specific features; for known tasks, the corresponding experts are directly reused. This design enables continuous optimization without retraining existing models. Experiments on OTS, Rain100H, and Snow100K demonstrate that DELNet surpasses state-of-the-art continual learning methods, achieving PSNR gains of 16%, 11%, and 12%, respectively. These results highlight the effectiveness, robustness, and efficiency of DELNet, which reduces retraining cost and enables practical deployment in real-world scenarios.

Method: Proposes Spatiotemporal-Semantic Contrastive Decoding: 1) Constructs negative features by deliberately disrupting spatiotemporal consistency and semantic associations of video features, 2) Uses contrastive decoding against original video features during inference to suppress hallucinations.

Result: Extensive experiments demonstrate the method effectively mitigates hallucination occurrences while preserving the model’s general video understanding and reasoning capabilities.

Conclusion: The proposed decoding strategy provides a more effective approach to mitigate video hallucinations in Video Large Language Models by addressing spatiotemporal and semantic correlations through contrastive decoding.

Abstract: Although Video Large Language Models perform remarkably well across tasks such as video understanding, question answering, and reasoning, they still suffer from the problem of hallucination, which refers to generating outputs that are inconsistent with explicit video content or factual evidence. However, existing decoding methods for mitigating video hallucinations, while considering the spatiotemporal characteristics of videos, mostly rely on heuristic designs. As a result, they fail to precisely capture the root causes of hallucinations and their fine-grained temporal and semantic correlations, leading to limited robustness and generalization in complex scenarios. To more effectively mitigate video hallucinations, we propose a novel decoding strategy termed Spatiotemporal-Semantic Contrastive Decoding. This strategy constructs negative features by deliberately disrupting the spatiotemporal consistency and semantic associations of video features, and suppresses video hallucinations through contrastive decoding against the original video features during inference. Extensive experiments demonstrate that our method not only effectively mitigates the occurrence of hallucinations, but also preserves the general video understanding and reasoning capabilities of the model.

Method: Introduced PhoStream benchmark with 5,572 open-ended QA pairs from 578 videos across 4 scenarios and 10 capabilities. Built using Automated Generative Pipeline with human verification, evaluated with Online Inference Pipeline and LLM-as-a-Judge for open-ended responses.

Result: Models show temporal asymmetry: perform well on Instant and Backward tasks (Gemini 3 Pro >80), but drop sharply on Forward tasks (16.40) due to early responses before required cues appear. Reveals fundamental limitation in timing decisions.

Conclusion: Current MLLMs struggle to decide when to speak, not just what to say. PhoStream provides crucial evaluation for real-world mobile assistant capabilities in streaming audio-visual contexts.

Abstract: Multimodal Large Language Models excel at offline audio-visual understanding, but their ability to serve as mobile assistants in continuous real-world streams remains underexplored. In daily phone use, mobile assistants must track streaming audio-visual inputs and respond at the right time, yet existing benchmarks are often restricted to multiple-choice questions or use shorter videos. In this paper, we introduce PhoStream, the first mobile-centric streaming benchmark that unifies on-screen and off-screen scenarios to evaluate video, audio, and temporal reasoning. PhoStream contains 5,572 open-ended QA pairs from 578 videos across 4 scenarios and 10 capabilities. We build it with an Automated Generative Pipeline backed by rigorous human verification, and evaluate models using a realistic Online Inference Pipeline and LLM-as-a-Judge evaluation for open-ended responses. Experiments reveal a temporal asymmetry in LLM-judged scores (0-100): models perform well on Instant and Backward tasks (Gemini 3 Pro exceeds 80), but drop sharply on Forward tasks (16.40), largely due to early responses before the required visual and audio cues appear. This highlights a fundamental limitation: current MLLMs struggle to decide when to speak, not just what to say. Code and datasets used in this work will be made publicly accessible at https://github.com/Lucky-Lance/PhoStream.

Method: Proposes MIFOMO built on a remote sensing foundation model pre-trained across large-scale RS problems. Uses coalescent projection to quickly adapt the foundation model while freezing the backbone, mixup domain adaptation to handle extreme domain discrepancies, and label smoothing for noisy pseudo-labels.

Result: MIFOMO beats prior arts with up to 14% margin improvement in performance, demonstrating significant advantages over existing methods.

Conclusion: The foundation model approach with mixup domain adaptation and coalescent projection effectively addresses CDFSL challenges in HSI classification, outperforming previous methods by substantial margins.

Abstract: Although cross-domain few-shot learning (CDFSL) for hyper-spectral image (HSI) classification has attracted significant research interest, existing works often rely on an unrealistic data augmentation procedure in the form of external noise to enlarge the sample size, thus greatly simplifying the issue of data scarcity. They involve a large number of parameters for model updates, being prone to the overfitting problem. To the best of our knowledge, none has explored the strength of the foundation model, having strong generalization power to be quickly adapted to downstream tasks. This paper proposes the MIxup FOundation MOdel (MIFOMO) for CDFSL of HSI classifications. MIFOMO is built upon the concept of a remote sensing (RS) foundation model, pre-trained across a large scale of RS problems, thus featuring generalizable features. The notion of coalescent projection (CP) is introduced to quickly adapt the foundation model to downstream tasks while freezing the backbone network. The concept of mixup domain adaptation (MDM) is proposed to address the extreme domain discrepancy problem. Last but not least, the label smoothing concept is implemented to cope with noisy pseudo-label problems. Our rigorous experiments demonstrate the advantage of MIFOMO, where it beats prior arts with up to 14% margin. The source code of MIFOMO is open-sourced in https://github.com/Naeem- Paeedeh/MIFOMO for reproducibility and convenient further study.

Method: Created dataset from French orthophotos and topographic building data spanning 28 departments, with 25 for training and 3 geographically disjoint departments for evaluation. Includes binary building change masks and instance-level annotations.

Result: Released FOTBCD-Binary (28,000 before/after image pairs) and FOTBCD-Instances (instance-level subset). Benchmarking shows geographic diversity improves cross-domain generalization compared to LEVIR-CD+ and WHU-CD.

Conclusion: FOTBCD enables large-scale benchmarking for building change detection with geographic domain shift, demonstrating that dataset-level geographic diversity enhances cross-domain generalization.

Abstract: We introduce FOTBCD, a large-scale building change detection dataset derived from authoritative French orthophotos and topographic building data provided by IGN France. Unlike existing benchmarks that are geographically constrained to single cities or limited regions, FOTBCD spans 28 departments across mainland France, with 25 used for training and three geographically disjoint departments held out for evaluation. The dataset covers diverse urban, suburban, and rural environments at 0.2m/pixel resolution. We publicly release FOTBCD-Binary, a dataset comprising approximately 28,000 before/after image pairs with pixel-wise binary building change masks, each associated with patch-level spatial metadata. The dataset is designed for large-scale benchmarking and evaluation under geographic domain shift, with validation and test samples drawn from held-out departments and manually verified to ensure label quality. In addition, we publicly release FOTBCD-Instances, a publicly available instance-level annotated subset comprising several thousand image pairs, which illustrates the complete annotation schema used in the full instance-level version of FOTBCD. Using a fixed reference baseline, we benchmark FOTBCD-Binary against LEVIR-CD+ and WHU-CD, providing strong empirical evidence that geographic diversity at the dataset level is associated with improved cross-domain generalization in building change detection.

Method: Proposes TTSA3R with two complementary modules: 1) Temporal Adaptive Update Module regulates update magnitude by analyzing temporal state evolution patterns, and 2) Spatial Contextual Update Module localizes spatial regions requiring updates through observation-state alignment and scene dynamics. These signals are fused to determine state updating strategies.

Result: Extensive experiments demonstrate effectiveness in diverse 3D tasks. The method exhibits only 15% error increase compared to over 200% degradation in baseline models on extended sequences, significantly improving long-term reconstruction stability.

Conclusion: TTSA3R effectively addresses catastrophic forgetting in streaming 3D reconstruction by leveraging both temporal state evolution and spatial observation quality for adaptive state updates, achieving superior long-term stability.

Abstract: Streaming recurrent models enable efficient 3D reconstruction by maintaining persistent state representations. However, they suffer from catastrophic memory forgetting over long sequences due to balancing historical information with new observations. Recent methods alleviate this by deriving adaptive signals from attention perspective, but they operate on single dimensions without considering temporal and spatial consistency. To this end, we propose a training-free framework termed TTSA3R that leverages both temporal state evolution and spatial observation quality for adaptive state updates in 3D reconstruction. In particular, we devise a Temporal Adaptive Update Module that regulates update magnitude by analyzing temporal state evolution patterns. Then, a Spatial Contextual Update Module is introduced to localize spatial regions that require updates through observation-state alignment and scene dynamics. These complementary signals are finally fused to determine the state updating strategies. Extensive experiments demonstrate the effectiveness of TTSA3R in diverse 3D tasks. Moreover, our method exhibits only 15% error increase compared to over 200% degradation in baseline models on extended sequences, significantly improving long-term reconstruction stability. Our codes will be available soon.

Method: Proposes UniGeo with two key components: 1) Geometry-aware learning module that establishes learnable mapping from spatial relationships to feature weights for explicit geometric feature enhancement. 2) Dynamic channel gating mechanism using learnable channel-wise weighting to adaptively optimize features from sparse 3D U-Net network.

Result: Extensive experiments on six different indoor scene datasets validate superior performance of the proposed method.

Conclusion: UniGeo effectively addresses geometric relationship modeling in sparse point cloud scenes and enhances feature representation for improved 3D indoor object detection.

Abstract: The growing adoption of robotics and augmented reality in real-world applications has driven considerable research interest in 3D object detection based on point clouds. While previous methods address unified training across multiple datasets, they fail to model geometric relationships in sparse point cloud scenes and ignore the feature distribution in significant areas, which ultimately restricts their performance. To deal with this issue, a unified 3D indoor detection framework, called UniGeo, is proposed. To model geometric relations in scenes, we first propose a geometry-aware learning module that establishes a learnable mapping from spatial relationships to feature weights, which enabes explicit geometric feature enhancement. Then, to further enhance point cloud feature representation, we propose a dynamic channel gating mechanism that leverages learnable channel-wise weighting. This mechanism adaptively optimizes features generated by the sparse 3D U-Net network, significantly enhancing key geometric information. Extensive experiments on six different indoor scene datasets clearly validate the superior performance of our method.

Method: Systematic empirical analysis of scaling behavior with different design choices: normalization paradigms (division vs subtraction), depthwise convolution for locality, and extension of gating mechanisms to bidirectional setting with KV gate. LINA model built entirely on linear attention.

Result: LINA achieves competitive performance: 2.18 FID on ImageNet (1.4B params) and 0.74 on GenEval (1.5B params). Single linear attention module reduces FLOPs by ~61% compared to softmax attention.

Conclusion: Division-based normalization scales better than subtraction-based for linear generative transformers. Locality modeling via convolution is crucial for autoregressive generation. KV gating enables flexible memory management. LINA demonstrates efficient high-fidelity image generation.

Abstract: Autoregressive models with continuous tokens form a promising paradigm for visual generation, especially for text-to-image (T2I) synthesis, but they suffer from high computational cost. We study how to design compute-efficient linear attention within this framework. Specifically, we conduct a systematic empirical analysis of scaling behavior with respect to parameter counts under different design choices, focusing on (1) normalization paradigms in linear attention (division-based vs. subtraction-based) and (2) depthwise convolution for locality augmentation. Our results show that although subtraction-based normalization is effective for image classification, division-based normalization scales better for linear generative transformers. In addition, incorporating convolution for locality modeling plays a crucial role in autoregressive generation, consistent with findings in diffusion models. We further extend gating mechanisms, commonly used in causal linear attention, to the bidirectional setting and propose a KV gate. By introducing data-independent learnable parameters to the key and value states, the KV gate assigns token-wise memory weights, enabling flexible memory management similar to forget gates in language models. Based on these findings, we present LINA, a simple and compute-efficient T2I model built entirely on linear attention, capable of generating high-fidelity 1024x1024 images from user instructions. LINA achieves competitive performance on both class-conditional and T2I benchmarks, obtaining 2.18 FID on ImageNet (about 1.4B parameters) and 0.74 on GenEval (about 1.5B parameters). A single linear attention module reduces FLOPs by about 61 percent compared to softmax attention. Code and models are available at: https://github.com/techmonsterwang/LINA.

Method: The paper adapts S.R. Ranganathan’s classification principles to create the vTelos CV annotation methodology, which provides a systematic framework for designing high-quality computer vision datasets.

Result: Experimental evidence shows improvements in CV annotation quality and accuracy when using the vTelos methodology, validating its effectiveness in addressing the Semantic Gap Problem.

Conclusion: Ranganathan’s classification principles offer a valuable theoretical foundation for addressing the Semantic Gap Problem in computer vision, and the vTelos methodology provides practical improvements in dataset design and annotation quality.

Abstract: The Semantic Gap Problem (SGP) in Computer Vision (CV) arises from the misalignment between visual and lexical semantics leading to flawed CV dataset design and CV benchmarks. This paper proposes that classification principles of S.R. Ranganathan can offer a principled starting point to address SGP and design high-quality CV datasets. We elucidate how these principles, suitably adapted, underpin the vTelos CV annotation methodology. The paper also briefly presents experimental evidence showing improvements in CV annotation and accuracy, thereby, validating vTelos.

Method: Proposes Alignment and Disentanglement method: 1) basic concept alignment using contrastive self-supervised learning (MoCo/DINO), 2) enhances attribution ability with Infomax loss for representation disentanglement, theoretically explained through canonical correlation analysis decomposition.

Result: On real-world benchmarks (AbC), the unsupervised method surprisingly outperforms supervised methods.

Conclusion: Provides a fresh perspective on synthetic image attribution by showing unsupervised methods can outperform supervised approaches, eliminating need for costly paired annotations.

Abstract: As the quality of synthetic images improves, identifying the underlying concepts of model-generated images is becoming increasingly crucial for copyright protection and ensuring model transparency. Existing methods achieve this attribution goal by training models using annotated pairs of synthetic images and their original training sources. However, obtaining such paired supervision is challenging, as it requires either well-designed synthetic concepts or precise annotations from millions of training sources. To eliminate the need for costly paired annotations, in this paper, we explore the possibility of unsupervised synthetic image attribution. We propose a simple yet effective unsupervised method called Alignment and Disentanglement. Specifically, we begin by performing basic concept alignment using contrastive self-supervised learning. Next, we enhance the model’s attribution ability by promoting representation disentanglement with the Infomax loss. This approach is motivated by an interesting observation: contrastive self-supervised models, such as MoCo and DINO, inherently exhibit the ability to perform simple cross-domain alignment. By formulating this observation as a theoretical assumption on cross-covariance, we provide a theoretical explanation of how alignment and disentanglement can approximate the concept-matching process through a decomposition of the canonical correlation analysis objective. On the real-world benchmarks, AbC, we show that our unsupervised method surprisingly outperforms the supervised methods. As a starting point, we expect our intuitive insights and experimental findings to provide a fresh perspective on this challenging task.

Method: Proposes ExpAlign with Expectation Alignment Head performing attention-based soft MIL pooling over token-region similarities. Uses energy-based multi-scale consistency regularization including Top-K multi-positive contrastive objective and Geometry-Aware Consistency Objective from Lagrangian-constrained free-energy minimization.

Result: Achieves 36.2 AP_r on LVIS minival split, outperforming other SOTA methods at comparable model scale. Consistently improves open-vocabulary detection and zero-shot instance segmentation, especially on long-tail categories. Remains lightweight and inference-efficient.

Conclusion: ExpAlign provides a theoretically grounded framework for vision-language alignment that enables implicit token and instance selection without additional annotations, achieving strong performance in open-vocabulary grounding tasks.

Abstract: Open-vocabulary grounding requires accurate vision-language alignment under weak supervision, yet existing methods either rely on global sentence embeddings that lack fine-grained expressiveness or introduce token-level alignment with explicit supervision or heavy cross-attention designs. We propose ExpAlign, a theoretically grounded vision-language alignment framework built on a principled multiple instance learning formulation. ExpAlign introduces an Expectation Alignment Head that performs attention-based soft MIL pooling over token-region similarities, enabling implicit token and instance selection without additional annotations. To further stabilize alignment learning, we develop an energy-based multi-scale consistency regularization scheme, including a Top-K multi-positive contrastive objective and a Geometry-Aware Consistency Objective derived from a Lagrangian-constrained free-energy minimization. Extensive experiments show that ExpAlign consistently improves open-vocabulary detection and zero-shot instance segmentation, particularly on long-tail categories. Most notably, it achieves 36.2 AP$_r$ on the LVIS minival split, outperforming other state-of-the-art methods at comparable model scale, while remaining lightweight and inference-efficient.

Method: Proposes VisionTrim with two plug-and-play modules: 1) Dominant Vision Token Selection (DVTS) preserves essential tokens via global-local view, and 2) Text-Guided Vision Complement (TGVC) enables context-aware token merging guided by textual cues.

Result: Extensive experiments across diverse image and video multimodal benchmarks demonstrate performance superiority, advancing practical MLLM deployment in real-world applications.

Conclusion: VisionTrim provides an effective training-free solution for accelerating MLLMs while maintaining performance through intelligent visual token reduction and text-guided merging.

Abstract: Multimodal large language models (MLLMs) suffer from high computational costs due to excessive visual tokens, particularly in high-resolution and video-based scenarios. Existing token reduction methods typically focus on isolated pipeline components and often neglect textual alignment, leading to performance degradation. In this paper, we propose VisionTrim, a unified framework for training-free MLLM acceleration, integrating two effective plug-and-play modules: 1) the Dominant Vision Token Selection (DVTS) module, which preserves essential visual tokens via a global-local view, and 2) the Text-Guided Vision Complement (TGVC) module, which facilitates context-aware token merging guided by textual cues. Extensive experiments across diverse image and video multimodal benchmarks demonstrate the performance superiority of our VisionTrim, advancing practical MLLM deployment in real-world applications. The code is available at: https://github.com/hanxunyu/VisionTrim.

Method: Proposes Pass-Bands Optimizer (PBO), a plug-and-play module with only two learnable parameters that optimizes the temporal pass-band toward task-relevant motion bands. Includes a lightweight consistency constraint to preserve semantics and boundaries, suppressing static components to effectively high-pass the stream.

Result: PBO yields over ten percentage points improvement on UCF101, and delivers consistent significant gains on more complex multi-modal action recognition and weakly supervised video anomaly detection tasks.

Conclusion: PBO offers a new perspective for SNN-based video processing and understanding by addressing the fundamental pass-band mismatch, enabling SNNs to better handle dynamic tasks with minimal computational overhead.

Abstract: Spiking neural networks (SNNs) have gained traction in vision due to their energy efficiency, bio-plausibility, and inherent temporal processing. Yet, despite this temporal capacity, most progress concentrates on static image benchmarks, and SNNs still underperform on dynamic video tasks compared to artificial neural networks (ANNs). In this work, we diagnose a fundamental pass-band mismatch: Standard spiking dynamics behave as a temporal low pass that emphasizes static content while attenuating motion bearing bands, where task relevant information concentrates in dynamic tasks. This phenomenon explains why SNNs can approach ANNs on static tasks yet fall behind on tasks that demand richer temporal understanding.To remedy this, we propose the Pass-Bands Optimizer (PBO), a plug-and-play module that optimizes the temporal pass-band toward task-relevant motion bands. PBO introduces only two learnable parameters, and a lightweight consistency constraint that preserves semantics and boundaries, incurring negligible computational overhead and requires no architectural changes. PBO deliberately suppresses static components that contribute little to discrimination, effectively high passing the stream so that spiking activity concentrates on motion bearing content. On UCF101, PBO yields over ten percentage points improvement. On more complex multi-modal action recognition and weakly supervised video anomaly detection, PBO delivers consistent and significant gains, offering a new perspective for SNN based video processing and understanding.

Method: Proposes Visual Personalization Turing Test (VPTT) paradigm with three components: 1) VPTT-Bench (10k-persona benchmark), 2) VPRAG (visual retrieval-augmented generator), and 3) VPTT Score (text-only metric calibrated against human and VLM judgments).

Result: High correlation across human, VLM, and VPTT evaluations, validating VPTT Score as reliable perceptual proxy. VPRAG achieves best alignment-originality balance for personalized generative AI.

Conclusion: VPTT offers scalable, privacy-safe foundation for evaluating and generating personalized visual content based on perceptual indistinguishability rather than identity replication.

Abstract: We introduce the Visual Personalization Turing Test (VPTT), a new paradigm for evaluating contextual visual personalization based on perceptual indistinguishability, rather than identity replication. A model passes the VPTT if its output (image, video, 3D asset, etc.) is indistinguishable to a human or calibrated VLM judge from content a given person might plausibly create or share. To operationalize VPTT, we present the VPTT Framework, integrating a 10k-persona benchmark (VPTT-Bench), a visual retrieval-augmented generator (VPRAG), and the VPTT Score, a text-only metric calibrated against human and VLM judgments. We show high correlation across human, VLM, and VPTT evaluations, validating the VPTT Score as a reliable perceptual proxy. Experiments demonstrate that VPRAG achieves the best alignment-originality balance, offering a scalable and privacy-safe foundation for personalized generative AI.

Method: 1) Synthesize region-level OOV prompts by sampling from low-likelihood regions of class-conditional Gaussian distributions in hidden space; 2) Mine pseudo-OOV images using Dirichlet-based gradient attribution that interprets attribution gradients as Dirichlet evidence to estimate prediction uncertainty; 3) Construct OOV decision boundary through low-density prior constraint using Gaussian kernel density estimation.

Result: Experimental results show significant improvement in OOV detection performance in zero-shot scenes compared to previous methods.

Conclusion: The proposed OOVDet framework effectively detects predefined classes while reliably rejecting undefined ones in zero-shot scenarios by leveraging low-density assumptions and uncertainty estimation techniques.

Abstract: Zero-shot out-of-vocabulary detection (ZS-OOVD) aims to accurately recognize objects of in-vocabulary (IV) categories provided at zero-shot inference, while simultaneously rejecting undefined ones (out-of-vocabulary, OOV) that lack corresponding category prompts. However, previous methods are prone to overfitting the IV classes, leading to the OOV or undefined classes being misclassified as IV ones with a high confidence score. To address this issue, this paper proposes a zero-shot OOV detector (OOVDet), a novel framework that effectively detects predefined classes while reliably rejecting undefined ones in zero-shot scenes. Specifically, due to the model’s lack of prior knowledge about the distribution of OOV data, we synthesize region-level OOV prompts by sampling from the low-likelihood regions of the class-conditional Gaussian distributions in the hidden space, motivated by the assumption that unknown semantics are more likely to emerge in low-density areas of the latent space. For OOV images, we further propose a Dirichlet-based gradient attribution mechanism to mine pseudo-OOV image samples, where the attribution gradients are interpreted as Dirichlet evidence to estimate prediction uncertainty, and samples with high uncertainty are selected as pseudo-OOV images. Building on these synthesized OOV prompts and pseudo-OOV images, we construct the OOV decision boundary through a low-density prior constraint, which regularizes the optimization of OOV classes using Gaussian kernel density estimation in accordance with the above assumption. Experimental results show that our method significantly improves the OOV detection performance in zero-shot scenes. The code is available at https://github.com/binyisu/OOV-detector.

Method: Uses a clean unified ViT-based model for coarse 3D geometry recovery, adds pixel-level supervision to optimize fine-grained details, and employs modular data annotation strategy to enrich training data and enhance robustness.

Result: Achieves over 100 FPS inference while substantially improving pose estimation accuracy on multiple benchmark datasets compared to previous SMPLX-based approaches.

Conclusion: PEAR provides a preprocessing-free, real-time framework for expressive human mesh recovery that addresses key limitations of existing methods in speed and fine-grained detail accuracy.

Abstract: Reconstructing detailed 3D human meshes from a single in-the-wild image remains a fundamental challenge in computer vision. Existing SMPLX-based methods often suffer from slow inference, produce only coarse body poses, and exhibit misalignments or unnatural artifacts in fine-grained regions such as the face and hands. These issues make current approaches difficult to apply to downstream tasks. To address these challenges, we propose PEAR-a fast and robust framework for pixel-aligned expressive human mesh recovery. PEAR explicitly tackles three major limitations of existing methods: slow inference, inaccurate localization of fine-grained human pose details, and insufficient facial expression capture. Specifically, to enable real-time SMPLX parameter inference, we depart from prior designs that rely on high resolution inputs or multi-branch architectures. Instead, we adopt a clean and unified ViT-based model capable of recovering coarse 3D human geometry. To compensate for the loss of fine-grained details caused by this simplified architecture, we introduce pixel-level supervision to optimize the geometry, significantly improving the reconstruction accuracy of fine-grained human details. To make this approach practical, we further propose a modular data annotation strategy that enriches the training data and enhances the robustness of the model. Overall, PEAR is a preprocessing-free framework that can simultaneously infer EHM-s (SMPLX and scaled-FLAME) parameters at over 100 FPS. Extensive experiments on multiple benchmark datasets demonstrate that our method achieves substantial improvements in pose estimation accuracy compared to previous SMPLX-based approaches. Project page: https://wujh2001.github.io/PEAR

Method: Reformulates vision-language alignment as conditional semantic comparison using bi-directional multiple-choice learning (Bi-MCQ) with Image-to-Text and Text-to-Image tasks, using affirmative/negative/mixed prompts and direction-specific Cross-Attention fusion modules

Result: Improves negation understanding by up to 0.47 AUC over zero-shot CARZero, achieves up to 0.08 gain on positive-negative combined evaluation, and reduces affirmative-negative AUC gap by average 0.12 compared to InfoNCE fine-tuning

Conclusion: Objective reformulation through conditional semantic comparison can substantially enhance negation understanding in medical vision-language models

Abstract: Recent vision-language models (VLMs) achieve strong zero-shot performance via large-scale image-text pretraining and have been widely adopted in medical image analysis. However, existing VLMs remain notably weak at understanding negated clinical statements, largely due to contrastive alignment objectives that treat negation as a minor linguistic variation rather than a meaning-inverting operator. In multi-label settings, prompt-based InfoNCE fine-tuning further reinforces easy-positive image-prompt alignments, limiting effective learning of disease absence. To overcome these limitations, we reformulate vision-language alignment as a conditional semantic comparison problem, which is instantiated through a bi-directional multiple-choice learning framework(Bi-MCQ). By jointly training Image-to-Text and Text-to-Image MCQ tasks with affirmative, negative, and mixed prompts, our method implements fine-tuning as conditional semantic comparison instead of global similarity maximization. We further introduce direction-specific Cross-Attention fusion modules to address asymmetric cues required by bi-directional reasoning and reduce alignment interference. Experiments on ChestXray14, Open-I, CheXpert, and PadChest show that Bi-MCQ improves negation understanding by up to 0.47 AUC over the zero-shot performance of the state-of-the-art CARZero model, while achieving up to a 0.08 absolute gain on positive-negative combined (PNC) evaluation. Additionally, Bi-MCQ reduces the affirmative-negative AUC gap by an average of 0.12 compared to InfoNCE-based fine-tuning, demonstrating that objective reformulation can substantially enhance negation understanding in medical VLMs.

Method: DAVIS is a post-hoc technique that enriches feature vectors by incorporating channel-wise variance and dominant (maximum) activations from activation maps before GAP, addressing the information loss from standard pooling operations.

Result: DAVIS achieves significant improvements across diverse architectures: 48.26% FPR95 reduction on CIFAR-10 with ResNet-18, 38.13% on CIFAR-100 with ResNet-34, and 26.83% on ImageNet-1k with MobileNet-v2, setting new benchmarks for OOD detection.

Conclusion: The overlooked statistics from activation maps before GAP are highly discriminative for OOD detection, and DAVIS provides a principled basis for moving beyond mean-based representations in OOD detection.

Abstract: Detecting out-of-distribution (OOD) inputs is a critical safeguard for deploying machine learning models in the real world. However, most post-hoc detection methods operate on penultimate feature representations derived from global average pooling (GAP) – a lossy operation that discards valuable distributional statistics from activation maps prior to global average pooling. We contend that these overlooked statistics, particularly channel-wise variance and dominant (maximum) activations, are highly discriminative for OOD detection. We introduce DAVIS, a simple and broadly applicable post-hoc technique that enriches feature vectors by incorporating these crucial statistics, directly addressing the information loss from GAP. Extensive evaluations show DAVIS sets a new benchmark across diverse architectures, including ResNet, DenseNet, and EfficientNet. It achieves significant reductions in the false positive rate (FPR95), with improvements of 48.26% on CIFAR-10 using ResNet-18, 38.13% on CIFAR-100 using ResNet-34, and 26.83% on ImageNet-1k benchmarks using MobileNet-v2. Our analysis reveals the underlying mechanism for this improvement, providing a principled basis for moving beyond the mean in OOD detection.

Method: Unifies knowledge distillation and quantization-aware training under Information Bottleneck principle. Uses confidence-gated decoupled distillation to filter unreliable supervision, relational centered kernel alignment to transfer visual token structures, and adaptive Lagrangian controller to balance fidelity against capacity constraints.

Result: INT4 models consistently outperform FP16 baselines (e.g., LLaVA-1.5-7B: 70.1 vs. 66.8 on SQA; Qwen2-VL-2B: 76.9 vs. 72.6 on MMBench), nearly matching teacher performance. Achieves 3× throughput with 54% memory reduction using real INT4 kernels.

Conclusion: GRACE provides a principled framework that significantly outperforms existing quantization methods, making it a compelling solution for resource-constrained deployment of Vision-Language Models.

Abstract: Vision-Language Models (VLMs) achieve strong multimodal performance but are costly to deploy, and post-training quantization often causes significant accuracy loss. Despite its potential, quantization-aware training for VLMs remains underexplored. We propose GRACE, a framework unifying knowledge distillation and QAT under the Information Bottleneck principle: quantization constrains information capacity while distillation guides what to preserve within this budget. Treating the teacher as a proxy for task-relevant information, we introduce confidence-gated decoupled distillation to filter unreliable supervision, relational centered kernel alignment to transfer visual token structures, and an adaptive controller via Lagrangian relaxation to balance fidelity against capacity constraints. Across extensive benchmarks on LLaVA and Qwen families, our INT4 models consistently outperform FP16 baselines (e.g., LLaVA-1.5-7B: 70.1 vs. 66.8 on SQA; Qwen2-VL-2B: 76.9 vs. 72.6 on MMBench), nearly matching teacher performance. Using real INT4 kernel, we achieve 3$\times$ throughput with 54% memory reduction. This principled framework significantly outperforms existing quantization methods, making GRACE a compelling solution for resource-constrained deployment.

Method: Constructed using DINOv3-based hierarchical clustering for balanced sampling and Gemini-powered dense captioning; proposed multi-modal evaluation protocol with VLM-based semantic reasoning and novel Multi-Scale Representation Metric based on SAM3 segmentation and morphological erosion.

Result: Strong agreement with human judgments (Kendall’s τ of 0.833 vs. 0.611 for SSIM), establishing a robust benchmark for VTON evaluation with 20 fine-grained garment categories.

Conclusion: OpenVTON-Bench provides a comprehensive, reliable benchmark for VTON systems with interpretable multi-dimensional evaluation that better aligns with human perception than traditional metrics.

Abstract: Recent advances in diffusion models have significantly elevated the visual fidelity of Virtual Try-On (VTON) systems, yet reliable evaluation remains a persistent bottleneck. Traditional metrics struggle to quantify fine-grained texture details and semantic consistency, while existing datasets fail to meet commercial standards in scale and diversity. We present OpenVTON-Bench, a large-scale benchmark comprising approximately 100K high-resolution image pairs (up to $1536 \times 1536$). The dataset is constructed using DINOv3-based hierarchical clustering for semantically balanced sampling and Gemini-powered dense captioning, ensuring a uniform distribution across 20 fine-grained garment categories. To support reliable evaluation, we propose a multi-modal protocol that measures VTON quality along five interpretable dimensions: background consistency, identity fidelity, texture fidelity, shape plausibility, and overall realism. The protocol integrates VLM-based semantic reasoning with a novel Multi-Scale Representation Metric based on SAM3 segmentation and morphological erosion, enabling the separation of boundary alignment errors from internal texture artifacts. Experimental results show strong agreement with human judgments (Kendall’s $τ$ of 0.833 vs. 0.611 for SSIM), establishing a robust benchmark for VTON evaluation.

Method: Uses memory-efficient continuous 3D Gaussian representation with four modules: 1) LiDAR Depth Feature Aggregation (LDFA) with depth-wise deformable sampling, 2) Entropy-Based Feature Smoothing (EBFS) to mitigate domain noise, 3) Adaptive Camera-LiDAR Fusion (ACLF) with uncertainty-aware reweighting, and 4) Gauss-Mamba Head leveraging Selective State Space Models for global context with linear complexity.

Result: Achieves state-of-the-art performance on Occ3D (49.4% mIoU), SurroundOcc (28.9% mIoU), and SemanticKITTI (25.2% mIoU) benchmarks, with superior robustness across challenging rainy and nighttime conditions.

Conclusion: GaussianOcc3D effectively bridges camera and LiDAR modalities through continuous 3D Gaussian representation, addressing key challenges in multi-modal 3D semantic occupancy prediction while maintaining computational efficiency.

Abstract: 3D semantic occupancy prediction is a pivotal task in autonomous driving, providing a dense and fine-grained understanding of the surrounding environment, yet single-modality methods face trade-offs between camera semantics and LiDAR geometry. Existing multi-modal frameworks often struggle with modality heterogeneity, spatial misalignment, and the representation crisis–where voxels are computationally heavy and BEV alternatives are lossy. We present GaussianOcc3D, a multi-modal framework bridging camera and LiDAR through a memory-efficient, continuous 3D Gaussian representation. We introduce four modules: (1) LiDAR Depth Feature Aggregation (LDFA), using depth-wise deformable sampling to lift sparse signals onto Gaussian primitives; (2) Entropy-Based Feature Smoothing (EBFS) to mitigate domain noise; (3) Adaptive Camera-LiDAR Fusion (ACLF) with uncertainty-aware reweighting for sensor reliability; and (4) a Gauss-Mamba Head leveraging Selective State Space Models for global context with linear complexity. Evaluations on Occ3D, SurroundOcc, and SemanticKITTI benchmarks demonstrate state-of-the-art performance, achieving mIoU scores of 49.4%, 28.9%, and 25.2% respectively. GaussianOcc3D exhibits superior robustness across challenging rainy and nighttime conditions.

Method: Proposes ImgCoT that replaces textual CoT reconstruction with visual CoT obtained by rendering reasoning steps into images, substituting linguistic bias with spatial inductive bias. Also introduces loose ImgCoT, a hybrid approach that augments visual latent tokens with key textual reasoning steps selected based on low token log-likelihood.

Result: Extensive experiments across multiple datasets and LLMs demonstrate the effectiveness of both versions of ImgCoT in capturing global reasoning structure while maintaining fine-grained details with fewer tokens than complete CoT.

Conclusion: Visual representations of reasoning chains enable better abstraction of global reasoning structure compared to text-based compression, and hybrid approaches combining visual structure with key textual details offer efficient reasoning compression.

Abstract: Compressing long chains of thought (CoT) into compact latent tokens is crucial for efficient reasoning with large language models (LLMs). Recent studies employ autoencoders to achieve this by reconstructing textual CoT from latent tokens, thus encoding CoT semantics. However, treating textual CoT as the reconstruction target forces latent tokens to preserve surface-level linguistic features (e.g., word choice and syntax), introducing a strong linguistic inductive bias that prioritizes linguistic form over reasoning structure and limits logical abstraction. Thus, we propose ImgCoT that replaces the reconstruction target from textual CoT to the visual CoT obtained by rendering CoT into images. This substitutes linguistic bias with spatial inductive bias, i.e., a tendency to model spatial layouts of the reasoning steps in visual CoT, enabling latent tokens to better capture global reasoning structure. Moreover, although visual latent tokens encode abstract reasoning structure, they may blur reasoning details. We thus propose a loose ImgCoT, a hybrid reasoning that augments visual latent tokens with a few key textual reasoning steps, selected based on low token log-likelihood. This design allows LLMs to retain both global reasoning structure and fine-grained reasoning details with fewer tokens than the complete CoT. Extensive experiments across multiple datasets and LLMs demonstrate the effectiveness of the two versions of ImgCoT.

Method: Created Lingua-SafetyBench with 100,440 harmful image-text pairs across 10 languages, explicitly partitioned into image-dominant and text-dominant subsets. Evaluated 11 open-source VLLMs and conducted controlled studies on the Qwen series to analyze safety performance across languages and modalities.

Result: Revealed consistent safety asymmetry: image-dominant risks yield higher attack success rates (ASR) in high-resource languages, while text-dominant risks are more severe in non-high-resource languages. Scaling and version upgrades reduce ASR overall but disproportionately benefit high-resource languages, widening the safety gap.

Conclusion: Current safety alignment approaches are insufficient for multilingual multimodal settings. Language- and modality-aware safety alignment beyond mere scaling is necessary to address the identified safety asymmetries across languages and modalities.

Abstract: Robust safety of vision-language large models (VLLMs) under joint multilingual and multimodal inputs remains underexplored. Existing benchmarks are typically multilingual but text-only, or multimodal but monolingual. Recent multilingual multimodal red-teaming efforts render harmful prompts into images, yet rely heavily on typography-style visuals and lack semantically grounded image-text pairs, limiting coverage of realistic cross-modal interactions. We introduce Lingua-SafetyBench, a benchmark of 100,440 harmful image-text pairs across 10 languages, explicitly partitioned into image-dominant and text-dominant subsets to disentangle risk sources. Evaluating 11 open-source VLLMs reveals a consistent asymmetry: image-dominant risks yield higher ASR in high-resource languages, while text-dominant risks are more severe in non-high-resource languages. A controlled study on the Qwen series shows that scaling and version upgrades reduce Attack Success Rate (ASR) overall but disproportionately benefit HRLs, widening the gap between HRLs and Non-HRLs under text-dominant risks. This underscores the necessity of language- and modality-aware safety alignment beyond mere scaling.To facilitate reproducibility and future research, we will publicly release our benchmark, model checkpoints, and source code.The code and dataset will be available at https://github.com/zsxr15/Lingua-SafetyBench.Warning: this paper contains examples with unsafe content.

Method: Proposes StreamSense with: 1) Lightweight streaming encoder for most timestamps, 2) Selective routing to VLM expert for hard/ambiguous cases, 3) Decision deferral when context is insufficient, 4) Training with cross-modal contrastive loss for audio/visual/text alignment and IoU-weighted loss to mitigate label interference across segment boundaries.

Result: Achieves higher accuracy than VLM-only streaming on social streaming detection tasks (sentiment classification, hate content moderation) while only occasionally invoking the VLM, reducing average latency and compute.

Conclusion: Selective escalation and deferral are effective primitives for understanding streaming social tasks, enabling efficient real-time multimodal analysis with improved accuracy-compute tradeoffs.

Abstract: Live streaming platforms require real-time monitoring and reaction to social signals, utilizing partial and asynchronous evidence from video, text, and audio. We propose StreamSense, a streaming detector that couples a lightweight streaming encoder with selective routing to a Vision-Language Model (VLM) expert. StreamSense handles most timestamps with the lightweight streaming encoder, escalates hard/ambiguous cases to the VLM, and defers decisions when context is insufficient. The encoder is trained using (i) a cross-modal contrastive term to align visual/audio cues with textual signals, and (ii) an IoU-weighted loss that down-weights poorly overlapping target segments, mitigating label interference across segment boundaries. We evaluate StreamSense on multiple social streaming detection tasks (e.g., sentiment classification and hate content moderation), and the results show that StreamSense achieves higher accuracy than VLM-only streaming while only occasionally invoking the VLM, thereby reducing average latency and compute. Our results indicate that selective escalation and deferral are effective primitives for understanding streaming social tasks. Code is publicly available on GitHub.

Method: Proposes FaceDefense with: 1) Diffusion loss to enhance adversarial example efficacy, 2) Directional facial attribute editing to restore perturbation-induced distortions, 3) Two-phase alternating optimization to generate final perturbed faces.

Result: Extensive experiments show FaceDefense significantly outperforms existing methods in both imperceptibility and defense effectiveness, achieving superior trade-off between protection and visual quality.

Conclusion: FaceDefense provides an effective proactive defense against diffusion-based face swapping attacks, addressing the core trade-off between protection strength and visual imperceptibility through novel diffusion loss and attribute editing techniques.

Abstract: Diffusion-based face swapping achieves state-of-the-art performance, yet it also exacerbates the potential harm of malicious face swapping to violate portraiture right or undermine personal reputation. This has spurred the development of proactive defense methods. However, existing approaches face a core trade-off: large perturbations distort facial structures, while small ones weaken protection effectiveness. To address these issues, we propose FaceDefense, an enhanced proactive defense framework against diffusion-based face swapping. Our method introduces a new diffusion loss to strengthen the defensive efficacy of adversarial examples, and employs a directional facial attribute editing to restore perturbation-induced distortions, thereby enhancing visual imperceptibility. A two-phase alternating optimization strategy is designed to generate final perturbed face images. Extensive experiments show that FaceDefense significantly outperforms existing methods in both imperceptibility and defense effectiveness, achieving a superior trade-off.

Method: Evaluated two Vision Language Models on extracting structured knowledge from troubleshooting guides. Compared two prompting strategies: standard instruction-guided prompting versus an augmented approach that cues troubleshooting layout patterns.

Result: Revealed model-specific trade-offs between layout sensitivity and semantic robustness, providing insights for practical deployment decisions.

Conclusion: VLMs show potential for automating knowledge extraction from industrial troubleshooting guides, but careful consideration of prompting strategies and model selection is needed based on specific requirements for layout understanding versus semantic accuracy.

Abstract: Industrial troubleshooting guides encode diagnostic procedures in flowchart-like diagrams where spatial layout and technical language jointly convey meaning. To integrate this knowledge into operator support systems, which assist shop-floor personnel in diagnosing and resolving equipment issues, the information must first be extracted and structured for machine interpretation. However, when performed manually, this extraction is labor-intensive and error-prone. Vision Language Models offer potential to automate this process by jointly interpreting visual and textual meaning, yet their performance on such guides remains underexplored. This paper evaluates two VLMs on extracting structured knowledge, comparing two prompting strategies: standard instruction-guided versus an augmented approach that cues troubleshooting layout patterns. Results reveal model-specific trade-offs between layout sensitivity and semantic robustness, informing practical deployment decisions.

Method: RAD stores anomaly-free features in a memory bank and detects anomalies through multi-level retrieval, matching test patches against the memory without any training. The approach is training-free and uses retrieval scores that theoretically upper-bound reconstruction-residual scores.

Result: RAD achieves state-of-the-art performance across four benchmarks (MVTec-AD, VisA, Real-IAD, 3D-ADAM) under standard and few-shot settings. On MVTec-AD, it reaches 96.7% Pixel AUROC with just one anomaly-free image.

Conclusion: The findings overturn the assumption that multi-class unsupervised anomaly detection requires task-specific training, showing that state-of-the-art performance is achievable with memory-based retrieval approaches.

Abstract: Current state-of-the-art multi-class unsupervised anomaly detection (MUAD) methods rely on training encoder-decoder models to reconstruct anomaly-free features. We first show these approaches have an inherent fidelity-stability dilemma in how they detect anomalies via reconstruction residuals. We then abandon the reconstruction paradigm entirely and propose Retrieval-based Anomaly Detection (RAD). RAD is a training-free approach that stores anomaly-free features in a memory and detects anomalies through multi-level retrieval, matching test patches against the memory. Experiments demonstrate that RAD achieves state-of-the-art performance across four established benchmarks (MVTec-AD, VisA, Real-IAD, 3D-ADAM) under both standard and few-shot settings. On MVTec-AD, RAD reaches 96.7% Pixel AUROC with just a single anomaly-free image compared to 98.5% of RAD’s full-data performance. We further prove that retrieval-based scores theoretically upper-bound reconstruction-residual scores. Collectively, these findings overturn the assumption that MUAD requires task-specific training, showing that state-of-the-art anomaly detection is feasible with memory-based retrieval. Our code is available at https://github.com/longkukuhi/RAD.

Method: Proposes Demosaicing-guided Color Correlation Training (DCCT) framework that simulates CFA sampling patterns to decompose color images into single-channel inputs and remaining channels as ground-truth targets. Uses self-supervised U-Net to model conditional distribution of missing channels via mixture of logistic functions.

Result: DCCT achieves state-of-the-art generalization and robustness, significantly outperforming prior methods across over 20 unseen generators. Theoretical analysis shows it targets provable distributional differences in color-correlation features.

Conclusion: Exploiting intrinsic camera imaging pipeline properties provides a more robust approach to AI-generated image detection that generalizes better across different generative models than artifact-based methods.

Abstract: As realistic AI-generated images threaten digital authenticity, we address the generalization failure of generative artifact-based detectors by exploiting the intrinsic properties of the camera imaging pipeline. Concretely, we investigate color correlations induced by the color filter array (CFA) and demosaicing, and propose a Demosaicing-guided Color Correlation Training (DCCT) framework for AI-generated image detection. By simulating the CFA sampling pattern, we decompose each color image into a single-channel input (as the condition) and the remaining two channels as the ground-truth targets (for prediction). A self-supervised U-Net is trained to model the conditional distribution of the missing channels from the given one, parameterized via a mixture of logistic functions. Our theoretical analysis reveals that DCCT targets a provable distributional difference in color-correlation features between photographic and AI-generated images. By leveraging these distinct features to construct a binary classifier, DCCT achieves state-of-the-art generalization and robustness, significantly outperforming prior methods across over 20 unseen generators.

Method: Fine-tune a state-of-the-art deep stereo network (MonSter) that leverages monocular depth priors on a curated dataset of diachronic image pairs from DFC2019 remote sensing challenge, containing both synchronic and diachronic pairs under diverse conditions.

Result: The approach consistently surpasses classical pipelines and unadapted deep stereo models on both synchronic and diachronic settings in experiments on multi-date WorldView-3 imagery, recovering accurate geometry despite strong appearance changes.

Conclusion: Fine-tuning on temporally diverse images with monocular priors enables 3D reconstruction from previously incompatible acquisition dates, addressing the diachronic stereo matching challenge in satellite imagery.

Abstract: Recent advances in image-based satellite 3D reconstruction have progressed along two complementary directions. On one hand, multi-date approaches using NeRF or Gaussian-splatting jointly model appearance and geometry across many acquisitions, achieving accurate reconstructions on opportunistic imagery with numerous observations. On the other hand, classical stereoscopic reconstruction pipelines deliver robust and scalable results for simultaneous or quasi-simultaneous image pairs. However, when the two images are captured months apart, strong seasonal, illumination, and shadow changes violate standard stereoscopic assumptions, causing existing pipelines to fail. This work presents the first Diachronic Stereo Matching method for satellite imagery, enabling reliable 3D reconstruction from temporally distant pairs. Two advances make this possible: (1) fine-tuning a state-of-the-art deep stereo network that leverages monocular depth priors, and (2) exposing it to a dataset specifically curated to include a diverse set of diachronic image pairs. In particular, we start from a pretrained MonSter model, trained initially on a mix of synthetic and real datasets such as SceneFlow and KITTI, and fine-tune it on a set of stereo pairs derived from the DFC2019 remote sensing challenge. This dataset contains both synchronic and diachronic pairs under diverse seasonal and illumination conditions. Experiments on multi-date WorldView-3 imagery demonstrate that our approach consistently surpasses classical pipelines and unadapted deep stereo models on both synchronic and diachronic settings. Fine-tuning on temporally diverse images, together with monocular priors, proves essential for enabling 3D reconstruction from previously incompatible acquisition dates. Left image (winter) Right image (autumn) DSM geometry Ours (1.23 m) Zero-shot (3.99 m) LiDAR GT Figure 1. Output geometry for a winter-autumn image pair from Omaha (OMA 331 test scene). Our method recovers accurate geometry despite the diachronic nature of the pair, exhibiting strong appearance changes, which cause existing zero-shot methods to fail. Missing values due to perspective shown in black. Mean altitude error in parentheses; lower is better.

Method: Proposes FarmMind framework with reasoning-query mechanism that: 1) analyzes root causes of segmentation ambiguities through reasoning, 2) determines what type of auxiliary image needs to be queried based on this analysis, and 3) dynamically queries external auxiliary images to compensate for insufficient information in single input images.

Result: Extensive experiments demonstrate superior segmentation performance and stronger generalization ability compared with existing methods. Source code and dataset are publicly available.

Conclusion: FarmMind breaks through limitations of static segmentation paradigm by introducing reasoning-query mechanism that mimics human expert thinking when faced with segmentation ambiguity, enabling more comprehensive reasoning through dynamic auxiliary image queries.

Abstract: Existing methods for farmland remote sensing image (FRSI) segmentation generally follow a static segmentation paradigm, where analysis relies solely on the limited information contained within a single input patch. Consequently, their reasoning capability is limited when dealing with complex scenes characterized by ambiguity and visual uncertainty. In contrast, human experts, when interpreting remote sensing images in such ambiguous cases, tend to actively query auxiliary images (such as higher-resolution, larger-scale, or temporally adjacent data) to conduct cross-verification and achieve more comprehensive reasoning. Inspired by this, we propose a reasoning-query-driven dynamic segmentation framework for FRSIs, named FarmMind. This framework breaks through the limitations of the static segmentation paradigm by introducing a reasoning-query mechanism, which dynamically and on-demand queries external auxiliary images to compensate for the insufficient information in a single input image. Unlike direct queries, this mechanism simulates the thinking process of human experts when faced with segmentation ambiguity: it first analyzes the root causes of segmentation ambiguities through reasoning, and then determines what type of auxiliary image needs to be queried based on this analysis. Extensive experiments demonstrate that FarmMind achieves superior segmentation performance and stronger generalization ability compared with existing methods. The source code and dataset used in this work are publicly available at: https://github.com/WithoutOcean/FarmMind.

Method: Systematic evaluation of ten representative LVLMs using the PeSOTIF dataset, which is specifically curated for long-tail traffic scenarios and environmental degradations. Performance is quantitatively compared against a classical YOLO-based detector approach.

Result: Top-performing LVLMs (e.g., Gemini 3, Doubao) surpass the YOLO baseline in recall by over 25% in complex natural scenarios, showing superior robustness to visual degradation. However, the baseline retains an advantage in geometric precision for synthetic perturbations, revealing a critical trade-off between semantic reasoning and geometric regression.

Conclusion: LVLMs demonstrate complementary strengths to conventional detectors, supporting their use as high-level safety validators in SOTIF-oriented automated driving systems. The findings highlight the value of semantic reasoning for robustness in adverse conditions while acknowledging the continued importance of geometric precision in certain scenarios.

Abstract: Reliable environmental perception remains one of the main obstacles for safe operation of automated vehicles. Safety of the Intended Functionality (SOTIF) concerns safety risks from perception insufficiencies, particularly under adverse conditions where conventional detectors often falter. While Large Vision-Language Models (LVLMs) demonstrate promising semantic reasoning, their quantitative effectiveness for safety-critical 2D object detection is underexplored. This paper presents a systematic evaluation of ten representative LVLMs using the PeSOTIF dataset, a benchmark specifically curated for long-tail traffic scenarios and environmental degradations. Performance is quantitatively compared against the classical perception approach, a YOLO-based detector. Experimental results reveal a critical trade-off: top-performing LVLMs (e.g., Gemini 3, Doubao) surpass the YOLO baseline in recall by over 25% in complex natural scenarios, exhibiting superior robustness to visual degradation. Conversely, the baseline retains an advantage in geometric precision for synthetic perturbations. These findings highlight the complementary strengths of semantic reasoning versus geometric regression, supporting the use of LVLMs as high-level safety validators in SOTIF-oriented automated driving systems.

Method: Proposes native visual tokenization that enforces causal dependencies during tokenization. NativeTok framework includes: (1) Meta Image Transformer (MIT) for latent image modeling, and (2) Mixture of Causal Expert Transformer (MoCET) where lightweight expert blocks generate single tokens conditioned on prior tokens and latent features. Uses Hierarchical Native Training strategy that updates only new expert blocks for efficiency.

Result: Extensive experiments demonstrate the effectiveness of NativeTok in achieving efficient reconstruction while embedding relational constraints within token sequences, addressing the mismatch between tokenization and generation stages.

Conclusion: Native visual tokenization with causal dependencies during tokenization improves image generation coherence by ensuring token sequences have inherent relational structure, making the generation stage more effective.

Abstract: VQ-based image generation typically follows a two-stage pipeline: a tokenizer encodes images into discrete tokens, and a generative model learns their dependencies for reconstruction. However, improved tokenization in the first stage does not necessarily enhance the second-stage generation, as existing methods fail to constrain token dependencies. This mismatch forces the generative model to learn from unordered distributions, leading to bias and weak coherence. To address this, we propose native visual tokenization, which enforces causal dependencies during tokenization. Building on this idea, we introduce NativeTok, a framework that achieves efficient reconstruction while embedding relational constraints within token sequences. NativeTok consists of: (1) a Meta Image Transformer (MIT) for latent image modeling, and (2) a Mixture of Causal Expert Transformer (MoCET), where each lightweight expert block generates a single token conditioned on prior tokens and latent features. We further design a Hierarchical Native Training strategy that updates only new expert blocks, ensuring training efficiency. Extensive experiments demonstrate the effectiveness of NativeTok.

Method: Neural Clothing Tryer (NCT) framework using diffusion models with semantic enhancement and controlling modules. Includes semantic-enhanced module with visual-language encoder for aligned cross-modal features, and semantic controlling module for maintaining garment details while editing model attributes.

Result: Extensive experiments on open benchmarks demonstrate superior performance of NCT framework in preserving garment semantics and details while enabling flexible model customization.

Conclusion: NCT effectively addresses the Cu-VTON task by leveraging diffusion models with semantic enhancement, enabling personalized virtual try-on with preserved garment details and flexible model editing.

Abstract: This work aims to address a novel Customized Virtual Try-ON (Cu-VTON) task, enabling the superimposition of a specified garment onto a model that can be customized in terms of appearance, posture, and additional attributes. Compared with traditional VTON task, it enables users to tailor digital avatars to their individual preferences, thereby enhancing the virtual fitting experience with greater flexibility and engagement. To address this task, we introduce a Neural Clothing Tryer (NCT) framework, which exploits the advanced diffusion models equipped with semantic enhancement and controlling modules to better preserve semantic characterization and textural details of the garment and meanwhile facilitating the flexible editing of the model’s postures and appearances. Specifically, NCT introduces a semantic-enhanced module to take semantic descriptions of garments and utilizes a visual-language encoder to learn aligned features across modalities. The aligned features are served as condition input to the diffusion model to enhance the preservation of the garment’s semantics. Then, a semantic controlling module is designed to take the garment image, tailored posture image, and semantic description as input to maintain garment details while simultaneously editing model postures, expressions, and various attributes. Extensive experiments on the open available benchmark demonstrate the superior performance of the proposed NCT framework.

Method: Use post-hoc slimming (uniform width reduction of pretrained encoders) to measure representational redundancy across six state-of-the-art RS FMs on four downstream classification tasks, comparing with CV models like MAE trained on ImageNet.

Result: RS FMs maintain over 71% relative accuracy at 1% FLOPs after slimming, while CV MAE retains less than 10% accuracy - a sevenfold difference supporting the hypothesis of early overparameterization in RS.

Conclusion: RS foundation models distribute task-relevant information with high redundancy, challenging prevailing scaling paradigms and establishing post-hoc slimmability as both practical deployment strategy and diagnostic tool.

Abstract: Large-scale foundation models (FMs) in remote sensing (RS) are developed based on the paradigms established in computer vision (CV) and have shown promise for various Earth observation applications. However, the direct transfer of scaling assumptions from CV to RS has not been adequately examined. We hypothesize that RS FMs enter an overparameterized regime at substantially smaller scales than their CV counterparts, where increasing parameter count primarily induces redundant representations rather than qualitatively new abstractions. To test this hypothesis, we use post-hoc slimming, where we uniformly reduce the width of pretrained encoder, as a tool to measure representational redundancy across six state-of-the-art RS FMs on four downstream classification tasks. Our findings reveal a significant contrast with those in the CV domain: while a post-hoc slimmed masked autoencoder (MAE) trained on ImageNet retains less than 10% accuracy at 1% FLOPs, RS FMs maintain over 71% relative accuracy at the same budget. This sevenfold difference provides strong empirical support for our hypothesis. We further demonstrate that learned slimmable training can improve both Momentum Contrast (MoCo)- and MAE- based models. In addition, through the explained variance ratio and the feature correlation analysis, we provide mechanistic explanations showing that RS FMs distribute task-relevant information with high redundancy. Our findings establish post-hoc slimmability as both a practical deployment strategy for resource-constrained environments and a diagnostic tool that challenges the prevailing scaling paradigm in RS. Upon acceptance, we will publish all code.

Method: Proposes DyMo with: 1) A novel selection algorithm that maximizes multimodal task-relevant information using task loss as a tractable proxy; 2) A principled reward function for modality selection; 3) A flexible network architecture for arbitrary modality combinations; 4) A tailored training strategy for robust representation learning.

Result: Extensive experiments on diverse natural and medical image datasets show DyMo significantly outperforms state-of-the-art incomplete/dynamic MDL methods across various missing-data scenarios.

Conclusion: DyMo provides a principled solution to the discard-imputation dilemma in incomplete multimodal learning by dynamically selecting reliable recovered modalities at inference time, fully exploring task-relevant information beyond conventional approaches.

Abstract: Multimodal deep learning (MDL) has achieved remarkable success across various domains, yet its practical deployment is often hindered by incomplete multimodal data. Existing incomplete MDL methods either discard missing modalities, risking the loss of valuable task-relevant information, or recover them, potentially introducing irrelevant noise, leading to the discarding-imputation dilemma. To address this dilemma, in this paper, we propose DyMo, a new inference-time dynamic modality selection framework that adaptively identifies and integrates reliable recovered modalities, fully exploring task-relevant information beyond the conventional discard-or-impute paradigm. Central to DyMo is a novel selection algorithm that maximizes multimodal task-relevant information for each test sample. Since direct estimation of such information at test time is intractable due to the unknown data distribution, we theoretically establish a connection between information and the task loss, which we compute at inference time as a tractable proxy. Building on this, a novel principled reward function is proposed to guide modality selection. In addition, we design a flexible multimodal network architecture compatible with arbitrary modality combinations, alongside a tailored training strategy for robust representation learning. Extensive experiments on diverse natural and medical image datasets show that DyMo significantly outperforms state-of-the-art incomplete/dynamic MDL methods across various missing-data scenarios. Our code is available at https://github.com//siyi-wind/DyMo.

Method: Uses Neural Radiance Fields (NeRF) with RGB images, includes specific image capture considerations for proper illumination, employs low light loss for poorly lit understory, and proposes two approaches to remove occluding canopy elements by controlling per-ray integration.

Result: Enables person detection for search and rescue comparable to thermal AOS, and shows potential for forest inventory tasks like tree counting. Provides cost-effective, high-resolution alternative to specialized sensors.

Conclusion: The approach offers a practical, cost-effective solution for various forest applications using only conventional RGB cameras, positioning it as a viable alternative to expensive specialized sensors.

Abstract: Mapping the terrain and understory hidden beneath dense forest canopies is of great interest for numerous applications such as search and rescue, trail mapping, forest inventory tasks, and more. Existing solutions rely on specialized sensors: either heavy, costly airborne LiDAR, or Airborne Optical Sectioning (AOS), which uses thermal synthetic aperture photography and is tailored for person detection. We introduce a novel approach for the reconstruction of canopy-free, photorealistic ground views using only conventional RGB images. Our solution is based on the celebrated Neural Radiance Fields (NeRF), a recent 3D reconstruction method. Additionally, we include specific image capture considerations, which dictate the needed illumination to successfully expose the scene beneath the canopy. To better cope with the poorly lit understory, we employ a low light loss. Finally, we propose two complementary approaches to remove occluding canopy elements by controlling per-ray integration procedure. To validate the value of our approach, we present two possible downstream tasks. For the task of search and rescue (SAR), we demonstrate that our method enables person detection which achieves promising results compared to thermal AOS (using only RGB images). Additionally, we show the potential of our approach for forest inventory tasks like tree counting. These results position our approach as a cost-effective, high-resolution alternative to specialized sensors for SAR, trail mapping, and forest-inventory tasks.

Method: Proposes a conditional compatibility learning framework that leverages vision-language representations to model subject-context relationships under limited supervision. Introduces CAAD-3K benchmark that isolates contextual anomalies by controlling subject identity while varying context.

Result: The method substantially outperforms existing approaches on CAAD-3K and achieves state-of-the-art performance on MVTec-AD and VisA datasets, demonstrating that modeling context dependence complements traditional structural anomaly detection.

Conclusion: Contextual anomaly detection is important for real-world applications where abnormality depends on context. The proposed vision-language approach effectively models subject-context compatibility and improves anomaly detection performance across multiple benchmarks.

Abstract: Anomaly detection is often formulated under the assumption that abnormality is an intrinsic property of an observation, independent of context. This assumption breaks down in many real-world settings, where the same object or action may be normal or anomalous depending on latent contextual factors (e.g., running on a track versus on a highway). We revisit \emph{contextual anomaly detection}, classically defined as context-dependent abnormality, and operationalize it in the visual domain, where anomaly labels depend on subject–context compatibility rather than intrinsic appearance. To enable systematic study of this setting, we introduce CAAD-3K, a benchmark that isolates contextual anomalies by controlling subject identity while varying context. We further propose a conditional compatibility learning framework that leverages vision–language representations to model subject–context relationships under limited supervision. Our method substantially outperforms existing approaches on CAAD-3K and achieves state-of-the-art performance on MVTec-AD and VisA, demonstrating that modeling context dependence complements traditional structural anomaly detection. Our code and dataset will be publicly released.

Method: Proposes DINO-SAE with: 1) Hierarchical Convolutional Patch Embedding module to enhance local structure/texture preservation, 2) Cosine Similarity Alignment objective that enforces semantic consistency while allowing flexible feature magnitudes, and 3) Riemannian Flow Matching to train Diffusion Transformers directly on the hyperspherical latent manifold of SSL-based foundation models.

Result: Achieves state-of-the-art reconstruction quality on ImageNet-1K with 0.37 rFID and 26.2 dB PSNR, while maintaining strong semantic alignment to pretrained VFM. The Riemannian Flow Matching-based DiT achieves efficient convergence with gFID of 3.47 at 80 epochs.

Conclusion: DINO-SAE successfully bridges semantic representation and pixel-level reconstruction by leveraging the hyperspherical nature of SSL-based foundation model representations, enabling high-fidelity image generation while preserving semantic information from pretrained vision models.

Abstract: Recent studies have explored using pretrained Vision Foundation Models (VFMs) such as DINO for generative autoencoders, showing strong generative performance. Unfortunately, existing approaches often suffer from limited reconstruction fidelity due to the loss of high-frequency details. In this work, we present the DINO Spherical Autoencoder (DINO-SAE), a framework that bridges semantic representation and pixel-level reconstruction. Our key insight is that semantic information in contrastive representations is primarily encoded in the direction of feature vectors, while forcing strict magnitude matching can hinder the encoder from preserving fine-grained details. To address this, we introduce Hierarchical Convolutional Patch Embedding module that enhances local structure and texture preservation, and Cosine Similarity Alignment objective that enforces semantic consistency while allowing flexible feature magnitudes for detail retention. Furthermore, leveraging the observation that SSL-based foundation model representations intrinsically lie on a hypersphere, we employ Riemannian Flow Matching to train a Diffusion Transformer (DiT) directly on this spherical latent manifold. Experiments on ImageNet-1K demonstrate that our approach achieves state-of-the-art reconstruction quality, reaching 0.37 rFID and 26.2 dB PSNR, while maintaining strong semantic alignment to the pretrained VFM. Notably, our Riemannian Flow Matching-based DiT exhibits efficient convergence, achieving a gFID of 3.47 at 80 epochs.

Method: Proposes a robust anomaly detection framework that integrates limited anomaly supervision into adaptive deviation learning. Uses a composite anomaly score with three components: deviation score (statistical irregularity), entropy-based uncertainty score (predictive inconsistency), and segmentation-based score (spatial abnormality). Incorporates a small set of labeled anomalies during training while mitigating contamination influence through adaptive instance weighting.

Result: Extensive experiments on MVTec and VisA benchmarks show the framework outperforms state-of-the-art baselines, achieving strong detection and localization performance, interpretability, and robustness under various levels of data contamination.

Conclusion: The proposed framework effectively addresses practical limitations of existing anomaly detection methods by incorporating limited anomaly supervision and handling data contamination, resulting in reliable performance for real-world industrial applications.

Abstract: Visual anomaly detection in real-world industrial settings faces two major limitations. First, most existing methods are trained on purely normal data or on unlabeled datasets assumed to be predominantly normal, presuming the absence of contamination, an assumption that is rarely satisfied in practice. Second, they assume no access to labeled anomaly samples, limiting the model from learning discriminative characteristics of true anomalies. Therefore, these approaches often struggle to distinguish anomalies from normal instances, resulting in reduced detection and weak localization performance. In real-world applications, where training data are frequently contaminated with anomalies, such methods fail to deliver reliable performance. In this work, we propose a robust anomaly detection framework that integrates limited anomaly supervision into the adaptive deviation learning paradigm. We introduce a composite anomaly score that combines three complementary components: a deviation score capturing statistical irregularity, an entropy-based uncertainty score reflecting predictive inconsistency, and a segmentation-based score highlighting spatial abnormality. This unified scoring mechanism enables accurate detection and supports gradient-based localization, providing intuitive and explainable visual evidence of anomalous regions. Following the few-anomaly paradigm, we incorporate a small set of labeled anomalies during training while simultaneously mitigating the influence of contaminated samples through adaptive instance weighting. Extensive experiments on the MVTec and VisA benchmarks demonstrate that our framework outperforms state-of-the-art baselines and achieves strong detection and localization performance, interpretability, and robustness under various levels of data contamination.

Method: Used two MDE models (Dense Prediction Transformers and Depth Anything) on 220 camera trap videos of wild chimpanzees, combined with multiple distance sampling strategies to generate detection distance estimates for population density and abundance inference.

Result: Calibrated DPT consistently outperformed Depth Anything in distance estimation accuracy and downstream density inference. Both models showed systematic biases, overestimating distances and consequently underestimating density/abundance compared to manual methods. Overall approach yielded population estimates within 22% of traditional methods.

Conclusion: MDE-driven camera trap distance sampling is a viable practical alternative to manual distance estimation for ecological monitoring, though animal detection failures across distance ranges remain a primary accuracy limitation.

Abstract: The estimation of abundance and density in unmarked populations of great apes relies on statistical frameworks that require animal-to-camera distance measurements. In practice, acquiring these distances depends on labour-intensive manual interpretation of animal observations across large camera trap video corpora. This study introduces and evaluates an only sparsely explored alternative: the integration of computer vision-based monocular depth estimation (MDE) pipelines directly into ecological camera trap workflows for great ape conservation. Using a real-world dataset of 220 camera trap videos documenting a wild chimpanzee population, we combine two MDE models, Dense Prediction Transformers and Depth Anything, with multiple distance sampling strategies. These components are used to generate detection distance estimates, from which population density and abundance are inferred. Comparative analysis against manually derived ground-truth distances shows that calibrated DPT consistently outperforms Depth Anything. This advantage is observed in both distance estimation accuracy and downstream density and abundance inference. Nevertheless, both models exhibit systematic biases. We show that, given complex forest environments, they tend to overestimate detection distances and consequently underestimate density and abundance relative to conventional manual approaches. We further find that failures in animal detection across distance ranges are a primary factor limiting estimation accuracy. Overall, this work provides a case study that shows MDE-driven camera trap distance sampling is a viable and practical alternative to manual distance estimation. The proposed approach yields population estimates within 22% of those obtained using traditional methods.

Method: Proposes Q-Hawkeye with two main components: 1) Uncertainty-Aware Dynamic Optimization that estimates predictive uncertainty using variance of predicted scores across multiple rollouts and uses this uncertainty to reweight each sample’s update strength, and 2) Perception-Aware Optimization that constructs paired inputs of degraded images and their originals with an Implicit Perception Loss to ground quality judgments in visual evidence.

Result: Extensive experiments demonstrate that Q-Hawkeye outperforms state-of-the-art methods and generalizes better across multiple datasets.

Conclusion: Q-Hawkeye provides a more reliable RL-based visual policy optimization framework for IQA by addressing both uncertainty and perception limitations, with code and models to be made available.

Abstract: Image Quality Assessment (IQA) predicts perceptual quality scores consistent with human judgments. Recent RL-based IQA methods built on MLLMs focus on generating visual quality descriptions and scores, ignoring two key reliability limitations: (i) although the model’s prediction stability varies significantly across training samples, existing GRPO-based methods apply uniform advantage weighting, thereby amplifying noisy signals from unstable samples in gradient updates; (ii) most works emphasize text-grounded reasoning over images while overlooking the model’s visual perception ability of image content. In this paper, we propose Q-Hawkeye, an RL-based reliable visual policy optimization framework that redesigns the learning signal through unified Uncertainty-Aware Dynamic Optimization and Perception-Aware Optimization. Q-Hawkeye estimates predictive uncertainty using the variance of predicted scores across multiple rollouts and leverages this uncertainty to reweight each sample’s update strength, stabilizing policy optimization. To strengthen perceptual reliability, we construct paired inputs of degraded images and their original images and introduce an Implicit Perception Loss that constrains the model to ground its quality judgments in genuine visual evidence. Extensive experiments demonstrate that Q-Hawkeye outperforms state-of-the-art methods and generalizes better across multiple datasets. The code and models will be made available.

Method: Propose SLImE (Semantic Leakage from Image Embeddings) - a lightweight inference framework using locally trained semantic retriever with off-the-shelf models, without task-specific decoder training. Validates semantic leakage through aligned embeddings to retrieved tags, symbolic representations, and coherent descriptions.

Result: Demonstrates consistent recovery of semantic information across diverse inference tasks and embedding models (GEMINI, COHERE, NOMIC, CLIP), revealing fundamental vulnerability in image embeddings where preserved semantic neighborhoods enable semantic leakage.

Conclusion: Image embeddings have inherent privacy vulnerabilities due to preserved semantic neighborhood structures under alignment, challenging current assumptions about embedding privacy and highlighting challenges for privacy preservation.

Abstract: Image embeddings are generally assumed to pose limited privacy risk. We challenge this assumption by formalizing semantic leakage as the ability to recover semantic structures from compressed image embeddings. Surprisingly, we show that semantic leakage does not require exact reconstruction of the original image. Preserving local semantic neighborhoods under embedding alignment is sufficient to expose the intrinsic vulnerability of image embeddings. Crucially, this preserved neighborhood structure allows semantic information to propagate through a sequence of lossy mappings. Based on this conjecture, we propose Semantic Leakage from Image Embeddings (SLImE), a lightweight inference framework that reveals semantic information from standalone compressed image embeddings, incorporating a locally trained semantic retriever with off-the-shelf models, without training task-specific decoders. We thoroughly validate each step of the framework empirically, from aligned embeddings to retrieved tags, symbolic representations, and grammatical and coherent descriptions. We evaluate SLImE across a range of open and closed embedding models, including GEMINI, COHERE, NOMIC, and CLIP, and demonstrate consistent recovery of semantic information across diverse inference tasks. Our results reveal a fundamental vulnerability in image embeddings, whereby the preservation of semantic neighborhoods under alignment enables semantic leakage, highlighting challenges for privacy preservation.1

Method: Two-stage hierarchical visual budgeting: 1) Frame-Level Budgeting identifies keyframes based on visual dynamics and relevance, 2) Token-Level Budgeting allocates tokens in two phases - Core Tokens for high-relevance content and Context Tokens selected via batched Maximal Marginal Relevance algorithm.

Result: Extensive experiments show Triage improves inference speed, reduces memory footprint, and maintains or surpasses baseline performance on various video reasoning benchmarks.

Conclusion: Triage effectively addresses computational challenges in video VLMs through efficient resource allocation, offering a practical plug-and-play solution for video reasoning tasks.

Abstract: Vision-Language Models (VLMs) face significant computational challenges in video processing due to massive data redundancy, which creates prohibitively long token sequences. To address this, we introduce Triage, a training-free, plug-and-play framework that reframes video reasoning as a resource allocation problem via hierarchical visual budgeting. Its first stage, Frame-Level Budgeting, identifies keyframes by evaluating their visual dynamics and relevance, generating a strategic prior based on their importance scores. Guided by this prior, the second stage, Token-Level Budgeting, allocates tokens in two phases: it first secures high-relevance Core Tokens, followed by diverse Context Tokens selected with an efficient batched Maximal Marginal Relevance (MMR) algorithm. Extensive experiments demonstrate that Triage improves inference speed and reduces memory footprint, while maintaining or surpassing the performance of baselines and other methods on various video reasoning benchmarks.

Method: Investigates Stable Diffusion and CycleGAN as generative AI methods for dataset expansion. Focuses on segmenting combine harvester components in thermal images for subsequent defect detection. Uses these generative models to create synthetic defective samples to balance the dataset.

Result: Dataset expansion using Stable Diffusion yields the most significant improvement, enhancing segmentation performance by 4.6%, achieving a Mean Intersection over Union (Mean IoU) of 84.6%.

Conclusion: Generative AI, particularly Stable Diffusion, shows strong potential for addressing dataset imbalance in industrial optical quality control, significantly improving segmentation performance for defect detection tasks.

Abstract: Supervised machine learning algorithms play a crucial role in optical quality control within industrial production. These approaches require representative datasets for effective model training. However, while non-defective components are frequent, defective parts are rare in production, resulting in highly imbalanced datasets that adversely impact model performance. Existing strategies to address this challenge, such as specialized loss functions or traditional data augmentation techniques, have limitations, including the need for careful hyperparameter tuning or the alteration of only simple image features. Therefore, this work explores the potential of generative artificial intelligence (GenAI) as an alternative method for expanding limited datasets and enhancing supervised machine learning performance. Specifically, we investigate Stable Diffusion and CycleGAN as image generation models, focusing on the segmentation of combine harvester components in thermal images for subsequent defect detection. Our results demonstrate that dataset expansion using Stable Diffusion yields the most significant improvement, enhancing segmentation performance by 4.6 %, resulting in a Mean Intersection over Union (Mean IoU) of 84.6 %.

Method: Proposes automated annotation using ArUco markers’ built-in recognition modules that provide ID and positional information. Uses this automatic annotation to train a YOLO-based deep learning model for marker recognition.

Result: The YOLO-based model trained with automatically annotated data improves recognition performance compared to conventional image processing, especially for images affected by blur or defocus. Automatic annotation reduces human effort and ensures consistent labeling quality.

Conclusion: Automated annotation using ArUco markers enables efficient training of robust deep learning models for marker recognition in factory automation, overcoming manual annotation bottlenecks while improving performance under challenging conditions.

Abstract: Factory automation has become increasingly important due to labor shortages, leading to the introduction of autonomous mobile robots for tasks such as material transportation. Markers are commonly used for robot self-localization and object identification. In the RoboCup Logistics League (RCLL), ArUco markers are employed both for robot localization and for identifying processing modules. Conventional recognition relies on OpenCV-based image processing, which detects black-and-white marker patterns. However, these methods often fail under noise, motion blur, defocus, or varying illumination conditions. Deep-learning-based recognition offers improved robustness under such conditions, but requires large amounts of annotated data. Annotation must typically be done manually, as the type and position of objects cannot be detected automatically, making dataset preparation a major bottleneck. In contrast, ArUco markers include built-in recognition modules that provide both ID and positional information, enabling automatic annotation. This paper proposes an automated annotation method for training deep-learning models on ArUco marker images. By leveraging marker detection results obtained from the ArUco module, the proposed approach eliminates the need for manual labeling. A YOLO-based model is trained using the automatically annotated dataset, and its performance is evaluated under various conditions. Experimental results demonstrate that the proposed method improves recognition performance compared with conventional image-processing techniques, particularly for images affected by blur or defocus. Automatic annotation also reduces human effort and ensures consistent labeling quality. Future work will investigate the relationship between confidence thresholds and recognition performance.

Method: GaussianSVR uses 3D Gaussian representations for high-fidelity reconstruction, employs simulated forward slice acquisition for self-supervised training, and implements multi-resolution training to optimize Gaussian parameters and spatial transformations.

Result: GaussianSVR outperforms baseline methods on fetal MR volumetric reconstruction tasks.

Conclusion: The proposed self-supervised framework enables efficient and accurate 3D fetal MR reconstruction without requiring ground truth volumes.

Abstract: Reconstructing 3D fetal MR volumes from motion-corrupted stacks of 2D slices is a crucial and challenging task. Conventional slice-to-volume reconstruction (SVR) methods are time-consuming and require multiple orthogonal stacks for reconstruction. While learning-based SVR approaches have significantly reduced the time required at the inference stage, they heavily rely on ground truth information for training, which is inaccessible in practice. To address these challenges, we propose GaussianSVR, a self-supervised framework for slice-to-volume reconstruction. GaussianSVR represents the target volume using 3D Gaussian representations to achieve high-fidelity reconstruction. It leverages a simulated forward slice acquisition model to enable self-supervised training, alleviating the need for ground-truth volumes. Furthermore, to enhance both accuracy and efficiency, we introduce a multi-resolution training strategy that jointly optimizes Gaussian parameters and spatial transformations across different resolution levels. Experiments show that GaussianSVR outperforms the baseline methods on fetal MR volumetric reconstruction. Code will be available upon acceptance.

Method: Train separate models on annotations from single experts, then aggregate their predictions to emulate consensus, capturing inter-rater variability more effectively than mixed annotation training.

Result: Experiments on colorectal organoid dataset with two expert annotators show improved calibration performance while maintaining comparable detection accuracy.

Conclusion: Explicitly modeling rater disagreement through rater-specific ensembles leads to more trustworthy object detectors in biomedical imaging.

Abstract: Deep learning-based object detectors have achieved impressive performance in microscopy imaging, yet their confidence estimates often lack calibration, limiting their reliability for biomedical applications. In this work, we introduce a new approach to improve model calibration by leveraging multi-rater annotations. We propose to train separate models on the annotations from single experts and aggregate their predictions to emulate consensus. This improves upon label sampling strategies, where models are trained on mixed annotations, and offers a more principled way to capture inter-rater variability. Experiments on a colorectal organoid dataset annotated by two experts demonstrate that our rater-specific ensemble strategy improves calibration performance while maintaining comparable detection accuracy. These findings suggest that explicitly modelling rater disagreement can lead to more trustworthy object detectors in biomedical imaging.

Method: OSGA uses variance-based data selection to pick one informative sample, then learns a single steering vector with contrastive objective and generative anchor regularization. The resulting vector is universally applied at a certain layer during inference without modifying model parameters.

Result: Experiments across multiple benchmarks show that a single OSGA-optimized steering vector consistently improves hallucination mitigation and safety enhancement with negligible overhead.

Conclusion: One-shot steering with OSGA provides a practical and scalable solution for reliable VLMs, demonstrating that effective steering vectors can generalize across inputs when tasks share semantic alignment.

Abstract: Vision Language Models (VLMs) achieve strong performance on multimodal tasks but still suffer from hallucination and safety-related failures that persist even at scale. Steering offers a lightweight technique to improve model performance. However, steering, whether input-dependent or input-independent, achieves a meaningful trade-off between efficiency and effectiveness. In this work, we observe that steering vectors can generalize across inputs when tasks share aligned semantic intent. Based on this insight, we propose \textbf{OSGA} (\textbf{O}ne-shot \textbf{S}teering with \textbf{G}enerative \textbf{A}nchor), an input-independent framework that improves model performance with a single optimization instance. OSGA first selects an informative sample via a variance-based data selection strategy and learns a single steering vector with a contrastive objective with generative anchor regularization. The resulting vector can be universally applied at a certain layer during inference time without modifying model parameters. Experiments across multiple benchmarks show that a single OSGA-optimized steering vector consistently improves hallucination mitigation and safety enhancement with negligible overhead, highlighting one-shot steering as a practical and scalable solution for reliable VLMs.

Method: Entity-centric formulation replaces image-to-image retrieval with compact hierarchy of geographic entities (country, region, subregion, city) embedded in Hyperbolic space. Uses Geo-Weighted Hyperbolic contrastive learning incorporating haversine distance into contrastive objective.

Result: Establishes new SOTA on OSV5M benchmark: reduces mean geodesic error by 19.5%, improves fine-grained subregion accuracy by 43%. Uses only 240k entity embeddings vs over 5M image embeddings in previous methods.

Conclusion: Geometry-aware hierarchical embeddings provide scalable and conceptually new alternative for global image geolocation, enabling interpretable predictions and efficient inference.

Abstract: Visual geolocalization, the task of predicting where an image was taken, remains challenging due to global scale, visual ambiguity, and the inherently hierarchical structure of geography. Existing paradigms rely on either large-scale retrieval, which requires storing a large number of image embeddings, grid-based classifiers that ignore geographic continuity, or generative models that diffuse over space but struggle with fine detail. We introduce an entity-centric formulation of geolocation that replaces image-to-image retrieval with a compact hierarchy of geographic entities embedded in Hyperbolic space. Images are aligned directly to country, region, subregion, and city entities through Geo-Weighted Hyperbolic contrastive learning by directly incorporating haversine distance into the contrastive objective. This hierarchical design enables interpretable predictions and efficient inference with 240k entity embeddings instead of over 5 million image embeddings on the OSV5M benchmark, on which our method establishes a new state-of-the-art performance. Compared to the current methods in the literature, it reduces mean geodesic error by 19.5%, while improving the fine-grained subregion accuracy by 43%. These results demonstrate that geometry-aware hierarchical embeddings provide a scalable and conceptually new alternative for global image geolocation.

Method: CoSA represents point clouds as compact combinations of class-specific prototypes in a shared low-dimensional semantic space. Adversarial perturbations are optimized within a low-rank subspace to induce coherent, architecture-agnostic variations, suppressing model-dependent noise and constraining perturbations to semantically meaningful directions.

Result: Extensive experiments on multiple datasets and network architectures show CoSA consistently outperforms state-of-the-art transferable attacks while maintaining competitive imperceptibility and robustness under common defense strategies.

Conclusion: The compact subspace perspective enables more transferable adversarial attacks on point clouds by operating in a shared semantic space that captures architecture-agnostic variations, improving generalization across different models.

Abstract: Transferable adversarial attacks on point clouds remain challenging, as existing methods often rely on model-specific gradients or heuristics that limit generalization to unseen architectures. In this paper, we rethink adversarial transferability from a compact subspace perspective and propose CoSA, a transferable attack framework that operates within a shared low-dimensional semantic space. Specifically, each point cloud is represented as a compact combination of class-specific prototypes that capture shared semantic structure, while adversarial perturbations are optimized within a low-rank subspace to induce coherent and architecture-agnostic variations. This design suppresses model-dependent noise and constrains perturbations to semantically meaningful directions, thereby improving cross-model transferability without relying on surrogate-specific artifacts. Extensive experiments on multiple datasets and network architectures demonstrate that CoSA consistently outperforms state-of-the-art transferable attacks, while maintaining competitive imperceptibility and robustness under common defense strategies. Codes will be made public upon paper acceptance.

Method: Uses motion cues from scene flow of static objects to detect rotational misalignment. Combines neural scene flow prior for flow estimation with a dual-branch detection network that fuses learned global flow features with handcrafted geometric descriptors. Performs two binary classification tasks: global misalignment detection and axis-specific misalignment detection for each rotational axis.

Result: Experiments on nuScenes dataset demonstrate robust miscalibration detection capability, establishing a benchmark for sensor-to-vehicle miscalibration detection without requiring additional sensors.

Conclusion: FlowCalib provides the first framework for detecting LiDAR-to-vehicle miscalibration using scene flow, addressing a critical gap in autonomous driving safety by identifying root-cause sensor misalignments rather than just correcting downstream errors.

Abstract: Accurate sensor-to-vehicle calibration is essential for safe autonomous driving. Angular misalignments of LiDAR sensors can lead to safety-critical issues during autonomous operation. However, current methods primarily focus on correcting sensor-to-sensor errors without considering the miscalibration of individual sensors that cause these errors in the first place. We introduce FlowCalib, the first framework that detects LiDAR-to-vehicle miscalibration using motion cues from the scene flow of static objects. Our approach leverages the systematic bias induced by rotational misalignment in the flow field generated from sequential 3D point clouds, eliminating the need for additional sensors. The architecture integrates a neural scene flow prior for flow estimation and incorporates a dual-branch detection network that fuses learned global flow features with handcrafted geometric descriptors. These combined representations allow the system to perform two complementary binary classification tasks: a global binary decision indicating whether misalignment is present and separate, axis-specific binary decisions indicating whether each rotational axis is misaligned. Experiments on the nuScenes dataset demonstrate FlowCalib’s ability to robustly detect miscalibration, establishing a benchmark for sensor-to-vehicle miscalibration detection.

Method: SEAL presents a unified framework supporting both event segmentation and open-vocabulary mask classification at instance-level and part-level granularity. The model uses a parameter-efficient architecture and includes a variant for generic spatiotemporal OV-EIS that doesn’t require visual prompts.

Result: Extensive experiments on four curated benchmarks show SEAL largely outperforms proposed baselines in both performance and inference speed. The framework handles coarse to fine class configurations and instance to part-level semantic granularity.

Conclusion: SEAL is the first framework addressing Open-Vocabulary Event Instance Segmentation, enabling comprehensive event sensor understanding with free-form language capabilities across multiple granularity levels.

Abstract: Scene understanding with free-form language has been widely explored within diverse modalities such as images, point clouds, and LiDAR. However, related studies on event sensors are scarce or narrowly centered on semantic-level understanding. We introduce SEAL, the first Semantic-aware Segment Any Events framework that addresses Open-Vocabulary Event Instance Segmentation (OV-EIS). Given the visual prompt, our model presents a unified framework to support both event segmentation and open-vocabulary mask classification at multiple levels of granularity, including instance-level and part-level. To enable thorough evaluation on OV-EIS, we curate four benchmarks that cover label granularity from coarse to fine class configurations and semantic granularity from instance-level to part-level understanding. Extensive experiments show that our SEAL largely outperforms proposed baselines in terms of performance and inference speed with a parameter-efficient architecture. In the Appendix, we further present a simple variant of our SEAL achieving generic spatiotemporal OV-EIS that does not require any visual prompts from users in the inference. Check out our project page in https://0nandon.github.io/SEAL

Method: Three technical innovations: 1) Lightness prior anchored guided relighting diffusion for stable intermediate results, 2) Hybrid Motion-Adaptive Lighting Smoothing Filter using optical flow for temporal stability without motion blur, 3) LAB-based Detail Fusion module to preserve high-frequency details.

Result: Extensive experiments show Hi-Light significantly outperforms state-of-the-art methods in both qualitative and quantitative comparisons, producing stable, highly detailed relit videos.

Conclusion: Hi-Light provides a robust, training-free solution for high-quality video relighting with temporal stability and detail preservation, along with a new evaluation metric for lighting consistency.

Abstract: Video relighting offers immense creative potential and commercial value but is hindered by challenges, including the absence of an adequate evaluation metric, severe light flickering, and the degradation of fine-grained details during editing. To overcome these challenges, we introduce Hi-Light, a novel, training-free framework for high-fidelity, high-resolution, robust video relighting. Our approach introduces three technical innovations: lightness prior anchored guided relighting diffusion that stabilises intermediate relit video, a Hybrid Motion-Adaptive Lighting Smoothing Filter that leverages optical flow to ensure temporal stability without introducing motion blur, and a LAB-based Detail Fusion module that preserves high-frequency detail information from the original video. Furthermore, to address the critical gap in evaluation, we propose the Light Stability Score, the first quantitative metric designed to specifically measure lighting consistency. Extensive experiments demonstrate that Hi-Light significantly outperforms state-of-the-art methods in both qualitative and quantitative comparisons, producing stable, highly detailed relit videos.

Method: Introduces Med-Scout framework using RL with intrinsic geometric logic from unlabeled medical images. Uses three proxy tasks for supervision: Hierarchical Scale Localization, Topological Jigsaw Reconstruction, and Anomaly Consistency Detection. Also presents Med-Scout-Bench benchmark for evaluation.

Result: Med-Scout significantly mitigates geometric blindness, outperforming leading proprietary and open-source MLLMs by over 40% on the benchmark. Enhanced geometric perception generalizes to broader medical understanding with superior results on radiological and comprehensive medical VQA tasks.

Conclusion: The framework successfully addresses geometric blindness in medical MLLMs through unsupervised geometric learning, improving both geometric perception and overall medical understanding without costly expert annotations.

Abstract: Despite recent Multimodal Large Language Models (MLLMs)’ linguistic prowess in medical diagnosis, we find even state-of-the-art MLLMs suffer from a critical perceptual deficit: geometric blindness. This failure to ground outputs in objective geometric constraints leads to plausible yet factually incorrect hallucinations, rooted in training paradigms that prioritize linguistic fluency over geometric fidelity. This paper introduces Med-Scout, a novel framework that “cures” this blindness via Reinforcement Learning (RL) that leverages the intrinsic geometric logic latent within unlabeled medical images. Instead of relying on costly expert annotations, Med-Scout derives verifiable supervision signals through three strategic proxy tasks: Hierarchical Scale Localization, Topological Jigsaw Reconstruction, and Anomaly Consistency Detection. To rigorously quantify this deficit, we present Med-Scout-Bench, a new benchmark specifically designed to evaluate geometric perception. Extensive evaluations show that Med-Scout significantly mitigates geometric blindness, outperforming leading proprietary and open-source MLLMs by over 40% on our benchmark. Furthermore, this enhanced geometric perception generalizes to broader medical understanding, achieving superior results on radiological and comprehensive medical VQA tasks.

Method: Proposes Region-Normalized DPO (RN-DPO), a segmentation-aware objective that normalizes preference updates by the size of the disagreement region between masks. Uses proposals from a supervised base segmenter trained on small labeled set, with preference pairs mined from noisy judges.

Result: RN-DPO improves sustained performance and stabilizes preference-based fine-tuning across two medical datasets and multiple regimes, outperforming standard DPO and strong baselines without requiring additional pixel annotations.

Conclusion: RN-DPO effectively leverages noisy automatic quality signals for segmentation improvement by reducing leverage of harmful comparisons through region-normalized updates, enabling scalable model refinement without expensive annotations.

Abstract: While dense pixel-wise annotations remain the gold standard for medical image segmentation, they are costly to obtain and limit scalability. In contrast, many deployed systems already produce inexpensive automatic quality-control (QC) signals like model agreement, uncertainty measures, or learned mask-quality scores which can be used for further model training without additional ground-truth annotation. However, these signals can be noisy and biased, making preference-based fine-tuning susceptible to harmful updates. We study Direct Preference Optimization (DPO) for segmentation from such noisy judges using proposals generated by a supervised base segmenter trained on a small labeled set. We find that outcomes depend strongly on how preference pairs are mined: selecting the judge’s top-ranked proposal can improve peak performance when the judge is reliable, but can amplify harmful errors under weaker judges. We propose Region-Normalized DPO (RN-DPO), a segmentation-aware objective which normalizes preference updates by the size of the disagreement region between masks, reducing the leverage of harmful comparisons and improving optimization stability. Across two medical datasets and multiple regimes, RN-DPO improves sustained performance and stabilizes preference-based fine-tuning, outperforming standard DPO and strong baselines without requiring additional pixel annotations.

Method: Proposes Video-o3 framework with: 1) Task-Decoupled Attention Masking to mitigate attention dispersion from heterogeneous reasoning and tool-calling, 2) Verifiable Trajectory-Guided Reward to balance exploration coverage with reasoning efficiency, and 3) Seeker-173K dataset of 173K tool-interaction trajectories for training.

Result: Achieves 72.1% accuracy on MLVU and 46.5% on Video-Holmes, substantially outperforming state-of-the-art methods and demonstrating strong multi-hop evidence-seeking and reasoning capabilities.

Conclusion: Video-o3 validates the effectiveness of native tool invocation in long-video scenarios and demonstrates superior performance through iterative discovery of salient visual evidence.

Abstract: Existing multimodal large language models for long-video understanding predominantly rely on uniform sampling and single-turn inference, limiting their ability to identify sparse yet critical evidence amid extensive redundancy. We introduce Video-o3, a novel framework that supports iterative discovery of salient visual clues, fine-grained inspection of key segments, and adaptive termination once sufficient evidence is acquired. Technically, we address two core challenges in interleaved tool invocation. First, to mitigate attention dispersion induced by the heterogeneity of reasoning and tool-calling, we propose Task-Decoupled Attention Masking, which isolates per-step concentration while preserving shared global context. Second, to control context length growth in multi-turn interactions, we introduce a Verifiable Trajectory-Guided Reward that balances exploration coverage with reasoning efficiency. To support training at scale, we further develop a data synthesis pipeline and construct Seeker-173K, comprising 173K high-quality tool-interaction trajectories for effective supervised and reinforcement learning. Extensive experiments show that Video-o3 substantially outperforms state-of-the-art methods, achieving 72.1% accuracy on MLVU and 46.5% on Video-Holmes. These results demonstrate Video-o3’s strong multi-hop evidence-seeking and reasoning capabilities, and validate the effectiveness of native tool invocation in long-video scenarios.

Method: Created ShotFinder benchmark with 1,210 samples across 20 categories using large models with human verification; proposed three-stage retrieval pipeline: query expansion via video imagination, candidate retrieval with search engine, and description-guided temporal localization.

Result: Experiments show significant gap to human performance, with clear imbalance across constraints: temporal localization is relatively tractable, while color and visual style remain major challenges.

Conclusion: Open-domain video shot retrieval is a critical capability that multimodal large models have yet to overcome, revealing limitations in current video understanding systems.

Abstract: In recent years, large language models (LLMs) have made rapid progress in information retrieval, yet existing research has mainly focused on text or static multimodal settings. Open-domain video shot retrieval, which involves richer temporal structure and more complex semantics, still lacks systematic benchmarks and analysis. To fill this gap, we introduce ShotFinder, a benchmark that formalizes editing requirements as keyframe-oriented shot descriptions and introduces five types of controllable single-factor constraints: Temporal order, Color, Visual style, Audio, and Resolution. We curate 1,210 high-quality samples from YouTube across 20 thematic categories, using large models for generation with human verification. Based on the benchmark, we propose ShotFinder, a text-driven three-stage retrieval and localization pipeline: (1) query expansion via video imagination, (2) candidate video retrieval with a search engine, and (3) description-guided temporal localization. Experiments on multiple closed-source and open-source models reveal a significant gap to human performance, with clear imbalance across constraints: temporal localization is relatively tractable, while color and visual style remain major challenges. These results reveal that open-domain video shot retrieval is still a critical capability that multimodal large models have yet to overcome.

Method: Created DoraVQA dataset (5,344 QA pairs) from 8 seasons of Dora the Explorer with timestamp alignment. Fine-tuned Qwen2 and Qwen3 using Group Relative Policy Optimization (GRPO) on 38 hours of educational videos, leveraging the structured context-question-pause-answer format.

Result: Achieved 8-14 point improvements on DoraVQA, state-of-the-art 86.16% on CVBench, with strong transfer to Video-MME and NExT-QA. Shows effective generalization from narrow pedagogical content to broad multimodal understanding.

Conclusion: VLMs can learn robust reasoning from structured educational content, demonstrating that content structure matters as much as content scale for improving multimodal reasoning capabilities.

Abstract: Vision-language models (VLMs) demonstrate impressive performance on standard video understanding benchmarks yet fail systematically on simple reasoning tasks that preschool children can solve, including counting, spatial reasoning, and compositional understanding. We hypothesize that the pedagogically-structured content of educational videos provides an ideal training signal for improving these capabilities. We introduce DoraVQA, a dataset of 5,344 question-answer pairs automatically extracted from 8 seasons of Dora the Explorer with precise timestamp alignment. Each episode follows a consistent \textit{context-question-pause-answer} structure that creates a self-contained learning environment analogous to interactive tutoring. We fine-tune both Qwen2 and Qwen3 using Group Relative Policy Optimization (GRPO), leveraging the clear correctness signals and structured reasoning traces inherent in educational content. Despite training exclusively on 38 hours of children’s educational videos, our approach achieves improvements of 8-14 points on DoraVQA and state-of-the-art 86.16% on CVBench, with strong transfer to Video-MME and NExT-QA, demonstrating effective generalization from narrow pedagogical content to broad multimodal understanding. Through cross-domain benchmarks, we show that VLMs can perform tasks that require robust reasoning learned from structured educational content, suggesting that content structure matters as much as content scale.

Method: Uses Brownian Distance Covariance to capture linear/nonlinear dependencies via pairwise distances for dynamic adaptation without training. Combines with attribute-enhanced prompting, dynamic clustering, and pseudo-label refinement.

Result: Significantly reduces computational cost while achieving state-of-the-art performance in domain and cross-dataset generalization across diverse datasets.

Conclusion: TaTa provides efficient, stable test-time adaptation for VLMs without training or backpropagation, enhancing real-world applicability through statistical dependency measures and enhanced prompting.

Abstract: Vision-language models suffer performance degradation under domain shift, limiting real-world applicability. Existing test-time adaptation methods are computationally intensive, rely on back-propagation, and often focus on single modalities. To address these issues, we propose Training-free Test-Time Adaptation with Brownian Distance Covariance (TaTa). TaTa leverages Brownian Distance Covariance-a powerful statistical measure that captures both linear and nonlinear dependencies via pairwise distances-to dynamically adapt VLMs to new domains without training or back-propagation. This not only improves efficiency but also enhances stability by avoiding disruptive weight updates. TaTa further integrates attribute-enhanced prompting to improve vision-language inference with descriptive visual cues. Combined with dynamic clustering and pseudo-label refinement, it effectively recalibrates the model for novel visual contexts. Experiments across diverse datasets show that TaTa significantly reduces computational cost while achieving state-of-the-art performance in domain and cross-dataset generalization.

Method: Evaluated two OSOD models (GroundingDINO and YOLO-E) on real-world XR images, simulating diverse user prompting behaviors using vision-language models. Considered four prompt types: standard, underdetailed, overdetailed, and pragmatically ambiguous. Examined impact of two enhancement strategies on these prompts.

Result: Both models show stable performance under underdetailed and standard prompts, but suffer degradation under ambiguous prompts. Overdetailed prompts primarily affect GroundingDINO. Prompt enhancement substantially improves robustness under ambiguity, yielding gains exceeding 55% mIoU and 41% average confidence.

Conclusion: Proposes several prompting strategies and prompt enhancement methods for OSOD models in XR environments based on findings about prompt-conditioned robustness.

Abstract: Open-set object detection (OSOD) localizes objects while identifying and rejecting unknown classes at inference. While recent OSOD models perform well on benchmarks, their behavior under realistic user prompting remains underexplored. In interactive XR settings, user-generated prompts are often ambiguous, underspecified, or overly detailed. To study prompt-conditioned robustness, we evaluate two OSOD models, GroundingDINO and YOLO-E, on real-world XR images and simulate diverse user prompting behaviors using vision-language models. We consider four prompt types: standard, underdetailed, overdetailed, and pragmatically ambiguous, and examine the impact of two enhancement strategies on these prompts. Results show that both models exhibit stable performance under underdetailed and standard prompts, while they suffer degradation under ambiguous prompts. Overdetailed prompts primarily affect GroundingDINO. Prompt enhancement substantially improves robustness under ambiguity, yielding gains exceeding 55% mIoU and 41% average confidence. Based on the findings, we propose several prompting strategies and prompt enhancement methods for OSOD models in XR environments.

Method: VideoGPA uses a geometry foundation model to automatically generate dense preference signals, then applies Direct Preference Optimization (DPO) to guide video diffusion models toward 3D consistency without human annotations.

Result: VideoGPA significantly improves temporal stability, physical plausibility, and motion coherence using minimal preference pairs, consistently outperforming state-of-the-art baselines in extensive experiments.

Conclusion: The self-supervised framework effectively addresses 3D consistency issues in video generation by leveraging geometric preference alignment, offering a data-efficient solution for improving video diffusion models.

Abstract: While recent video diffusion models (VDMs) produce visually impressive results, they fundamentally struggle to maintain 3D structural consistency, often resulting in object deformation or spatial drift. We hypothesize that these failures arise because standard denoising objectives lack explicit incentives for geometric coherence. To address this, we introduce VideoGPA (Video Geometric Preference Alignment), a data-efficient self-supervised framework that leverages a geometry foundation model to automatically derive dense preference signals that guide VDMs via Direct Preference Optimization (DPO). This approach effectively steers the generative distribution toward inherent 3D consistency without requiring human annotations. VideoGPA significantly enhances temporal stability, physical plausibility, and motion coherence using minimal preference pairs, consistently outperforming state-of-the-art baselines in extensive experiments.

Method: Two key designs: 1) Encoding face identity into a series of feature maps instead of a single image token to retain more facial details, and 2) A disentangled integration strategy to balance text and image guidance during generation, reducing conflicts between reference faces and text prompts.

Result: Extensive experiments demonstrate effectiveness on various applications including human image personalization, face swapping under language prompts, and making virtual characters into real people, with superior identity preservation and instruction following compared to prior methods.

Conclusion: FlashFace provides a practical solution for photo personalization with better identity fidelity and text instruction compliance, enabling diverse applications in human image customization.

Abstract: This work presents FlashFace, a practical tool with which users can easily personalize their own photos on the fly by providing one or a few reference face images and a text prompt. Our approach is distinguishable from existing human photo customization methods by higher-fidelity identity preservation and better instruction following, benefiting from two subtle designs. First, we encode the face identity into a series of feature maps instead of one image token as in prior arts, allowing the model to retain more details of the reference faces (e.g., scars, tattoos, and face shape ). Second, we introduce a disentangled integration strategy to balance the text and image guidance during the text-to-image generation process, alleviating the conflict between the reference faces and the text prompts (e.g., personalizing an adult into a “child” or an “elder”). Extensive experimental results demonstrate the effectiveness of our method on various applications, including human image personalization, face swapping under language prompts, making virtual characters into real people, etc. Project Page: https://jshilong.github.io/flashface-page.

Method: Three components: (1) synthetic data generation with 3D scanning, articulation rigging, and physically-based rendering; (2) pose estimation framework combining tool detection with pose and articulation estimation; (3) training strategy using synthetic and real unannotated video data with domain adaptation and automatically generated pseudo-labels.

Result: Demonstrates good performance and real-world applicability on real open surgery data, showing potential for integration into medical augmented reality and robotic systems.

Conclusion: The framework successfully addresses surgical instrument pose estimation challenges and eliminates need for extensive manual annotation, enabling practical applications in medical AR and robotics.

Abstract: This work presents a framework for monocular 6D pose estimation of surgical instruments in open surgery, addressing challenges such as object articulations, specularity, occlusions, and synthetic-to-real domain adaptation. The proposed approach consists of three main components: $(1)$ synthetic data generation pipeline that incorporates 3D scanning of surgical tools with articulation rigging and physically-based rendering; $(2)$ a tailored pose estimation framework combining tool detection with pose and articulation estimation; and $(3)$ a training strategy on synthetic and real unannotated video data, employing domain adaptation with automatically generated pseudo-labels. Evaluations conducted on real data of open surgery demonstrate the good performance and real-world applicability of the proposed framework, highlighting its potential for integration into medical augmented reality and robotic systems. The approach eliminates the need for extensive manual annotation of real surgical data.

Method: Developed STAGE: 1) First GenAI image creator with accurate 3D pose control to generate controlled image pairs, 2) Novel evaluation strategy to isolate single factors, 3) Generated benchmarks to audit popular pose estimators’ sensitivity to various factors.

Result: Natural variations (gender, ethnicity, age, clothing, location, weather) can severely degrade pose estimator performance. The study raises doubts about current pose estimators’ readiness for open-world deployment due to robustness issues.

Conclusion: STAGE provides a systematic way to audit 3D human pose estimators and quantify their robustness issues. The toolkit highlights significant performance degradation from natural variations and aims to establish a benchmark for measuring these problems.

Abstract: For safety-critical applications, it is crucial to audit 3D human pose estimators before deployment. Will the system break down if the weather or the clothing changes? Is it robust regarding gender and age? To answer these questions and more, we need controlled studies with images that differ in a single attribute, but real benchmarks cannot provide such pairs. We thus present STAGE, a GenAI data toolkit for auditing 3D human pose estimators. For STAGE, we develop the first GenAI image creator with accurate 3D pose control and propose a novel evaluation strategy to isolate and quantify the effects of single factors such as gender, ethnicity, age, clothing, location, and weather. Enabled by STAGE, we generate a series of benchmarks to audit, for the first time, the sensitivity of popular pose estimators towards such factors. Our results show that natural variations can severely degrade pose estimator performance, raising doubts about their readiness for open-world deployment. We aim to highlight these robustness issues and establish STAGE as a benchmark to quantify them.

Method: Proposes ARB-LLM with alternating refined binarization (ARB) algorithm to progressively update binarization parameters, reducing quantization error. Extends to ARB-X and ARB-RC variants considering calibration data and column deviation. Refines weight partition with column-group bitmap (CGB) strategy.

Result: ARB-LLM significantly outperforms state-of-the-art binarization methods for LLMs. ARB-LLM_RC is the first binary PTQ method to surpass FP16 models of the same size.

Conclusion: ARB-LLM effectively addresses key challenges in LLM binarization through refined algorithms and column-aware strategies, enabling practical deployment of compressed LLMs without sacrificing performance.

Abstract: Large Language Models (LLMs) have greatly pushed forward advancements in natural language processing, yet their high memory and computational demands hinder practical deployment. Binarization, as an effective compression technique, can shrink model weights to just 1 bit, significantly reducing the high demands on computation and memory. However, current binarization methods struggle to narrow the distribution gap between binarized and full-precision weights, while also overlooking the column deviation in LLM weight distribution. To tackle these issues, we propose ARB-LLM, a novel 1-bit post-training quantization (PTQ) technique tailored for LLMs. To narrow the distribution shift between binarized and full-precision weights, we first design an alternating refined binarization (ARB) algorithm to progressively update the binarization parameters, which significantly reduces the quantization error. Moreover, considering the pivot role of calibration data and the column deviation in LLM weights, we further extend ARB to ARB-X and ARB-RC. In addition, we refine the weight partition strategy with column-group bitmap (CGB), which further enhance performance. Equipping ARB-X and ARB-RC with CGB, we obtain ARB-LLM$\text{X}$ and ARB-LLM$\text{RC}$ respectively, which significantly outperform state-of-the-art (SOTA) binarization methods for LLMs. As a binary PTQ method, our ARB-LLM$_\text{RC}$ is the first to surpass FP16 models of the same size. The code and models will be available at https://github.com/ZHITENGLI/ARB-LLM.

Method: Proposes 2DMamba, a 2D selective SSM framework that incorporates 2D spatial structure into Mamba with a highly optimized hardware-aware operator, balancing spatial continuity and computational efficiency.

Result: On 10 public WSI datasets: improvements up to 2.48% AUC, 3.11% F1, 2.47% accuracy, 5.52% C-index. For natural images: 0.5-0.7 mIoU improvement on ADE20k segmentation, 0.2% accuracy improvement on ImageNet-1K.

Conclusion: 2DMamba successfully addresses the spatial discrepancy problem in extending 1D SSMs to vision tasks while maintaining computational efficiency, demonstrating strong performance across medical imaging and natural image tasks.

Abstract: Efficiently modeling large 2D contexts is essential for various fields including Giga-Pixel Whole Slide Imaging (WSI) and remote sensing. Transformer-based models offer high parallelism but face challenges due to their quadratic complexity for handling long sequences. Recently, Mamba introduced a selective State Space Model (SSM) with linear complexity and high parallelism, enabling effective and efficient modeling of wide context in 1D sequences. However, extending Mamba to vision tasks, which inherently involve 2D structures, results in spatial discrepancies due to the limitations of 1D sequence processing. On the other hand, current 2D SSMs inherently model 2D structures but they suffer from prohibitively slow computation due to the lack of efficient parallel algorithms. In this work, we propose 2DMamba, a novel 2D selective SSM framework that incorporates the 2D spatial structure of images into Mamba, with a highly optimized hardware-aware operator, adopting both spatial continuity and computational efficiency. We validate the versatility of our approach on both WSIs and natural images. Extensive experiments on 10 public datasets for WSI classification and survival analysis show that 2DMamba improves up to 2.48% in AUC, 3.11% in F1 score, 2.47% in accuracy and 5.52% in C-index. Additionally, integrating our method with VMamba for natural imaging yields 0.5 to 0.7 improvements in mIoU on the ADE20k semantic segmentation dataset, and 0.2% accuracy improvement on ImageNet-1K classification dataset. Our code is available at https://github.com/AtlasAnalyticsLab/2DMamba.

Method: Comparative analysis of native and non-native multimodal VLMs, examining information flow patterns, embedding separations in residual streams, and conducting ablation studies on key tokens.

Result: Native VLMs have more separated image/text embeddings and rely on a single post-image token for visual information transfer, while non-native models use distributed communication through multiple tokens. Ablating the single token significantly hurts performance.

Conclusion: Different multimodal training approaches lead to distinct visual information processing architectures, with native models using a narrow gate mechanism that enables fine-grained control through targeted interventions.

Abstract: Recent advances in multimodal training have significantly improved the integration of image understanding and generation within a unified model. This study investigates how vision-language models (VLMs) handle image-understanding tasks, focusing on how visual information is processed and transferred to the textual domain. We compare native multimodal VLMs, models trained from scratch on multimodal data to generate both text and images, and non-native multimodal VLMs, models adapted from pre-trained large language models or capable of generating only text, highlighting key differences in information flow. We find that in native multimodal VLMs, image and text embeddings are more separated within the residual stream. Moreover, VLMs differ in how visual information reaches text: non-native multimodal VLMs exhibit a distributed communication pattern, where information is exchanged through multiple image tokens, whereas models trained natively for joint image and text generation tend to rely on a single post-image token that acts as a narrow gate for visual information. We show that ablating this single token significantly deteriorates image-understanding performance, whereas targeted, token-level interventions reliably steer image semantics and downstream text with fine-grained control.

Method: Proposes a two-stage framework: 1) First stage generates low-resolution video with large parameters and sufficient NFEs for prompt fidelity, 2) Second stage uses flow matching to create nearly straight ODE trajectory between low and high resolutions, generating fine details with minimal NFEs. Includes careful degradation strategies for seamless stage connection.

Result: Achieves state-of-the-art high-resolution video generation with superior computational efficiency. Enables preview capability where users can adjust prompts before full-resolution generation, reducing computational costs and wait times.

Conclusion: FlashVideo successfully addresses computational efficiency challenges in high-resolution text-to-video generation through strategic two-stage design, balancing fidelity and quality while enabling practical commercial applications.

Abstract: DiT models have achieved great success in text-to-video generation, leveraging their scalability in model capacity and data scale. High content and motion fidelity aligned with text prompts, however, often require large model parameters and a substantial number of function evaluations (NFEs). Realistic and visually appealing details are typically reflected in high-resolution outputs, further amplifying computational demands-especially for single-stage DiT models. To address these challenges, we propose a novel two-stage framework, FlashVideo, which strategically allocates model capacity and NFEs across stages to balance generation fidelity and quality. In the first stage, prompt fidelity is prioritized through a low-resolution generation process utilizing large parameters and sufficient NFEs to enhance computational efficiency. The second stage achieves a nearly straight ODE trajectory between low and high resolutions via flow matching, effectively generating fine details and fixing artifacts with minimal NFEs. To ensure a seamless connection between the two independently trained stages during inference, we carefully design degradation strategies during the second-stage training. Quantitative and visual results demonstrate that FlashVideo achieves state-of-the-art high-resolution video generation with superior computational efficiency. Additionally, the two-stage design enables users to preview the initial output and accordingly adjust the prompt before committing to full-resolution generation, thereby significantly reducing computational costs and wait times as well as enhancing commercial viability.

Method: SPEED directly edits model parameters by searching for a null space where parameter updates don’t affect non-target concepts. Uses three strategies: Influence-based Prior Filtering (IPF) to retain most affected non-target concepts, Directed Prior Augmentation (DPA) to enrich retain set with semantic variations, and Invariant Equality Constraints (IEC) to preserve key T2I generation invariants.

Result: SPEED outperforms existing methods in non-target concept preservation while achieving efficient high-fidelity erasure. Can erase 100 concepts within only 5 seconds across multiple concept erasure tasks.

Conclusion: SPEED provides a scalable and precise solution for concept erasure in text-to-image diffusion models, addressing efficiency and quality preservation challenges through null space optimization with complementary strategies.

Abstract: Erasing concepts from large-scale text-to-image (T2I) diffusion models has become increasingly crucial due to the growing concerns over copyright infringement, offensive content, and privacy violations. In scalable applications, fine-tuning-based methods are time-consuming to precisely erase multiple target concepts, while real-time editing-based methods often degrade the generation quality of non-target concepts due to conflicting optimization objectives. To address this dilemma, we introduce SPEED, an efficient concept erasure approach that directly edits model parameters. SPEED searches for a null space, a model editing space where parameter updates do not affect non-target concepts, to achieve scalable and precise erasure. To facilitate accurate null space optimization, we incorporate three complementary strategies: Influence-based Prior Filtering (IPF) to selectively retain the most affected non-target concepts, Directed Prior Augmentation (DPA) to enrich the filtered retain set with semantically consistent variations, and Invariant Equality Constraints (IEC) to preserve key invariants during the T2I generation process. Extensive evaluations across multiple concept erasure tasks demonstrate that SPEED consistently outperforms existing methods in non-target preservation while achieving efficient and high-fidelity concept erasure, successfully erasing 100 concepts within only 5 seconds. Our code and models are available at: https://github.com/Ouxiang-Li/SPEED.

Method: 1) Hierarchical prompt-guided visual feature extraction at three complementary levels emphasizing question-relevant information; 2) Dynamic fusion and injection into LLM; 3) Data Efficient GRPO reinforcement learning strategy that identifies high-utility samples and maximizes per-instance contribution under limited supervision.

Result: FaVChat consistently outperforms existing VLLMs in extensive experiments, including zero-shot evaluations on four facial understanding tasks. A large-scale benchmark dataset FaVChat-170K was created with 60K facial videos and 170K QA pairs.

Conclusion: FaVChat successfully addresses the limitations of prompt-agnostic visual encoders in VLLMs by introducing targeted facial feature extraction and efficient learning strategies, enabling superior reasoning over subtle facial details.

Abstract: Existing video large language models (VLLMs) primarily leverage prompt agnostic visual encoders, which extract untargeted facial representations without awareness of the queried information, leading to the loss of task critical cues. To address this challenge, we propose FaVChat, the first VLLM designed for reasoning over subtle visual and dynamic facial cues. FaVChat introduces a hierarchical, prompt guided visual feature extraction framework that emphasizes question relevant information at three complementary levels. These multi level features are dynamically fused and injected into the LLM, enabling more accurate facial details reasoning To further improve learning efficiency under data scarcity, we propose Data Efficient GRPO, a reinforcement learning strategy that iteratively identifies high utility samples and maximizes the contribution of each instance via per instance utility estimation, substantially enhancing performance gains under limited supervision. We construct a large scale benchmark dataset FaVChat 170K, comprising approximately 60K high quality facial videos and 170K question answer pairs focusing on fine grained facial details. Extensive experiments, including zero shot evaluations on four facial understanding tasks, demonstrate that FaVChat consistently outperforms existing VLLMs.

Method: 1. Extract physical clues and contextual information from real-world accident reports. 2. Use a reliable physical simulator to replicate post-collision trajectories and build a collision trajectory dataset. 3. Fine-tune a language model to predict physically consistent trajectories from user descriptions. 4. Use Neural Radiance Fields (NeRF) to render high-quality backgrounds and merge with foreground vehicles.

Result: Experimental results show that AccidentSim produces videos that excel in both visual and physical authenticity compared to existing methods.

Conclusion: AccidentSim provides a novel framework for generating physically realistic vehicle collision videos by combining physical simulation with neural rendering, addressing limitations of existing video generation methods for autonomous driving research.

Abstract: Collecting real-world vehicle accident videos for autonomous driving research is challenging due to their rarity and complexity. While existing driving video generation methods may produce visually realistic videos, they often fail to deliver physically realistic simulations because they lack the capability to generate accurate post-collision trajectories. In this paper, we introduce AccidentSim, a novel framework that generates physically realistic vehicle collision videos by extracting and utilizing the physical clues and contextual information available in real-world vehicle accident reports. Specifically, AccidentSim leverages a reliable physical simulator to replicate post-collision vehicle trajectories from the physical and contextual information in the accident reports and to build a vehicle collision trajectory dataset. This dataset is then used to fine-tune a language model, enabling it to respond to user prompts and predict physically consistent post-collision trajectories across various driving scenarios based on user descriptions. Finally, we employ Neural Radiance Fields (NeRF) to render high-quality backgrounds, merging them with the foreground vehicles that exhibit physically realistic trajectories to generate vehicle collision videos. Experimental results demonstrate that the videos produced by AccidentSim excel in both visual and physical authenticity.

Method: Rigorous literature review framework covering over 80 VLA models published in the past three years, organized across five thematic pillars: conceptual foundations, architectural innovations, training strategies, inference accelerations, and application domains.

Result: Comprehensive analysis of VLA landscape including progress in architectural designs, efficient training methods, real-time inference techniques, and diverse applications in autonomous vehicles, robotics, agriculture, and augmented reality.

Conclusion: VLA models represent a crucial step toward general-purpose embodied agents, with future directions pointing toward convergence of VLA models, VLMs, and agentic AI for socially aligned, adaptive robotics and artificial general intelligence.

Abstract: Vision-Language-Action (VLA) models mark a transformative advancement in artificial intelligence, aiming to unify perception, natural language understanding, and embodied action within a single computational framework. This foundational review presents a comprehensive synthesis of recent advancements in Vision-Language-Action models, systematically organized across five thematic pillars that structure the landscape of this rapidly evolving field. We begin by establishing the conceptual foundations of VLA systems, tracing their evolution from cross-modal learning architectures to generalist agents that tightly integrate vision-language models (VLMs), action planners, and hierarchical controllers. Our methodology adopts a rigorous literature review framework, covering over 80 VLA models published in the past three years. Key progress areas include architectural innovations, efficient training strategies, and real-time inference accelerations. We explore diverse application domains such as autonomous vehicles, medical and industrial robotics, precision agriculture, humanoid robotics, and augmented reality. We analyzed challenges and propose solutions including agentic adaptation and cross-embodiment planning. Furthermore, we outline a forward-looking roadmap where VLA models, VLMs, and agentic AI converge to strengthen socially aligned, adaptive, and general-purpose embodied agents. This work, is expected to serve as a foundational reference for advancing intelligent, real-world robotics and artificial general intelligence. The project repository is available on GitHub as https://github.com/Applied-AI-Research-Lab/Vision-Language-Action-Models-Concepts-Progress-Applications-and-Challenges. [Index Terms: Vision Language Action, VLA, Vision Language Models, VLMs, Action Tokenization, NLP]

Method: Uses Vision Language Model to detect target objects in direction-zoomed Street View Images. Successful detection indicates object visibility at that location. Constructs heterogeneous visibility graph to model complex observer-target interactions.

Result: 87% accuracy detecting visibility of six tall landmarks in global cities. Revealed contextual differences in landmark perception. Visibility graph uncovered connection patterns for Thames landmarks, with bridges accounting for ~30% of connections.

Conclusion: Method complements traditional LoS-based analysis, enables revealing connections between visual objects in urban environments, opens new research perspectives for urban planning, heritage conservation, and computational social science.

Abstract: Visibility analysis is one of the fundamental analytics methods in urban planning and landscape research, traditionally conducted through computational simulations based on the Line-of-Sight (LoS) principle. However, when assessing the visibility of named urban objects such as landmarks, geometric intersection alone fails to capture the contextual and perceptual dimensions of visibility as experienced in the real world. The study challenges the traditional LoS-based approaches by introducing a new, image-based visibility analysis method. Specifically, a Vision Language Model (VLM) is applied to detect the target object within a direction-zoomed Street View Image (SVI). Successful detection represents the object’s visibility at the corresponding SVI location. Further, a heterogeneous visibility graph is constructed to address the complex interaction between observers and target objects. In the first case study, the method proves its reliability in detecting the visibility of six tall landmark constructions in global cities, with an overall accuracy of 87%. Furthermore, it reveals broader contextual differences when the landmarks are perceived and experienced. In the second case, the proposed visibility graph uncovers the form and strength of connections for multiple landmarks along the River Thames in London, as well as the places where these connections occur. Notably, bridges on the River Thames account for approximately 30% of total connections. Our method complements and enhances traditional LoS-based visibility analysis, and showcases the possibility of revealing the prevalent connection of any visual objects in the urban environment. It opens up new research perspectives for urban planning, heritage conservation, and computational social science.

Method: Two-stage flow-based approach: 1) Precompute and cache results from an unconditional flow-based generative model that transforms Gaussian mixture to future motion density; 2) Use lightweight model to map historical trajectories to Gaussian mixture samples for conditional prediction.

Result: Achieves ~1ms inference time (4x faster than VAE methods, 30x faster than diffusion methods), improved density estimation accuracy, and comparable prediction accuracy to SOTA on Human3.6M and AMASS datasets.

Conclusion: CacheFlow enables fast, accurate 3D human motion prediction through efficient caching and two-stage architecture, making real-time applications feasible.

Abstract: Many density estimation techniques for 3D human motion prediction require a significant amount of inference time, often exceeding the duration of the predicted time horizon. To address the need for faster density estimation for 3D human motion prediction, we introduce a novel flow-based method for human motion prediction called CacheFlow. Unlike previous conditional generative models that suffer from poor time efficiency, CacheFlow takes advantage of an unconditional flow-based generative model that transforms a Gaussian mixture into the density of future motions. The results of the computation of the flow-based generative model can be precomputed and cached. Then, for conditional prediction, we seek a mapping from historical trajectories to samples in the Gaussian mixture. This mapping can be done by a much more lightweight model, thus saving significant computation overhead compared to a typical conditional flow model. In such a two-stage fashion and by caching results from the slow flow model computation, we build our CacheFlow without loss of prediction accuracy and model expressiveness. This inference process is completed in approximately one millisecond, making it 4 times faster than previous VAE methods and 30 times faster than previous diffusion-based methods on standard benchmarks such as Human3.6M and AMASS datasets. Furthermore, our method demonstrates improved density estimation accuracy and comparable prediction accuracy to a SOTA method on Human3.6M. Our code and models are available at https://github.com/meaten/CacheFlow.

Method: Theoretically demonstrate that structurally re-parameterized blocks induce spatially varying learning rates crucial for convergence. Introduce Rep3D framework that employs a lightweight modulation network to generate receptive-biased scaling masks, adaptively re-weighting kernel updates within a plain encoder architecture. This avoids multi-branch design complexity while ensuring robust local-to-global convergence.

Result: Extensive evaluations on five 3D segmentation benchmarks demonstrate that Rep3D consistently outperforms state-of-the-art transformer and fixed-prior baselines.

Conclusion: Rep3D provides an effective framework for stable optimization of large kernel convolutions in 3D volumetric analysis, unifying spatial inductive bias with optimization-aware learning and outperforming existing approaches.

Abstract: Large kernel convolutions offer a scalable alternative to vision transformers for high-resolution 3D volumetric analysis, yet naively increasing kernel size often leads to optimization instability. Motivated by the spatial bias inherent in effective receptive fields (ERFs), we theoretically demonstrate that structurally re-parameterized blocks induce spatially varying learning rates that are crucial for convergence. Leveraging this insight, we introduce Rep3D, a framework that employs a lightweight modulation network to generate receptive-biased scaling masks, adaptively re-weighting kernel updates within a plain encoder architecture. This approach unifies spatial inductive bias with optimization-aware learning, avoiding the complexity of multi-branch designs while ensuring robust local-to-global convergence. Extensive evaluations on five 3D segmentation benchmarks demonstrate that Rep3D consistently outperforms state-of-the-art transformer and fixed-prior baselines. The source code is publicly available at https://github.com/leeh43/Rep3D.

Method: Two-stage framework: (1) Complementary global and local scans with token merging during visual encoding to address redundancy, (2) Pruning at intermediate layers of the language model based on empirical studies of how visual tokens are processed.

Result: Achieves 2.91× speedup in prefilling and 10× FLOPs reduction for LLaVA-NeXT-7B while retaining 95.4% of original performance. Validated across four LVLMs on sixteen benchmarks, outperforming current SOTA methods.

Conclusion: VScan effectively reduces computational overhead of LVLMs while maintaining performance, enabling more efficient multimodal understanding for real-time applications.

Abstract: Recent Large Vision-Language Models (LVLMs) have advanced multi-modal understanding by incorporating finer-grained visual perception and encoding. However, such methods incur significant computational costs due to longer visual token sequences, posing challenges for real-time deployment. To mitigate this, prior studies have explored pruning unimportant visual tokens either at the output layer of the visual encoder or at the early layers of the language model. In this work, we revisit these design choices and reassess their effectiveness through comprehensive empirical studies of how visual tokens are processed throughout the visual encoding and language decoding stages. Guided by these insights, we propose VScan, a two-stage visual token reduction framework that addresses token redundancy by: (1) integrating complementary global and local scans with token merging during visual encoding, and (2) introducing pruning at intermediate layers of the language model. Extensive experimental results across four LVLMs validate the effectiveness of VScan in accelerating inference and demonstrate its superior performance over current state-of-the-arts on sixteen benchmarks. Notably, when applied to LLaVA-NeXT-7B, VScan achieves a 2.91$\times$ speedup in prefilling and a 10$\times$ reduction in FLOPs, while retaining 95.4% of the original performance. Code is available at https://github.com/Tencent/SelfEvolvingAgent/tree/main/VScan.

Method: Proposes first training-free weight update framework for concept removal in video diffusion models. Uses 5 paired safe/unsafe prompts to estimate a refusal vector integrated into model weights as closed-form update. Contrastive low-rank factorization disentangles target concept from unrelated semantics for selective suppression.

Result: Reduces unsafe generations on Open-Sora and ZeroScopeT2V models across T2VSafetyBench and SafeSora benchmarks with average reductions of 36.3% and 58.2% respectively, while preserving prompt alignment and video quality.

Conclusion: Establishes efficient and scalable solution for safe video generation without retraining or inference overhead, enabling concept removal through training-free weight updates.

Abstract: Video generative models achieve high-quality synthesis from natural-language prompts by leveraging large-scale web data. However, this training paradigm inherently exposes them to unsafe biases and harmful concepts, introducing the risk of generating undesirable or illicit content. To mitigate unsafe generations, existing machine unlearning approaches either rely on filtering, and can therefore be bypassed, or they update model weights, but with costly fine-tuning or training-free closed-form edits. We propose the first training-free weight update framework for concept removal in video diffusion models. From five paired safe/unsafe prompts, our method estimates a refusal vector and integrates it into the model weights as a closed-form update. A contrastive low-rank factorization further disentangles the target concept from unrelated semantics, it ensures a selective concept suppression and it does not harm generation quality. Our approach reduces unsafe generations on the Open-Sora and ZeroScopeT2V models across the T2VSafetyBench and SafeSora benchmarks, with average reductions of 36.3% and 58.2% respectively, while preserving prompt alignment and video quality. This establishes an efficient and scalable solution for safe video generation without retraining nor any inference overhead. Project page: https://www.pinlab.org/video-unlearning.

Method: Introduces Symmetrical Flow Matching (SymmFlow) with symmetric learning objective that jointly models forward and reverse transformations, ensuring bi-directional consistency while preserving entropy for diversity. Includes explicit semantic information retention across flows and supports flexible conditioning with both pixel-level (masks) and image-level (class labels) inputs.

Result: Achieves state-of-the-art FID scores of 11.9 on CelebAMask-HQ and 7.0 on COCO-Stuff with only 25 inference steps for semantic image synthesis. Also delivers competitive semantic segmentation results and shows promising classification capabilities.

Conclusion: SymmFlow successfully unifies generative and discriminative tasks within a single flow matching framework, enabling efficient one-step inference for segmentation/classification while maintaining high-quality image generation with flexible conditioning.

Abstract: Flow Matching has emerged as a powerful framework for learning continuous transformations between distributions, enabling high-fidelity generative modeling. This work introduces Symmetrical Flow Matching (SymmFlow), a new formulation that unifies semantic segmentation, classification, and image generation within a single model. Using a symmetric learning objective, SymmFlow models forward and reverse transformations jointly, ensuring bi-directional consistency, while preserving sufficient entropy for generative diversity. A new training objective is introduced to explicitly retain semantic information across flows, featuring efficient sampling while preserving semantic structure, allowing for one-step segmentation and classification without iterative refinement. Unlike previous approaches that impose strict one-to-one mapping between masks and images, SymmFlow generalizes to flexible conditioning, supporting both pixel-level and image-level class labels. Experimental results on various benchmarks demonstrate that SymmFlow achieves state-of-the-art performance on semantic image synthesis, obtaining FID scores of 11.9 on CelebAMask-HQ and 7.0 on COCO-Stuff with only 25 inference steps. Additionally, it delivers competitive results on semantic segmentation and shows promising capabilities in classification tasks.

Method: Three key modifications to ORB-SLAM3: (1) Replace ORB with self-supervised SuperPoint detector-descriptor, (2) Enforce spatially uniform keypoints via adaptive non-maximal suppression (ANMS), (3) Integrate lightweight NetVLAD place-recognition head for learning-based loop closure.

Result: On KITTI Odometry: Mean translational error reduced from 4.15% to 0.34%, mean rotational error from 0.0027 deg/m to 0.0010 deg/m. On EuRoC MAV: Roughly halves errors across all sequences (e.g., V2_03: 1.58% to 0.79%). Preserves real-time operation.

Conclusion: Fusing modern deep features (SuperPoint) with learned loop-closure (NetVLAD) significantly improves ORB-SLAM3 accuracy while maintaining real-time performance, demonstrating the value of integrating deep learning into traditional SLAM pipelines.

Abstract: Visual simultaneous localization and mapping (SLAM) must remain accurate under extreme viewpoint, scale and illumination variations. The widely adopted ORB-SLAM3 falters in these regimes because it relies on hand-crafted ORB keypoints. We introduce SuperPoint-SLAM3, a drop-in upgrade that (i) replaces ORB with the self-supervised SuperPoint detector–descriptor, (ii) enforces spatially uniform keypoints via adaptive non-maximal suppression (ANMS), and (iii) integrates a lightweight NetVLAD place-recognition head for learning-based loop closure. On the KITTI Odometry benchmark SuperPoint-SLAM3 reduces mean translational error from 4.15% to 0.34% and mean rotational error from 0.0027 deg/m to 0.0010 deg/m. On the EuRoC MAV dataset it roughly halves both errors across every sequence (e.g., V2_03: 1.58% -> 0.79%). These gains confirm that fusing modern deep features with a learned loop-closure module markedly improves ORB-SLAM3 accuracy while preserving its real-time operation. Implementation, pretrained weights and reproducibility scripts are available at https://github.com/shahram95/SuperPointSLAM3.

Method: SnS frames transformations as bi-objective optimization problems: for invariance, it seeks perturbations that maximally alter representations while preserving unit activation; for adversarial sensitivity, it reverses stretching and squeezing to perturb activation while minimizing representation changes.

Result: SnS revealed invariant transformations farther from reference images than affine transformations while better preserving target unit responses. Different processing stages produced different invariant images: pixel-level changes affected luminance/contrast, while mid/late-layer changes altered texture/pose. Robust networks showed interpretability drops when stretching deep layers.

Conclusion: SnS provides a systematic way to characterize invariant stimuli and adversarial vulnerabilities in visual systems, revealing important differences between standard and robust models in how interpretability changes across processing stages.

Abstract: Uncovering which feature combinations are encoded by visual units is critical to understanding how images are transformed into representations that support recognition. While existing feature visualization approaches typically infer a unit’s most exciting images, this is insufficient to reveal the manifold of transformations under which responses remain invariant, which is critical to generalization in vision. Here we introduce Stretch-and-Squeeze (SnS), a model-agnostic, gradient-free framework to systematically characterize a unit’s maximally invariant stimuli, and its vulnerability to adversarial perturbations, in both biological and artificial visual systems. SnS frames these transformations as bi-objective optimization problems. To probe invariance, SnS seeks image perturbations that maximally alter (stretch) the representation of a reference stimulus in a given processing stage while preserving unit activation downstream (squeeze). To probe adversarial sensitivity, stretching and squeezing are reversed to maximally perturb unit activation while minimizing changes to the upstream representation. Applied to CNNs, SnS revealed invariant transformations that were farther from a reference image in pixel-space than those produced by affine transformations, while more strongly preserving the target unit’s response. The discovered invariant images differed depending on the stage of the image representation used for optimization: pixel-level changes primarily affected luminance and contrast, while stretching mid- and late-layer representations mainly altered texture and pose. By measuring how well the hierarchical invariant images obtained for L2 robust networks were classified by humans and other observer networks, we discovered a substantial drop in their interpretability when the representation was stretched in deep layers, while the opposite trend was found for standard models.

Method: Reframes lighting estimation as chrome ball inpainting using Stable Diffusion XL. Uses iterative inpainting to compute median chrome ball as lighting prior, with Exposure LoRA for HDR generation. DiffusionLight-Turbo trains Turbo LoRA to directly predict averaged chrome balls for 60x speedup.

Result: Produces convincing light estimates across diverse settings with superior generalization to in-the-wild scenarios. DiffusionLight takes ~30 minutes per estimation, while DiffusionLight-Turbo reduces runtime to ~30 seconds with minimal quality loss.

Conclusion: The approach effectively leverages diffusion models for lighting estimation, with the turbo variant making it practical for real-world applications through significant speed improvements.

Abstract: We introduce a simple yet effective technique for estimating lighting from a single low-dynamic-range (LDR) image by reframing the task as a chrome ball inpainting problem. This approach leverages a pre-trained diffusion model, Stable Diffusion XL, to overcome the generalization failures of existing methods that rely on limited HDR panorama datasets. While conceptually simple, the task remains challenging because diffusion models often insert incorrect or inconsistent content and cannot readily generate chrome balls in HDR format. Our analysis reveals that the inpainting process is highly sensitive to the initial noise in the diffusion process, occasionally resulting in unrealistic outputs. To address this, we first introduce DiffusionLight, which uses iterative inpainting to compute a median chrome ball from multiple outputs to serve as a stable, low-frequency lighting prior that guides the generation of a high-quality final result. To generate high-dynamic-range (HDR) light probes, an Exposure LoRA is fine-tuned to create LDR images at multiple exposure values, which are then merged. While effective, DiffusionLight is time-intensive, requiring approximately 30 minutes per estimation. To reduce this overhead, we introduce DiffusionLight-Turbo, which reduces the runtime to about 30 seconds with minimal quality loss. This 60x speedup is achieved by training a Turbo LoRA to directly predict the averaged chrome balls from the iterative process. Inference is further streamlined into a single denoising pass using a LoRA swapping technique. Experimental results that show our method produces convincing light estimates across diverse settings and demonstrates superior generalization to in-the-wild scenarios. Our code is available at https://diffusionlight.github.io/turbo

Method: Proposes MAT (Map Association Transformer) with path-aware attention to handle spatial fluctuations and semantic disparities, and spatial attention to integrate noisy OP features via global context. Also introduces OMA dataset with 30K scenarios and NR P-R evaluation metric.

Result: MAT outperforms existing methods with 34 ms latency, enabling low-cost and up-to-date lane-level navigation. The method effectively handles many-to-one lane-to-road mappings despite spatial misalignment and OP map noise.

Conclusion: ONR provides a practical solution for lane-level navigation by combining SD map topology with OP map real-time geometry, making accurate navigation more accessible and adaptable to dynamic road conditions.

Abstract: Lane-level navigation is critical for geographic information systems and navigation-based tasks, offering finer-grained guidance than road-level navigation by standard definition (SD) maps. However, it currently relies on expansive global HD maps that cannot adapt to dynamic road conditions. Recently, online perception (OP) maps have become research hotspots, providing real-time geometry as an alternative, but lack the global topology needed for navigation. To address these issues, Online Navigation Refinement (ONR), a new mission is introduced that refines SD-map-based road-level routes into accurate lane-level navigation by associating SD maps with OP maps. The map-to-map association to handle many-to-one lane-to-road mappings under two key challenges: (1) no public dataset provides lane-to-road correspondences; (2) severe misalignment from spatial fluctuations, semantic disparities, and OP map noise invalidates traditional map matching. For these challenges, We contribute: (1) Online map association dataset (OMA), the first ONR benchmark with 30K scenarios and 2.6M annotated lane vectors; (2) MAT, a transformer with path-aware attention to aligns topology despite spatial fluctuations and semantic disparities and spatial attention for integrates noisy OP features via global context; and (3) NR P-R, a metric evaluating geometric and semantic alignment. Experiments show that MAT outperforms existing methods at 34 ms latency, enabling low-cost and up-to-date lane-level navigation.

Method: Analyze attention patterns in VLMs processing image series to identify sparse patterns. Categorize attention heads into Dense, Sink, Intra-Image, and Intra-Image+Sink types. Develop Triton-based GPU kernel that applies input-template-aware attention sparsity masks with zero runtime overhead.

Result: Achieves 1.8-3.2x speedup in attention computation for prompts of length 36K-300K. Generalizes across VLMs (Qwen2-VL, Qwen2.5-VL, Gemma 3) with only 0.78% absolute accuracy degradation on multi-image comprehension benchmarks.

Conclusion: BlindSight effectively optimizes multi-image VLM inference by leveraging attention sparsity. The approach advocates for designing efficient VLMs that combine sparse and dense layers inspired by these findings.

Abstract: Large vision-language models (VLMs) enable joint processing of text and images. However, incorporating vision data significantly increases the prompt length, resulting in a longer time to first token (TTFT). This bottleneck can be alleviated by leveraging the inherent sparsity in the attention computation. Analyzing these attention patterns in VLMs when processing a series of images, we observe the absence of inter-image attention in a substantial portion of layers. Based on this, we propose BlindSight: an approach to optimize multi-image VLM inference using an input-template-aware attention sparsity mask with no runtime overhead. We utilize a dataset to derive a prompt-agnostic categorization for attention heads: Dense, Sink, Intra-Image, and Intra-Image+Sink. We develop a Triton-based GPU kernel to leverage this sparsity. BlindSight achieves a 1.8-3.2x speedup in the attention computation (prompt length 36K-300K). BlindSight generalizes across VLMs (Qwen2-VL, Qwen2.5-VL, Gemma 3), with only a 0.78% absolute accuracy degradation on average on multi-image comprehension benchmarks. Finally, we advocate for the design of efficient VLMs that combine BlindSight-inspired sparse and dense layers.

Method: Investigated foundation models for mitotic figure classification using linear probing and LoRA adaptation. Compared against end-to-end trained CNNs and Vision Transformers. Evaluated data scaling laws and robustness to unseen tumor domains.

Result: LoRA-adapted foundation models outperformed linear probing, achieving performance close to 100% data availability with only 10% training data. LoRA adaptation of recent foundation models nearly closed the out-of-domain performance gap on unseen tumor domains.

Conclusion: LoRA adaptation of foundation models is highly effective for medical imaging tasks with limited data, offering superior performance and domain robustness compared to linear probing, though traditional full fine-tuning remains competitive.

Abstract: The performance of deep learning models is known to scale with data quantity and diversity. In pathology, as in many other medical imaging domains, the availability of labeled images for a specific task is often limited. Self-supervised learning techniques have enabled the use of vast amounts of unlabeled data to train large-scale neural networks, i.e., foundation models, that can address the limited data problem by providing semantically rich feature vectors that can generalize well to new tasks with minimal training effort increasing model performance and robustness. In this work, we investigate the use of foundation models for mitotic figure classification. The mitotic count, which can be derived from this classification task, is an independent prognostic marker for specific tumors and part of certain tumor grading systems. In particular, we investigate the data scaling laws on multiple current foundation models and evaluate their robustness to unseen tumor domains. Next to the commonly used linear probing paradigm, we also adapt the models using low-rank adaptation (LoRA) of their attention mechanisms. We compare all models against end-to-end-trained baselines, both CNNs and Vision Transformers. Our results demonstrate that LoRA-adapted foundation models provide superior performance to those adapted with standard linear probing, reaching performance levels close to 100% data availability with only 10% of training data. Furthermore, LoRA-adaptation of the most recent foundation models almost closes the out-of-domain performance gap when evaluated on unseen tumor domains. However, full fine-tuning of traditional architectures still yields competitive performance.

Method: Rechecked builds on existing label error detection methods by reviewing their error proposals through lightweight, crowd-sourced microtasks. The framework combines automated detection with human verification for scalable correction.

Result: Applied to KITTI dataset’s pedestrian class, Rechecked detected 18% of missing and inaccurate labels in original ground truth. Current detection methods combined with this framework can recover hundreds of errors with minimal human effort compared to annotation from scratch, though best methods still miss up to 66% of label errors.

Conclusion: The Rechecked framework enables scalable label error correction with reduced human effort, but significant gaps remain in detection methods, motivating further research. The released benchmark supports future work in this area.

Abstract: Object detection has advanced rapidly in recent years, driven by increasingly large and diverse datasets. However, label errors often compromise the quality of these datasets and affect the outcomes of training and benchmark evaluations. Although label error detection methods for object detection datasets now exist, they are typically validated only on synthetic benchmarks or via limited manual inspection. How to correct such errors systematically and at scale remains an open problem. We introduce a semi-automated framework for label error correction called Rechecked. Building on existing label error detection methods, their error proposals are reviewed with lightweight, crowd-sourced microtasks. We apply Rechecked to the class pedestrian in the KITTI dataset, for which we crowdsourced high-quality corrected annotations. We detect 18% of missing and inaccurate labels in the original ground truth. We show that current label error detection methods, when combined with our correction framework, can recover hundreds of errors with little human effort compared to annotation from scratch. However, even the best methods still miss up to 66% of the label errors, which motivates further research, now enabled by our released benchmark.

Method: Decomposes rigid transformation using Chasles’ theorem into translation along rotation axis and 2D rigid transformation. Uses rotation-only BnB search with range maximum query (RMQ) problems. Searches top-k candidate rotation axes via cube mapping, estimates translation through interval stabbing, and solves 2D registration as 1D rotation angle search with 2D RMQ using sweep line algorithm with segment tree.

Result: Demonstrates superior accuracy and efficiency over SOTA methods on indoor 3DMatch/3DLoMatch scanning and outdoor KITTI LiDAR datasets. Achieves polynomial time complexity and linear space complexity with number of points.

Conclusion: Proposed geometric maximum overlapping registration framework via rotation-only BnB search provides accurate and efficient point cloud registration with polynomial time complexity, outperforming existing methods on both indoor and outdoor datasets.

Abstract: Point cloud registration based on correspondences computes the rigid transformation that maximizes the number of inliers constrained within the noise threshold. Current state-of-the-art (SOTA) methods employing spatial compatibility graphs or branch-and-bound (BnB) search mainly focus on registration under high outlier ratios. However, graph-based methods require at least quadratic space and time complexity for graph construction, while multi-stage BnB search methods often suffer from inaccuracy due to local optima between decomposed stages. This paper proposes a geometric maximum overlapping registration framework via rotation-only BnB search. The rigid transformation is decomposed using Chasles’ theorem into a translation along rotation axis and a 2D rigid transformation. The optimal rotation axis and angle are searched via BnB, with residual parameters formulated as range maximum query (RMQ) problems. Firstly, the top-k candidate rotation axes are searched within a hemisphere parameterized by cube mapping, and the translation along each axis is estimated through interval stabbing of the correspondences projected onto that axis. Secondly, the 2D registration is relaxed to 1D rotation angle search with 2D RMQ of geometric overlapping for axis-aligned rectangles, which is solved deterministically in polynomial time using sweep line algorithm with segment tree. Experimental results on indoor 3DMatch/3DLoMatch scanning and outdoor KITTI LiDAR datasets demonstrate superior accuracy and efficiency over SOTA methods, while the time complexity is polynomial and the space complexity increases linearly with the number of points, even in the worst case.

Method: Used pretrained convolutional and transformer vision models adapted via transfer learning to predict group-level perceived fear for spider-related images, evaluated on new people and new images.

Result: Models achieved mean absolute error below 10 units on 0-100 fear scale, with predictions driven by spider-specific regions. Transformer models were data efficient and approached performance saturation with ~300 images.

Conclusion: Establishes transparent, data-driven fear estimation from images, laying groundwork for adaptive digital mental health tools.

Abstract: Phobias are common and impairing, and exposure therapy, which involves confronting patients with fear-provoking visual stimuli, is the most effective treatment. Scalable computerized exposure therapy requires automated prediction of fear directly from image content to adapt stimulus selection and treatment intensity. Whether such predictions can be made reliably and generalize across individuals and stimuli, however, remains unknown. Here we show that pretrained convolutional and transformer vision models, adapted via transfer learning, accurately predict group-level perceived fear for spider-related images, even when evaluated on new people and new images, achieving a mean absolute error (MAE) below 10 units on the 0-100 fear scale. Visual explanation analyses indicate that predictions are driven by spider-specific regions in the images. Learning-curve analyses show that transformer models are data efficient and approach performance saturation with the available data (~300 images). Prediction errors increase for very low and very high fear levels and within specific categories of images. These results establish transparent, data-driven fear estimation from images, laying the groundwork for adaptive digital mental health tools.

Method: The approach first mines latent knowledge from the MLLM itself, then injects it into the model via fine-tuning. During inference, conflict signals in predictions activate a self-reflection process that refines the final responses.

Result: DF-LLaVA achieves outstanding detection accuracy exceeding expert models while maintaining the interpretability offered by MLLMs, as confirmed by extensive experiments.

Conclusion: The proposed framework successfully unlocks the intrinsic discrimination potential of MLLMs for synthetic image detection, achieving both high accuracy and explainability.

Abstract: With the increasing prevalence of synthetic images, evaluating image authenticity and locating forgeries accurately while maintaining human interpretability remains a challenging task. Existing detection models primarily focus on simple authenticity classification, ultimately providing only a forgery probability or binary judgment, which offers limited explanatory insights into image authenticity. Moreover, while MLLM-based detection methods can provide more interpretable results, they still lag behind expert models in terms of pure authenticity classification accuracy. To address this, we propose DF-LLaVA, a novel and effective framework that unlocks the intrinsic discrimination potential of MLLMs. Our approach first mines latent knowledge from the MLLM itself and then injects it into the model via fine-tuning. During inference, conflict signals arising from the model’s predictions activate a self-reflection process, leading to the final refined responses. This framework allows LLaVA to achieve outstanding detection accuracy exceeding expert models while still maintaining the interpretability offered by MLLMs. Extensive experiments confirm the superiority of DF-LLaVA, achieving both high accuracy and explainability in synthetic image detection. Code is available online at: https://github.com/Eliot-Shen/DF-LLaVA.

Method: Proposes Core Space merging framework that merges LoRA-adapted models within a common alignment basis, with formal proof that projection into Core Space ensures no information loss.

Result: Significantly improves existing merging techniques, achieves state-of-the-art results on both vision and language tasks while using fraction of computational resources.

Conclusion: Core Space enables efficient merging of parameter-efficient adaptations while preserving low-rank efficiency, advancing model merging techniques for vision and language domains.

Abstract: In this paper, we address the challenges associated with merging low-rank adaptations of large neural networks. With the rise of parameter-efficient adaptation techniques, such as Low-Rank Adaptation (LoRA), model fine-tuning has become more accessible. While fine-tuning models with LoRA is highly efficient, existing merging methods often sacrifice this efficiency by merging fully-sized weight matrices. We propose the Core Space merging framework, which enables the merging of LoRA-adapted models within a common alignment basis, thereby preserving the efficiency of low-rank adaptation while substantially improving accuracy across tasks. We further provide a formal proof that projection into Core Space ensures no loss of information and provide a complexity analysis showing the efficiency gains. Extensive empirical results demonstrate that Core Space significantly improves existing merging techniques and achieves state-of-the-art results on both vision and language tasks while utilizing a fraction of the computational resources. Codebase is available at https://github.com/apanariello4/core-space-merging.

Method: The framework adapts frozen text-to-image diffusion backbones using structural causal modeling with two key strategies: 1) prompt-aligned injection that aligns causal attributes with textual embeddings for semantic control, and 2) conditioned token contrastive loss to disentangle attribute factors and reduce spurious correlations.

Result: Achieves state-of-the-art performance with up to 91% MAE reduction on Pendulum dataset for accurate attribute control and 87% FID reduction on ADNI dataset for high-fidelity MRI image generation, demonstrating robust counterfactual editing with faithful attribute modification and strong identity preservation.

Conclusion: Causal-Adapter enables robust, generalizable counterfactual image editing by explicitly modeling causal relationships, providing precise attribute control while preserving core image identity, outperforming previous prompt-engineering based approaches.

Abstract: We present Causal-Adapter, a modular framework that adapts frozen text-to-image diffusion backbones for counterfactual image generation. Our method enables causal interventions on target attributes, consistently propagating their effects to causal dependents without altering the core identity of the image. In contrast to prior approaches that rely on prompt engineering without explicit causal structure, Causal-Adapter leverages structural causal modeling augmented with two attribute regularization strategies: prompt-aligned injection, which aligns causal attributes with textual embeddings for precise semantic control, and a conditioned token contrastive loss to disentangle attribute factors and reduce spurious correlations. Causal-Adapter achieves state-of-the-art performance on both synthetic and real-world datasets, with up to 91% MAE reduction on Pendulum for accurate attribute control and 87% FID reduction on ADNI for high-fidelity MRI image generation. These results show that our approach enables robust, generalizable counterfactual editing with faithful attribute modification and strong identity preservation.

Method: Architectural enhancements including removal of Distribution Focal Loss, adoption of end-to-end NMS-free inference, integration of ProgLoss and Small-Target-Aware Label Assignment (STAL), and introduction of MuSGD optimizer for stable convergence. Supports multi-task framework for detection, segmentation, pose estimation, oriented detection, and classification.

Result: YOLO26 demonstrates strong performance on edge devices like NVIDIA Jetson Nano and Orin, outperforming previous YOLO versions (v8, v11, v12, v13) and transformer-based detectors (RF-DETR and RT-DETR) in real-time object detection benchmarks.

Conclusion: YOLO26 represents a significant advancement in real-time vision systems for edge deployment, offering improved efficiency, accuracy, and multi-task capabilities with practical applications across various industries.

Abstract: This study presents a comprehensive analysis of Ultralytics YOLO26(also called as YOLOv26), highlighting its key architectural enhancements and performance benchmarking for real-time object detection. YOLO26, released in September 2025, stands as the newest and most advanced member of the YOLO family, purpose-built to deliver efficiency, accuracy, and deployment readiness on edge and low-power devices. The paper sequentially details architectural innovations of YOLO26, including the removal of Distribution Focal Loss (DFL), adoption of end-to-end NMS-free inference, integration of ProgLoss and Small-Target-Aware Label Assignment (STAL), and the introduction of the MuSGD optimizer for stable convergence. Beyond architecture, the study positions YOLO26 as a multi-task framework, supporting object detection, instance segmentation, pose/keypoints estimation, oriented detection, and classification. We present performance benchmarks of YOLO26 on edge devices such as NVIDIA Jetson Nano and Orin, comparing its results with YOLOv8, YOLOv11, YOLOv12, YOLOv13, and transformer-based detectors(RF-DETR and RT-DETR). This paper further explores real-time deployment pathways, flexible export options (ONNX, TensorRT, CoreML, TFLite), and quantization for INT8/FP16. Practical use cases of YOLO26 across robotics, manufacturing, and IoT are highlighted to demonstrate cross-industry adaptability. Finally, insights on deployment efficiency and broader implications are discussed, with future directions for YOLO26 and the YOLO lineage outlined.

Method: Adapts Qwen2.5-VL through end-to-end training on 216K video instruction dataset using hardware-aware hybrid attention: dense attention for text and Native Sparse Attention (NSA) for video.

Result: Outperforms token-compression and training-free sparse baselines on long-video understanding, temporal reasoning, and spatial benchmarks. Achieves reliable scaling to 128K tokens with optimal global-local attention allocation.

Conclusion: VideoNSA effectively addresses context length limitations in video-language models through sparse attention, enabling better long-video understanding and temporal reasoning capabilities.

Abstract: Video understanding in multimodal language models remains limited by context length: models often miss key transition frames and struggle to maintain coherence across long time scales. To address this, we adapt Native Sparse Attention (NSA) to video-language models. Our method, VideoNSA, adapts Qwen2.5-VL through end-to-end training on a 216K video instruction dataset. We employ a hardware-aware hybrid approach to attention, preserving dense attention for text, while employing NSA for video. Compared to token-compression and training-free sparse baselines, VideoNSA achieves improved performance on long-video understanding, temporal reasoning, and spatial benchmarks. Further ablation analysis reveals four key findings: (1) reliable scaling to 128K tokens; (2) an optimal global-local attention allocation at a fixed budget; (3) task-dependent branch usage patterns; and (4) the learnable combined sparse attention help induce dynamic attention sinks. Project Page: https://enxinsong.com/VideoNSA-web/, Code: https://github.com/Espere-1119-Song/VideoNSA

Method: FrameOracle is a plug-and-play module trained via curriculum learning from weak proxy signals (cross-modal similarity) to stronger supervision using FrameOracle-41K, a large-scale VideoQA dataset with validated keyframe annotations specifying minimal sufficient frames per question.

Result: Extensive experiments across five VLMs and six benchmarks show FrameOracle reduces 16-frame inputs to 10.4 frames without accuracy loss, and reduces 64-frame candidates to 13.9 frames while improving accuracy by 1.5%, achieving state-of-the-art efficiency-accuracy trade-offs.

Conclusion: FrameOracle enables scalable video understanding by providing adaptive frame sampling that reduces computational costs while maintaining or improving performance, addressing a key bottleneck in video-language model efficiency.

Abstract: Vision-language models (VLMs) advance video understanding but operate under tight computational budgets, making performance dependent on selecting a small, high-quality subset of frames. Existing frame sampling strategies, such as uniform or fixed-budget selection, fail to adapt to variations in content density or task complexity. To address this, we present FrameOracle, a lightweight, plug-and-play module that predicts both (1) which frames are most relevant to a given query and (2) how many frames are needed. FrameOracle is trained via a curriculum that progresses from weak proxy signals, such as cross-modal similarity, to stronger supervision with FrameOracle-41K, the first large-scale VideoQA dataset with validated keyframe annotations specifying minimal sufficient frames per question. Extensive experiments across five VLMs and six benchmarks show that FrameOracle reduces 16-frame inputs to an average of 10.4 frames without accuracy loss. When starting from 64-frame candidates, it reduces inputs to 13.9 frames on average while improving accuracy by 1.5%, achieving state-of-the-art efficiency-accuracy trade-offs for scalable video understanding.

Method: Proposes Identity-GRPO, a human feedback-driven optimization pipeline: 1) constructs a video reward model trained on large-scale preference data with human-annotated and synthetic distortion data focused on human consistency, 2) employs a GRPO variant tailored for multi-human consistency optimization.

Result: Achieves up to 18.9% improvement in human consistency metrics over baseline methods, with extensive ablation studies showing impact of annotation quality and design choices on policy optimization.

Conclusion: Identity-GRPO effectively enhances multi-human identity preservation in video generation and provides actionable insights for aligning reinforcement learning with personalized video generation.

Abstract: While advanced methods like VACE and Phantom have advanced video generation for specific subjects in diverse scenarios, they struggle with multi-human identity preservation in dynamic interactions, where consistent identities across multiple characters are critical. To address this, we propose Identity-GRPO, a human feedback-driven optimization pipeline for refining multi-human identity-preserving video generation. First, we construct a video reward model trained on a large-scale preference dataset containing human-annotated and synthetic distortion data, with pairwise annotations focused on maintaining human consistency throughout the video. We then employ a GRPO variant tailored for multi-human consistency, which greatly enhances both VACE and Phantom. Through extensive ablation studies, we evaluate the impact of annotation quality and design choices on policy optimization. Experiments show that Identity-GRPO achieves up to 18.9% improvement in human consistency metrics over baseline methods, offering actionable insights for aligning reinforcement learning with personalized video generation.

Method: Proposes LoCoT2V-Bench constructed from real-world videos with multi-scene prompts and hierarchical metadata (character settings, camera behaviors). Also introduces LoCoT2V-Eval framework covering perceptual quality, text-video alignment, temporal quality, dynamic quality, and Human Expectation Realization Degree (HERD).

Result: Evaluation of 13 LVG models reveals strong perceptual quality and background consistency but weak fine-grained text-video alignment and character consistency. Shows pronounced capability disparities across evaluation dimensions.

Conclusion: Improving prompt faithfulness and identity preservation remains a key challenge for long-form video generation. The benchmark provides comprehensive evaluation for advancing long video generation research.

Abstract: Recent advances in text-to-video generation have achieved impressive performance on short clips, yet evaluating long-form generation under complex textual inputs remains a significant challenge. In response to this challenge, we present LoCoT2V-Bench, a benchmark for long video generation (LVG) featuring multi-scene prompts with hierarchical metadata (e.g., character settings and camera behaviors), constructed from collected real-world videos. We further propose LoCoT2V-Eval, a multi-dimensional framework covering perceptual quality, text-video alignment, temporal quality, dynamic quality, and Human Expectation Realization Degree (HERD), with an emphasis on aspects such as fine-grained text-video alignment and temporal character consistency. Experiments on 13 representative LVG models reveal pronounced capability disparities across evaluation dimensions, with strong perceptual quality and background consistency but markedly weaker fine-grained text-video alignment and character consistency. These findings suggest that improving prompt faithfulness and identity preservation remains a key challenge for long-form video generation.

Method: Introduces a negotiator during training to derive common representation from local representations of each modality’s agent. Uses sender-receiver pairs for feature transformation between local and common spaces. Employs structural, pragmatic, and distribution alignment losses for better knowledge distillation.

Result: Effectively reduces inherent domain gaps with various local representations, enabling better alignment and knowledge transfer between heterogeneous agents.

Conclusion: NegoCollab’s negotiated common representation approach successfully addresses heterogeneity challenges in collaborative perception by creating a more equitable representation space that bridges domain gaps between different agents.

Abstract: Collaborative perception improves task performance by expanding the perception range through information sharing among agents. . Immutable heterogeneity poses a significant challenge in collaborative perception, as participating agents may employ different and fixed perception models. This leads to domain gaps in the intermediate features shared among agents, consequently degrading collaborative performance. Aligning the features of all agents to a common representation can eliminate domain gaps with low training cost. However, in existing methods, the common representation is designated as the representation of a specific agent, making it difficult for agents with significant domain discrepancies from this specific agent to achieve proper alignment. This paper proposes NegoCollab, a heterogeneous collaboration method based on the negotiated common representation. It introduces a negotiator during training to derive the common representation from the local representations of each modality’s agent, effectively reducing the inherent domain gap with the various local representations. In NegoCollab, the mutual transformation of features between the local representation space and the common representation space is achieved by a pair of sender and receiver. To better align local representations to the common representation containing multimodal information, we introduce structural alignment loss and pragmatic alignment loss in addition to the distribution alignment loss to supervise the training. This enables the knowledge in the common representation to be fully distilled into the sender.

Method: Conducted systematic study of video-text alignment using modern encoders, proposed parametric test-time scaling laws to capture alignment behavior, investigated correlations between semantic alignment and downstream task performance, and examined temporal reasoning in relation to cross-modal alignment.

Result: Found that cross-modal alignment depends on richness of visual and text data at test time, with scaling laws showing strong predictive power. Strong alignment correlates with better performance on semantic and non-semantic downstream tasks. Temporal reasoning provides challenging testbed for vision-language models.

Conclusion: Video-text alignment serves as an informative zero-shot probe for evaluating representation power of encoders on spatio-temporal data, with implications for understanding general-purpose video representation and understanding.

Abstract: The alignment of representations from different modalities has recently been shown to provide insights on the structural similarities and downstream capabilities of different encoders across diverse data types. While significant progress has been made in aligning images with text, the temporal nature of video data remains largely unexplored in this context. In this work, we conduct the first comprehensive study of video-text representation alignment, probing the capabilities of modern video and language encoders. Our findings reveal several key insights. First, we demonstrate that cross-modal alignment highly depends on the richness of both visual (static images vs. multi-frame videos) and text (single caption vs. a collection) data provided at test time, especially when using state-of-the-art video encoders. We propose parametric test-time scaling laws that capture this behavior and show remarkable predictive power against empirical observations. Secondly, we investigate the correlation between semantic alignment and performance on both semantic and non-semantic downstream tasks, providing initial evidence that strong alignment against text encoders may be linked to general-purpose video representation and understanding. Finally, we correlate temporal reasoning with cross-modal alignment providing a challenging test-bed for vision and language models. Overall, our work introduces video-text alignment as an informative zero-shot way to probe the representation power of different encoders for spatio-temporal data. Project page can be found at https://video-prh.github.io/

Method: Three-module architecture: understanding model, texture module (for appearance synthesis with spatiotemporal modeling), and geometry module (providing explicit geometric constraints). Uses two-stage training strategy and joint modeling of scene understanding, novel view synthesis, and geometry estimation.

Result: Achieves state-of-the-art score of 55.4 on VSI-Bench benchmark, outperforming specialized 3D understanding models, while delivering strong performance in both novel view synthesis and 3D scene generation.

Conclusion: Omni-View successfully demonstrates that joint modeling of 3D scene understanding and generation tasks enables synergistic improvements, validating that generation facilitates understanding in 3D multimodal systems.

Abstract: This paper presents Omni-View, which extends the unified multimodal understanding and generation to 3D scenes based on multiview images, exploring the principle that “generation facilitates understanding”. Consisting of understanding model, texture module, and geometry module, Omni-View jointly models scene understanding, novel view synthesis, and geometry estimation, enabling synergistic interaction between 3D scene understanding and generation tasks. By design, it leverages the spatiotemporal modeling capabilities of its texture module responsible for appearance synthesis, alongside the explicit geometric constraints provided by its dedicated geometry module, thereby enriching the model’s holistic understanding of 3D scenes. Trained with a two-stage strategy, Omni-View achieves a state-of-the-art score of 55.4 on the VSI-Bench benchmark, outperforming existing specialized 3D understanding models, while simultaneously delivering strong performance in both novel view synthesis and 3D scene generation. The code and pretraiend models are open-sourced at https://github.com/AIDC-AI/Omni-View.

Method: MACEval employs a multi-agent network with role assignment, in-process data generation, and evaluation routing through cascaded agents. It uses interactive and autonomous evaluation modes with new longitudinal performance metrics.

Result: Extensive experiments on 23 large models demonstrate MACEval’s effectiveness in evaluating models dynamically while reducing evaluation overhead and lightening the evaluation process.

Conclusion: MACEval provides a novel approach to large model evaluation that addresses limitations of traditional benchmarks and could broaden future directions in model assessment.

Abstract: Hundreds of benchmarks dedicated to evaluating large models have been presented over the past few years. However, most of them remain closed-ended and are prone to overfitting due to the potential data contamination. Moreover, the increasing scale and scope of current benchmarks with transient metrics, as well as the heavily human-dependent curation procedure, pose significant challenges for timely maintenance and adaptation. In this paper, we introduce MACEval, a Multi-Agent Continual Evaluation network for dynamic evaluation of large models, and define new metrics to quantify performance longitudinally. MACEval employs an interactive and autonomous evaluation mode, utilizing role assignment, in-process data generation, and evaluation routing through a cascaded agent network. Extensive experiments on 23 large models demonstrate the effectiveness of MACEval, which also lightens the evaluation process and reduces a considerable amount of overhead. We hope that MACEval can broaden future directions of large model evaluation. Project page: https://github.com/zijianchen98/MACEval.

Method: Proposes asymmetric aggregation with geometric constraints (A²GC-VPR). Uses row-column normalization averaging with separate marginal calibration for asymmetric matching. Incorporates geometric constraints through learnable coordinate embeddings that compute compatibility scores fused with feature similarities, promoting spatially proximal features to cluster together.

Result: Superior performance demonstrated on MSLS, NordLand, and Pittsburgh datasets. Validates effectiveness in improving matching accuracy and robustness compared to state-of-the-art methods.

Conclusion: The proposed asymmetric aggregation method with geometric constraints effectively addresses distributional discrepancies in VPR, leading to improved performance and robustness in visual place recognition tasks.

Abstract: Visual Place Recognition (VPR) aims to match query images against a database using visual cues. State-of-the-art methods aggregate features from deep backbones to form global descriptors. Optimal transport-based aggregation methods reformulate feature-to-cluster assignment as a transport problem, but the standard Sinkhorn algorithm symmetrically treats source and target marginals, limiting effectiveness when image features and cluster centers exhibit substantially different distributions. We propose an asymmetric aggregation VPR method with geometric constraints for locally aggregated descriptors, called $A^2$GC-VPR. Our method employs row-column normalization averaging with separate marginal calibration, enabling asymmetric matching that adapts to distributional discrepancies in visual place recognition. Geometric constraints are incorporated through learnable coordinate embeddings, computing compatibility scores fused with feature similarities, thereby promoting spatially proximal features to the same cluster and enhancing spatial awareness. Experimental results on MSLS, NordLand, and Pittsburgh datasets demonstrate superior performance, validating the effectiveness of our approach in improving matching accuracy and robustness.

Method: The framework integrates atlas-based regional parcellation, centerline extraction, and graph construction to compute fifteen morphometric, topological, fractal, and geometric features from 3D TOF-MRA scans, enabling both global and regional analysis of vascular networks.

Result: Applied to 570 scans from the IXI dataset, CaravelMetrics produced reproducible vessel graphs that captured age- and sex-related variations and education-associated increases in vascular complexity, consistent with existing literature findings.

Conclusion: CaravelMetrics provides a scalable, automated solution for quantitative cerebrovascular analysis that can support large-scale studies of vascular health, aging, and potentially clinical applications in cerebrovascular disease assessment.

Abstract: We present CaravelMetrics, a computational framework for automated cerebrovascular analysis that models vessel morphology through skeletonization-derived graph representations. The framework integrates atlas-based regional parcellation, centerline extraction, and graph construction to compute fifteen morphometric, topological, fractal, and geometric features. The features can be estimated globally from the complete vascular network or regionally within arterial territories, enabling multiscale characterization of cerebrovascular organization. Applied to 570 3D TOF-MRA scans from the IXI dataset (ages 20-86), CaravelMetrics yields reproducible vessel graphs capturing age- and sex-related variations and education-associated increases in vascular complexity, consistent with findings reported in the literature. The framework provides a scalable and fully automated approach for quantitative cerebrovascular feature extraction, supporting normative modeling and population-level studies of vascular health and aging.

Method: Proposes VAT (Vision Action Transformer), which extends ViT architecture to process specialized action tokens with visual features across all transformer layers. This enables deep, progressive fusion of perception and action generation by utilizing the full “representation trajectory” of vision models.

Result: Achieves 98.15% average success rate across four LIBERO benchmarks for simulated manipulation tasks, establishing new state-of-the-art and outperforming prior methods like OpenVLA-OFT.

Conclusion: VAT demonstrates the critical importance of leveraging complete vision model representations for advancing robotic policy learning, presenting a powerful model for imitation learning that unlocks the full potential of Vision Transformers for robotics.

Abstract: In robot learning, Vision Transformers (ViTs) are standard for visual perception, yet most methods discard valuable information by using only the final layer’s features. We argue this provides an insufficient representation and propose the Vision Action Transformer (VAT), a novel architecture that is extended from ViT and unlocks the full feature hierarchy of ViT. VAT processes specialized action tokens with visual features across all transformer layers, enabling a deep and progressive fusion of perception and action generation. On a suite of simulated manipulation tasks, VAT achieves a 98.15% average success rate across four LIBERO benchmarks, establishing a new state-of-the-art by outperforming prior methods like OpenVLA-OFT. Our work presents not only a powerful model for imitation learning but also demonstrates the critical importance of leveraging the complete ‘‘representation trajectory’’ of vision models to advance robotic policy. The GitHub URL for the project code is https://github.com/sellerbubble/VAT.

Method: Empirical analysis reveals task-model misalignment: VLMs are good at semantic reasoning but poor at pixel artifacts, while vision models are the opposite. Proposes Task-Model Alignment principle with AlignGemini - a two-branch detector combining VLM (semantic supervision) and vision model (pixel-artifact supervision).

Result: AlignGemini improves average accuracy by 9.5% on in-the-wild benchmarks using simplified training data. Shows that clear specialization of each branch captures complementary cues for better detection.

Conclusion: Task-model alignment is an effective principle for generalizable AI-generated image detection. Different models are naturally suited to different subtasks (semantic consistency vs pixel artifacts), and combining them with clear specialization yields superior performance.

Abstract: Vision Language Models (VLMs) are increasingly used for detecting AI-generated images (AIGI). However, converting VLMs into reliable detectors is resource-intensive, and the resulting models often suffer from hallucination and poor generalization. To investigate the root cause, we conduct an empirical analysis and identify two consistent behaviors. First, fine-tuning VLMs with semantic supervision improves semantic discrimination and generalizes well to unseen data. Second, fine-tuning VLMs with pixel-artifact supervision leads to weak generalization. These findings reveal a fundamental task-model misalignment. VLMs are optimized for high-level semantic reasoning and lack inductive bias toward low-level pixel artifacts. In contrast, conventional vision models effectively capture pixel-level artifacts but are less sensitive to semantic inconsistencies. This indicates that different models are naturally suited to different subtasks. Based on this insight, we formulate AIGI detection as two orthogonal subtasks: semantic consistency checking and pixel-artifact detection. Neglecting either subtask leads to systematic detection failures. We further propose the Task-Model Alignment principle and instantiate it in a two-branch detector, AlignGemini. The detector combines a VLM trained with pure semantic supervision and a vision model trained with pure pixel-artifact supervision. By enforcing clear specialization, each branch captures complementary cues. Experiments on in-the-wild benchmarks show that AlignGemini improves average accuracy by 9.5 percent using simplified training data. These results demonstrate that task-model alignment is an effective principle for generalizable AIGI detection.

Method: MVP treats camera exposure triangle (ISO, shutter speed, aperture) as physical prompts. At inference: 1) acquires library of physical views per scene, 2) selects top-k sensor settings using source-affinity score, 3) evaluates each view under lightweight digital augmentations, 4) filters lowest-entropy subset, 5) aggregates predictions with hard voting. No gradients or model modifications needed.

Result: On ImageNet-ES and ImageNet-ES-Diverse, MVP outperforms digital-only TTA on single Auto-Exposure captures by up to 25.6 percentage points, with additional 3.4 pp gains over conventional sensor control + TTA pipelines. Effective even with reduced parameter sets for lower latency.

Conclusion: Measurement-time control through physical view selection and combination substantially improves VLM robustness in sensor-mediated environments, going beyond post-capture digital prompting.

Abstract: To extend the application of vision-language models (VLMs) from web images to sensor-mediated physical environments, we propose Multi-View Physical-prompt for Test-Time Adaptation (MVP), a forward-only framework that moves test-time adaptation (TTA) from tokens to photons by treating the camera exposure triangle–ISO, shutter speed, and aperture–as physical prompts. At inference, MVP acquires a library of physical views per scene, selects the top-k sensor settings using a source-affinity score, evaluates each retained view under lightweight digital augmentations, filters the lowest-entropy subset of augmented views, and aggregates predictions with Zero-temperature softmax (i.e., hard voting). This selection-then-vote design is simple, calibration-friendly, and requires no gradients or model modifications. On ImageNet-ES and ImageNet-ES-Diverse, MVP consistently outperforms digital-only TTA on single Auto-Exposure captures, by up to 25.6 percentage points (pp), and delivers up to 3.4 pp additional gains over pipelines that combine conventional sensor control with TTA. MVP remains effective under reduced parameter candidate sets that lower capture latency, demonstrating practicality. These results support the main claim that, beyond post-capture prompting, measurement-time control–selecting and combining real physical views–substantially improves robustness for VLMs.

Method: Modular, loosely coupled multi-expert architecture with adaptive GPU load balancing, distributed inference, dynamic module orchestration, and configurable parsing modes supporting holistic or modality-specific parsing.

Result: Achieves processing rate of up to 20 PDF pages per second on 8 x NVIDIA RTX 4090D GPUs, enabling cost-efficient inference across billions of pages for large-scale applications.

Conclusion: Uni-Parser provides scalable, industrial-grade document parsing that facilitates downstream applications like literature retrieval, chemical structure extraction, and corpus curation for training next-generation AI models.

Abstract: This technical report introduces Uni-Parser, an industrial-grade document parsing engine tailored for scientific literature and patents, delivering high throughput, robust accuracy, and cost efficiency. Unlike pipeline-based document parsing methods, Uni-Parser employs a modular, loosely coupled multi-expert architecture that preserves fine-grained cross-modal alignments across text, equations, tables, figures, and chemical structures, while remaining easily extensible to emerging modalities. The system incorporates adaptive GPU load balancing, distributed inference, dynamic module orchestration, and configurable modes that support either holistic or modality-specific parsing. Optimized for large-scale cloud deployment, Uni-Parser achieves a processing rate of up to 20 PDF pages per second on 8 x NVIDIA RTX 4090D GPUs, enabling cost-efficient inference across billions of pages. This level of scalability facilitates a broad spectrum of downstream applications, ranging from literature retrieval and summarization to the extraction of chemical structures, reaction schemes, and bioactivity data, as well as the curation of large-scale corpora for training next-generation large language models and AI4Science models.

Method: 1) ViT depth network with metric-scale supervision for globally consistent depths; 2) Multi-scale feature bundle-adjustment layer refines poses in feature space using learned pyramidal descriptors; 3) Incremental local-radiance-field hierarchy with hash-grid NeRFs allocated on-the-fly when view overlap falls below threshold.

Result: Achieves Absolute Trajectory Error of 0.001-0.021 m on Tanks and Temples benchmark (18x lower than BARF, 2x lower than NoPe-NeRF) with sub-pixel Relative Pose Error, enabling city-block-scale coverage on single GPU.

Conclusion: Metric-scale, drift-free 3D reconstruction and high-fidelity novel-view synthesis are achievable from single uncalibrated RGB camera through joint optimization of depth, pose, and radiance factors.

Abstract: Photorealistic 3-D reconstruction from monocular video collapses in large-scale scenes when depth, pose, and radiance are solved in isolation: scale-ambiguous depth yields ghost geometry, long-horizon pose drift corrupts alignment, and a single global NeRF cannot model hundreds of metres of content. We introduce a joint learning framework that couples all three factors and demonstrably overcomes each failure case. Our system begins with a Vision-Transformer (ViT) depth network trained with metric-scale supervision, giving globally consistent depths despite wide field-of-view variations. A multi-scale feature bundle-adjustment (BA) layer refines camera poses directly in feature space–leveraging learned pyramidal descriptors instead of brittle keypoints–to suppress drift on unconstrained trajectories. For scene representation, we deploy an incremental local-radiance-field hierarchy: new hash-grid NeRFs are allocated and frozen on-the-fly when view overlap falls below a threshold, enabling city-block-scale coverage on a single GPU. Evaluated on the Tanks and Temples benchmark, our method reduces Absolute Trajectory Error to 0.001-0.021 m across eight indoor-outdoor sequences–up to 18x lower than BARF and 2x lower than NoPe-NeRF–while maintaining sub-pixel Relative Pose Error. These results demonstrate that metric-scale, drift-free 3-D reconstruction and high-fidelity novel-view synthesis are achievable from a single uncalibrated RGB camera.

Method: Two-stage adaptive design: 1) Lightweight temporal extractor ranks data windows by saliency, 2) 4D hierarchical encoder (Hiera-JEPA) learns voxel-level representations only from top-k selected windows with ~70% patch masking.

Result: Achieves SOTA on seven public benchmarks while requiring only 4k pre-training sessions and ~30% GPU memory compared to traditional voxel-level methods.

Conclusion: SLIM-Brain successfully addresses data- and training-efficiency bottlenecks in fMRI foundation models, enabling effective atlas-free modeling with practical resource requirements.

Abstract: Foundation models are emerging as a powerful paradigm for fMRI analysis, but current approaches face a dual bottleneck of data- and training-efficiency. Atlas-based methods aggregate voxel signals into fixed regions of interest, reducing data dimensionality but discarding fine-grained spatial details, and requiring extremely large cohorts to train effectively as general-purpose foundation models. Atlas-free methods, on the other hand, operate directly on voxel-level information - preserving spatial fidelity but are prohibitively memory- and compute-intensive, making large-scale pre-training infeasible. We introduce SLIM-Brain (Sample-efficient, Low-memory fMRI Foundation Model for Human Brain), a new atlas-free foundation model that simultaneously improves both data- and training-efficiency. SLIM-Brain adopts a two-stage adaptive design: (i) a lightweight temporal extractor captures global context across full sequences and ranks data windows by saliency, and (ii) a 4D hierarchical encoder (Hiera-JEPA) learns fine-grained voxel-level representations only from the top-$k$ selected windows, while deleting about 70% masked patches. Extensive experiments across seven public benchmarks show that SLIM-Brain establishes new state-of-the-art performance on diverse tasks, while requiring only 4 thousand pre-training sessions and approximately 30% of GPU memory comparing to traditional voxel-level methods.

Method: Proposes DepthDirector with View-Content Dual-Stream Condition mechanism that injects both source video and warped depth sequence from target viewpoint into pretrained video generation model. Uses lightweight LoRA-based video diffusion adapter for training. Also creates MultiCam-WarpData dataset with 8K videos across 1K dynamic scenes using Unreal Engine 5.

Result: Outperforms existing methods in both camera controllability and visual quality. The framework enables faithful reproduction of dynamic scenes under novel camera trajectories while maintaining content consistency.

Conclusion: DepthDirector successfully addresses the camera control challenge by leveraging explicit 3D depth guidance, enabling precise camera trajectory alteration while preserving video content through better utilization of video diffusion models’ 3D understanding capabilities.

Abstract: Camera control has been extensively studied in conditioned video generation; however, performing precisely altering the camera trajectories while faithfully preserving the video content remains a challenging task. The mainstream approach to achieving precise camera control is warping a 3D representation according to the target trajectory. However, such methods fail to fully leverage the 3D priors of video diffusion models (VDMs) and often fall into the Inpainting Trap, resulting in subject inconsistency and degraded generation quality. To address this problem, we propose DepthDirector, a video re-rendering framework with precise camera controllability. By leveraging the depth video from explicit 3D representation as camera-control guidance, our method can faithfully reproduce the dynamic scene of an input video under novel camera trajectories. Specifically, we design a View-Content Dual-Stream Condition mechanism that injects both the source video and the warped depth sequence rendered under the target viewpoint into the pretrained video generation model. This geometric guidance signal enables VDMs to comprehend camera movements and leverage their 3D understanding capabilities, thereby facilitating precise camera control and consistent content generation. Next, we introduce a lightweight LoRA-based video diffusion adapter to train our framework, fully preserving the knowledge priors of VDMs. Additionally, we construct a large-scale multi-camera synchronized dataset named MultiCam-WarpData using Unreal Engine 5, containing 8K videos across 1K dynamic scenes. Extensive experiments show that DepthDirector outperforms existing methods in both camera controllability and visual quality. Our code and dataset will be publicly available.

Method: Developed systematic pipeline with data source selection, text refinement, visual diversification, and cross-modal cross-batch filtering to curate 100M high-quality Chinese image-text pairs from Common Crawl, including contemporary 2024-2025 data.

Result: DanQing consistently outperforms existing Chinese datasets across zero-shot classification, cross-modal retrieval, and Chinese LMM tasks when used for continued pretraining of SigLIP2 models, showing balanced semantic distribution and superior scaling.

Conclusion: DanQing addresses the critical data bottleneck for Chinese VLP, providing high-quality contemporary data that enables better multimodal understanding and will be open-sourced to advance Chinese vision-language research.

Abstract: Vision-Language Pre-training (VLP) models have achieved remarkable success by leveraging large-scale image-text pairs. While English-centric models like CLIP and SigLIP benefit from massive datasets (e.g., LAION-400M), the development of Chinese VLP remains bottlenecked by the lack of high-quality, large-scale open-source data. In this paper, we present DanQing, a large-scale Chinese cross-modal dataset containing 100 million high-quality image-text pairs curated from Common Crawl. To ensure superior data quality, we develop an effective systematic pipeline comprising data source selection, text refinement, visual diversification, and cross-modal cross-batch filtering, thereby effectively mitigating the intrinsic noise prevalent in web data. Notably, DanQing incorporates data from 2024-2025, enabling models to capture contemporary semantic trends and emerging concepts. Extensive experiments via continued pretraining of SigLIP2 models demonstrate that DanQing consistently outperforms existing Chinese datasets across diverse downstream tasks, including zero-shot classification, cross-modal retrieval, and Chinese-centric large multimodal model tasks. Furthermore, in-depth analysis of DanQing reveals that it exhibits a more balanced semantic distribution and superior scaling capability compared to existing datasets. To facilitate further research in Chinese vision-language pre-training, we will open-source the DanQing dataset under the Creative Common CC-BY 4.0 license.

Method: Three-stage framework: 1) Adaptive visual sampling to select candidate frames, 2) Label-driven keyframe anchoring using semantic labels to identify meaningful transitions, and 3) Contextual validation using a large language model to filter out generic content and ensure coherence.

Result: Despite sampling fewer than 5% of original frames, PRISM retains 84% semantic content and improves over baselines by up to 33%. The method generalizes well across procedural and domain-specific video tasks, achieving strong performance in both semantic alignment and precision.

Conclusion: PRISM effectively produces contextually coherent video summaries by integrating multimodal analysis with LLM-based validation, demonstrating strong performance in retaining semantic content while using minimal frames across various video domains.

Abstract: Video summarization helps turn long videos into clear, concise representations that are easier to review, document, and analyze, especially in high-stakes domains like surgical training. Prior work has progressed from using basic visual features like color, motion, and structural changes to using pre-trained vision-language models that can better understand what’s happening in the video (semantics) and capture temporal flow, resulting in more context-aware video summarization. We propose a three-stage framework, PRISM: Procedural Representation via Integrated Semantic and Multimodal analysis, that produces semantically grounded video summaries. PRISM combines adaptive visual sampling, label-driven keyframe anchoring, and contextual validation using a large language model (LLM). Our method ensures that selected frames reflect meaningful and procedural transitions while filtering out generic or hallucinated content, resulting in contextually coherent summaries across both domain-specific and instructional videos. We evaluate our method on instructional and activity datasets, using reference summaries for instructional videos. Despite sampling fewer than 5% of the original frames, our summaries retain 84% semantic content while improving over baselines by as much as 33%. Our approach generalizes across procedural and domain-specific video tasks, achieving strong performance with both semantic alignment and precision.

Method: Created ObjectVisA-120 dataset with 120 participants in VR street-crossing scenarios, featuring gaze data, object state-space, panoptic segmentation, depth, and vehicle keypoints. Proposed object-based similarity (oSIM) metric and developed SUMGraph model using Mamba U-Net with graph representation of critical scene objects.

Result: Explicit optimization for object-based attention improves oSIM performance and enhances model performance on common metrics. SUMGraph outperforms state-of-the-art visual attention prediction methods.

Conclusion: ObjectVisA-120 enables proper evaluation of object-based visual attention models, and explicitly modeling object-based attention leads to improved performance, advancing computational visual attention research.

Abstract: The object-based nature of human visual attention is well-known in cognitive science, but has only played a minor role in computational visual attention models so far. This is mainly due to a lack of suitable datasets and evaluation metrics for object-based attention. To address these limitations, we present ObjectVisA-120 – a novel 120-participant dataset of spatial street-crossing navigation in virtual reality specifically geared to object-based attention evaluations. The uniqueness of the presented dataset lies in the ethical and safety affiliated challenges that make collecting comparable data in real-world environments highly difficult. ObjectVisA-120 not only features accurate gaze data and a complete state-space representation of objects in the virtual environment, but it also offers variable scenario complexities and rich annotations, including panoptic segmentation, depth information, and vehicle keypoints. We further propose object-based similarity (oSIM) as a novel metric to evaluate the performance of object-based visual attention models, a previously unexplored performance characteristic. Our evaluations show that explicitly optimising for object-based attention not only improves oSIM performance but also leads to an improved model performance on common metrics. In addition, we present SUMGraph, a Mamba U-Net-based model, which explicitly encodes critical scene objects (vehicles) in a graph representation, leading to further performance improvements over several state-of-the-art visual attention prediction methods. The dataset, code and models will be publicly released.

Method: 1) Learns per-pixel Gaussian representation from few input images, 2) Uses transformer-based encoder to fuse DINOv3 and Stable Diffusion VAE features, 3) Extends Gaussian representations with per-Gaussian features, 4) Employs lightweight MLP-based dynamic network for real-time animation from expression codes, 5) Uses point maps from pre-trained reconstruction model for geometry supervision.

Result: Significantly outperforms existing methods in both rendering quality and inference efficiency, while supporting real-time dynamic avatar animation from minimal input images.

Conclusion: FastGHA enables efficient, high-quality head avatar generation from few images with real-time animation capabilities, addressing scalability and usability limitations of current methods.

Abstract: Despite recent progress in 3D Gaussian-based head avatar modeling, efficiently generating high fidelity avatars remains a challenge. Current methods typically rely on extensive multi-view capture setups or monocular videos with per-identity optimization during inference, limiting their scalability and ease of use on unseen subjects. To overcome these efficiency drawbacks, we propose FastGHA, a feed-forward method to generate high-quality Gaussian head avatars from only a few input images while supporting real-time animation. Our approach directly learns a per-pixel Gaussian representation from the input images, and aggregates multi-view information using a transformer-based encoder that fuses image features from both DINOv3 and Stable Diffusion VAE. For real-time animation, we extend the explicit Gaussian representations with per-Gaussian features and introduce a lightweight MLP-based dynamic network to predict 3D Gaussian deformations from expression codes. Furthermore, to enhance geometric smoothness of the 3D head, we employ point maps from a pre-trained large reconstruction model as geometry supervision. Experiments show that our approach significantly outperforms existing methods in both rendering quality and inference efficiency, while supporting real-time dynamic avatar animation.

Method: FMIR combines a foundation model-based feature encoder for anatomical structure extraction with a general registration head, trained using channel regularization strategy on just a single dataset to enhance generalization.

Result: FMIR achieves state-of-the-art in-domain performance while maintaining robust registration on out-of-domain images, demonstrating effective generalization with limited training resources.

Conclusion: The approach shows a viable path toward building generalizable medical imaging foundation models with limited resources, addressing a key limitation in clinical deployment of deep learning registration methods.

Abstract: Deep learning has revolutionized medical image registration by achieving unprecedented speeds, yet its clinical application is hindered by a limited ability to generalize beyond the training domain, a critical weakness given the typically small scale of medical datasets. In this paper, we introduce FMIR, a foundation model-based registration framework that overcomes this limitation.Combining a foundation model-based feature encoder for extracting anatomical structures with a general registration head, and trained with a channel regularization strategy on just a single dataset, FMIR achieves state-of-the-art(SOTA) in-domain performance while maintaining robust registration on out-of-domain images.Our approach demonstrates a viable path toward building generalizable medical imaging foundation models with limited resources. The code is available at https://github.com/Monday0328/FMIR.git.

Method: Proposes a two-phase approach: 1) Cold-start using foundation-model embeddings with clustering (including automatic cluster number selection and proportional sampling) to create diverse initial training set; 2) Uncertainty-based active learning with spatial diversity integration for subsequent sample selection. The method is designed to be interpretable with feature-space visualization.

Result: Evaluated on three medical imaging datasets (CheXmask, Montgomery, SynthStrip). Cold-start outperformed random selection: CheXmask Dice improved from 0.918 to 0.929, Hausdorff distance reduced from 32.41 to 27.66 mm. Combined entropy+diversity AL improved CheXmask Dice from 0.919 to 0.939, Hausdorff from 30.10 to 19.16 mm. Similar improvements on other datasets.

Conclusion: The proposed framework consistently outperforms baselines in low-data regimes, improving segmentation accuracy while reducing annotation burden through more effective sample selection strategies.

Abstract: Accurate segmentation annotations are critical for disease monitoring, yet manual labeling remains a major bottleneck due to the time and expertise required. Active learning (AL) alleviates this burden by prioritizing informative samples for annotation, typically through a diversity-based cold-start phase followed by uncertainty-driven selection. We propose a novel cold-start sampling strategy that combines foundation-model embeddings with clustering, including automatic selection of the number of clusters and proportional sampling across clusters, to construct a diverse and representative initial training. This is followed by an uncertainty-based AL framework that integrates spatial diversity to guide sample selection. The proposed method is intuitive and interpretable, enabling visualization of the feature-space distribution of candidate samples. We evaluate our approach on three datasets spanning X-ray and MRI modalities. On the CheXmask dataset, the cold-start strategy outperforms random selection, improving Dice from 0.918 to 0.929 and reducing the Hausdorff distance from 32.41 to 27.66 mm. In the AL setting, combined entropy and diversity selection improves Dice from 0.919 to 0.939 and reduces the Hausdorff distance from 30.10 to 19.16 mm. On the Montgomery dataset, cold-start gains are substantial, with Dice improving from 0.928 to 0.950 and Hausdorff distance decreasing from 14.22 to 9.38 mm. On the SynthStrip dataset, cold-start selection slightly affects Dice but reduces the Hausdorff distance from 9.43 to 8.69 mm, while active learning improves Dice from 0.816 to 0.826 and reduces the Hausdorff distance from 7.76 to 6.38 mm. Overall, the proposed framework consistently outperforms baseline methods in low-data regimes, improving segmentation accuracy.

Method: Hybrid workflow combining deep learning-based image modality classification with figure caption analysis to curate and categorize images from PubMed Central repository using specific keywords.

Result: Retrieved 7,772 images across 166 diagnoses; hybrid approach achieved 90.4% F-score; dataset validated by board-certified dermatopathologists; found current OpenAI image analysis inadequate for dermatopathology.

Conclusion: Developed DermpathNet, a large peer-reviewed open-access dermatopathology dataset with semi-automated curation workflow for educational and ML purposes.

Abstract: Accessing high-quality, open-access dermatopathology image datasets for learning and cross-referencing is a common challenge for clinicians and dermatopathology trainees. To establish a comprehensive open-access dermatopathology dataset for educational, cross-referencing, and machine-learning purposes, we employed a hybrid workflow to curate and categorize images from the PubMed Central (PMC) repository. We used specific keywords to extract relevant images, and classified them using a novel hybrid method that combined deep learning-based image modality classification with figure caption analyses. Validation on 651 manually annotated images demonstrated the robustness of our workflow, with an F-score of 89.6% for the deep learning approach, 61.0% for the keyword-based retrieval method, and 90.4% for the hybrid approach. We retrieved over 7,772 images across 166 diagnoses and released this fully annotated dataset, reviewed by board-certified dermatopathologists. Using our dataset as a challenging task, we found the current image analysis algorithm from OpenAI inadequate for analyzing dermatopathology images. In conclusion, we have developed a large, peer-reviewed, open-access dermatopathology image dataset, DermpathNet, which features a semi-automated curation workflow.

Method: Proposes Entropy-Guided k-Guard (ENkG) sampling that adapts token candidate sizes based on token-wise dispersion measured by entropy. Low-entropy regions use fewer candidates to suppress noise, high-entropy regions use more candidates to mitigate error compounding.

Result: Experiments show consistent improvements in perceptual quality and structural stability compared to static top-k/top-p strategies. The method is model-agnostic, training-free, and adds negligible overhead.

Conclusion: ENkG sampling effectively addresses the mismatch between LLM sampling strategies and video generation needs by adapting to token uncertainty, improving long-horizon video generation quality.

Abstract: Autoregressive (AR) architectures have achieved significant successes in LLMs, inspiring explorations for video generation. In LLMs, top-p/top-k sampling strategies work exceptionally well: language tokens have high semantic density and low redundancy, so a fixed size of token candidates already strikes a balance between semantic accuracy and generation diversity. In contrast, video tokens have low semantic density and high spatio-temporal redundancy. This mismatch makes static top-k/top-p strategies ineffective for video decoders: they either introduce unnecessary randomness for low-uncertainty regions (static backgrounds) or get stuck in early errors for high-uncertainty regions (foreground objects). Prediction errors will accumulate as more frames are generated and eventually severely degrade long-horizon quality. To address this, we propose Entropy-Guided k-Guard (ENkG) sampling, a simple yet effective strategy that adapts sampling to token-wise dispersion, quantified by the entropy of each token’s predicted distribution. ENkG uses adaptive token candidate sizes: for low-entropy regions, it employs fewer candidates to suppress redundant noise and preserve structural integrity; for high-entropy regions, it uses more candidates to mitigate error compounding. ENkG is model-agnostic, training-free, and adds negligible overhead. Experiments demonstrate consistent improvements in perceptual quality and structural stability compared to static top-k/top-p strategies.

Method: Proposes MARE with three key components: 1) Comprehensive reward functions using RLHF to generate text-spatially aligned reasoning content, 2) Multimodal alignment between vision and language for explainable detection, and 3) Forgery disentanglement module to separate intrinsic forgery traces from high-level facial semantics.

Result: MARE achieves state-of-the-art performance in both quantitative and qualitative evaluations. It demonstrates superior accuracy and reliability in deepfake detection while generating explainable reasoning content that aligns with human preferences.

Conclusion: MARE successfully addresses the need for explainable deepfake detection by leveraging vision-language models with multimodal alignment and reinforcement learning, providing both accurate detection and human-interpretable reasoning.

Abstract: Deepfake detection is a widely researched topic that is crucial for combating the spread of malicious content, with existing methods mainly modeling the problem as classification or spatial localization. The rapid advancements in generative models impose new demands on Deepfake detection. In this paper, we propose multimodal alignment and reinforcement for explainable Deepfake detection via vision-language models, termed MARE, which aims to enhance the accuracy and reliability of Vision-Language Models (VLMs) in Deepfake detection and reasoning. Specifically, MARE designs comprehensive reward functions, incorporating reinforcement learning from human feedback (RLHF), to incentivize the generation of text-spatially aligned reasoning content that adheres to human preferences. Besides, MARE introduces a forgery disentanglement module to capture intrinsic forgery traces from high-level facial semantics, thereby improving its authenticity detection capability. We conduct thorough evaluations on the reasoning content generated by MARE. Both quantitative and qualitative experimental results demonstrate that MARE achieves state-of-the-art performance in terms of accuracy and reliability.

Method: FunHSI performs functionality-aware contact reasoning to identify functional scene elements, reconstruct their 3D geometry, and model interactions via a contact graph. It uses vision-language models to synthesize humans performing tasks in images, estimates 3D body/hand poses, and refines configurations through stage-wise optimization for physical plausibility.

Result: FunHSI generates functionally correct and physically plausible human-scene interactions across diverse indoor and outdoor scenes, supporting both general interactions (like “sitting on a sofa”) and fine-grained functional interactions (like “increasing the room temperature”).

Conclusion: FunHSI provides a training-free, functionality-driven framework that enables functionally correct human-scene interactions from open-vocabulary prompts, addressing limitations of existing methods through explicit reasoning about object functionality and human-scene contact.

Abstract: Generating 3D humans that functionally interact with 3D scenes remains an open problem with applications in embodied AI, robotics, and interactive content creation. The key challenge involves reasoning about both the semantics of functional elements in 3D scenes and the 3D human poses required to achieve functionality-aware interaction. Unfortunately, existing methods typically lack explicit reasoning over object functionality and the corresponding human-scene contact, resulting in implausible or functionally incorrect interactions. In this work, we propose FunHSI, a training-free, functionality-driven framework that enables functionally correct human-scene interactions from open-vocabulary task prompts. Given a task prompt, FunHSI performs functionality-aware contact reasoning to identify functional scene elements, reconstruct their 3D geometry, and model high-level interactions via a contact graph. We then leverage vision-language models to synthesize a human performing the task in the image and estimate proposed 3D body and hand poses. Finally, the proposed 3D body configuration is refined via stage-wise optimization to ensure physical plausibility and functional correctness. In contrast to existing methods, FunHSI not only synthesizes more plausible general 3D interactions, such as “sitting on a sofa’’, while supporting fine-grained functional human-scene interactions, e.g., “increasing the room temperature’’. Extensive experiments demonstrate that FunHSI consistently generates functionally correct and physically plausible human-scene interactions across diverse indoor and outdoor scenes.

Method: Proposes a multimodal classification and matching framework that fuses video footage, antenna geometric features, and PCI signals. Introduces a dedicated training framework with Token Entropy Regularization to address cross-modal alignment challenges in pretraining.

Result: Experiments show that Token Entropy Regularization accelerates convergence and yields significant performance gains. Analysis reveals that the entropy of the first token is modality-dependent.

Conclusion: The proposed multimodal approach with specialized training techniques effectively addresses antenna affiliation identification, offering a practical solution for communication network optimization and maintenance.

Abstract: Accurate antenna affiliation identification is crucial for optimizing and maintaining communication networks. Current practice, however, relies on the cumbersome and error-prone process of manual tower inspections. We propose a novel paradigm shift that fuses video footage of base stations, antenna geometric features, and Physical Cell Identity (PCI) signals, transforming antenna affiliation identification into multi-modal classification and matching tasks. Publicly available pretrained transformers struggle with this unique task due to a lack of analogous data in the communications domain, which hampers cross-modal alignment. To address this, we introduce a dedicated training framework that aligns antenna images with corresponding PCI signals. To tackle the representation alignment challenge, we propose a novel Token Entropy Regularization module in the pretraining stage. Our experiments demonstrate that TER accelerates convergence and yields significant performance gains. Further analysis reveals that the entropy of the first token is modality-dependent. Code will be made available upon publication.

Method: Proposes UniMRG (Unified Multi-Representation Generation), an architecture-agnostic post-training method that trains UMMs to generate multiple intrinsic representations of input images: pixel (reconstruction), depth (geometry), and segmentation (structure), alongside standard visual understanding objectives.

Result: Extensive experiments show the method notably enhances fine-grained perception, reduces hallucinations, improves spatial understanding, and simultaneously boosts generation capabilities across diverse UMM architectures.

Conclusion: UniMRG successfully demonstrates that incorporating auxiliary generation tasks for multiple visual representations can significantly enhance the understanding capabilities of multimodal models, creating a mutually reinforcing cycle between understanding and generation.

Abstract: Unified Multimodal Models (UMMs) integrate both visual understanding and generation within a single framework. Their ultimate aspiration is to create a cycle where understanding and generation mutually reinforce each other. While recent post-training methods have successfully leveraged understanding to enhance generation, the reverse direction of utilizing generation to improve understanding remains largely unexplored. In this work, we propose UniMRG (Unified Multi-Representation Generation), a simple yet effective architecture-agnostic post-training method. UniMRG enhances the understanding capabilities of UMMs by incorporating auxiliary generation tasks. Specifically, we train UMMs to generate multiple intrinsic representations of input images, namely pixel (reconstruction), depth (geometry), and segmentation (structure), alongside standard visual understanding objectives. By synthesizing these diverse representations, UMMs capture complementary information regarding appearance, spatial relations, and structural layout. Consequently, UMMs develop a deeper and more comprehensive understanding of visual inputs. Extensive experiments across diverse UMM architectures demonstrate that our method notably enhances fine-grained perception, reduces hallucinations, and improves spatial understanding, while simultaneously boosting generation capabilities.

Method: Proposes Bi-Anchor Interpolation Solver with two components: 1) Bidirectional Temporal Perception where a lightweight SideNet learns future/historical velocities without retraining backbone, and 2) Bi-Anchor Velocity Integration using SideNet with two anchor velocities for batched high-order integration.

Result: On ImageNet-256², achieves generation quality comparable to 100+ NFEs Euler solver in just 10 NFEs, maintains high fidelity in 5 NFEs with negligible training costs, and enables seamless integration with existing pipelines.

Conclusion: BA-solver bridges the gap between training-free and training-based acceleration methods, offering plug-and-play versatility with significant speedup while maintaining generation quality.

Abstract: Flow Matching (FM) models have emerged as a leading paradigm for high-fidelity synthesis. However, their reliance on iterative Ordinary Differential Equation (ODE) solving creates a significant latency bottleneck. Existing solutions face a dichotomy: training-free solvers suffer from significant performance degradation at low Neural Function Evaluations (NFEs), while training-based one- or few-steps generation methods incur prohibitive training costs and lack plug-and-play versatility. To bridge this gap, we propose the Bi-Anchor Interpolation Solver (BA-solver). BA-solver retains the versatility of standard training-free solvers while achieving significant acceleration by introducing a lightweight SideNet (1-2% backbone size) alongside the frozen backbone. Specifically, our method is founded on two synergistic components: \textbf{1) Bidirectional Temporal Perception}, where the SideNet learns to approximate both future and historical velocities without retraining the heavy backbone; and 2) Bi-Anchor Velocity Integration, which utilizes the SideNet with two anchor velocities to efficiently approximate intermediate velocities for batched high-order integration. By utilizing the backbone to establish high-precision ``anchors’’ and the SideNet to densify the trajectory, BA-solver enables large interval sizes with minimized error. Empirical results on ImageNet-256^2 demonstrate that BA-solver achieves generation quality comparable to 100+ NFEs Euler solver in just 10 NFEs and maintains high fidelity in as few as 5 NFEs, incurring negligible training costs. Furthermore, BA-solver ensures seamless integration with existing generative pipelines, facilitating downstream tasks such as image editing.

Method: Uses hybrid representation coupling explicit geometric primitives with neural Gaussians, enabling decoupled geometry and appearance modeling with online initialization and optimization strategy.

Result: Improves dense mesh Chamfer-L2 by 18.52% over PGSR, surpasses ARTDECO by 1.31 dB PSNR, reconstructs ScanNetV2 scenes in under 100 seconds (5x faster than 2D Gaussian Splatting) while matching offline optimization quality.

Conclusion: PLANING achieves both high-quality reconstruction and computational efficiency, making it suitable for large-scale scene modeling and simulation-ready environments for embodied AI applications.

Abstract: Streaming reconstruction from monocular image sequences remains challenging, as existing methods typically favor either high-quality rendering or accurate geometry, but rarely both. We present PLANING, an efficient on-the-fly reconstruction framework built on a hybrid representation that loosely couples explicit geometric primitives with neural Gaussians, enabling geometry and appearance to be modeled in a decoupled manner. This decoupling supports an online initialization and optimization strategy that separates geometry and appearance updates, yielding stable streaming reconstruction with substantially reduced structural redundancy. PLANING improves dense mesh Chamfer-L2 by 18.52% over PGSR, surpasses ARTDECO by 1.31 dB PSNR, and reconstructs ScanNetV2 scenes in under 100 seconds, over 5x faster than 2D Gaussian Splatting, while matching the quality of offline per-scene optimization. Beyond reconstruction quality, the structural clarity and computational efficiency of PLANING make it well suited for a broad range of downstream applications, such as enabling large-scale scene modeling and simulation-ready environments for embodied AI. Project page: https://city-super.github.io/PLANING/ .

Method: JAF uses joint inference across a cohort of query-response pairs with belief propagation and ensemble learning principles. It employs a flexible locality-sensitive hashing (LSH) algorithm that integrates semantic embeddings, LLM-driven hash predicates, categorical supervision, and side information to select diverse exemplars for in-context learning.

Result: Validated on cloud misconfiguration triage in large-scale cloud environments, showing improved evaluation and feedback capabilities compared to isolated evaluation approaches.

Conclusion: JAF elevates judge agents from local evaluators to holistic learners by enabling cross-instance pattern recognition and collective perspective feedback, improving agent self-refinement through joint inference across related responses.

Abstract: Judge agents are fundamental to agentic AI frameworks: they provide automated evaluation, and enable iterative self-refinement of reasoning processes. We introduce JAF: Judge Agent Forest, a framework in which the judge agent conducts joint inference across a cohort of query–response pairs generated by a primary agent, rather than evaluating each in isolation. This paradigm elevates the judge from a local evaluator to a holistic learner: by simultaneously assessing related responses, the judge discerns cross-instance patterns and inconsistencies, whose aggregate feedback enables the primary agent to improve by viewing its own outputs through the judge’s collective perspective. Conceptually, JAF bridges belief propagation and ensemble-learning principles: overlapping in-context neighborhoods induce a knowledge-graph structure that facilitates propagation of critique, and repeated, randomized evaluations yield a robust ensemble of context-sensitive judgments. JAF can be instantiated entirely via ICL, with the judge prompted for each query using its associated primary-agent response plus a small, possibly noisy set of peer exemplars. While kNN in embedding space is a natural starting point for exemplars, this approach overlooks categorical structure, domain metadata, or nuanced distinctions accessible to modern LLMs. To overcome these limitations, we develop a flexible locality-sensitive hashing (LSH) algorithm that learns informative binary codes by integrating semantic embeddings, LLM-driven hash predicates, supervision from categorical labels, and relevant side information. These hash codes support efficient, interpretable, and relation-aware selection of diverse exemplars, and further optimize exploration of CoT reasoning paths. We validate JAF with an empirical study on the demanding task of cloud misconfigs triage in large-scale cloud environments.

Method: Three-component architecture: (1) task decomposition into dependency tree of atomic actions, (2) micro-agent sampling with parallel execution across diverse LLMs, (3) consensus voting with dynamic scaling that clusters outputs and selects winning cluster.

Result: Proven mathematical error reduction to O(p^{ceil(n/2)}), achieving 0.11% error with 5 agents (from 5% baseline) and 3.4 DPMO with 13 agents. 14,700x reliability improvement with 80% cost reduction across enterprise use cases.

Conclusion: Enterprise AI reliability emerges from principled redundancy and consensus rather than model scaling alone, establishing Six Sigma standards for LLM deployment.

Abstract: Large Language Models demonstrate remarkable capabilities yet remain fundamentally probabilistic, presenting critical reliability challenges for enterprise deployment. We introduce the Six Sigma Agent, a novel architecture that achieves enterprise-grade reliability through three synergistic components: (1) task decomposition into a dependency tree of atomic actions; (2) micro-agent sampling where each task is executed n times in parallel across diverse LLMs to generate independent outputs; and (3) consensus voting with dynamic scaling, clustering outputs and selecting the answer from the winning cluster with maximum votes. We prove that sampling n independent outputs with error rate p achieves system error O(p^{ceil(n/2)}), enabling exponential reliability gains. Even using cheaper models with 5% per-action error, consensus voting with 5 agents reduces error to 0.11%; dynamic scaling to 13 agents achieves 3.4 DPMO (Defects Per Million Opportunities), the Six Sigma standard. Evaluation across three enterprise use cases demonstrates a 14,700x reliability improvement over single-agent execution while reducing costs by 80%. Our work establishes that reliability in AI systems emerges from principled redundancy and consensus rather than model scaling alone.

Method: FLARE (Future-aware Lookahead with Reward Estimation) enforces explicit lookahead, value propagation, and limited commitment in a single model, allowing downstream outcomes to influence early decisions.

Result: FLARE consistently improves task performance and planning-level behavior across multiple benchmarks, agent frameworks, and LLM backbones, with LLaMA-8B+FLARE often outperforming GPT-4o with standard reasoning.

Conclusion: The results establish a clear distinction between reasoning and planning, showing that future-aware planning is essential for long-horizon tasks where step-wise reasoning fails.

Abstract: Large language model (LLM)-based agents exhibit strong step-by-step reasoning capabilities over short horizons, yet often fail to sustain coherent behavior over long planning horizons. We argue that this failure reflects a fundamental mismatch: step-wise reasoning induces a form of step-wise greedy policy that is adequate for short horizons but fails in long-horizon planning, where early actions must account for delayed consequences. From this planning-centric perspective, we study LLM-based agents in deterministic, fully structured environments with explicit state transitions and evaluation signals. Our analysis reveals a core failure mode of reasoning-based policies: locally optimal choices induced by step-wise scoring lead to early myopic commitments that are systematically amplified over time and difficult to recover from. We introduce FLARE (Future-aware Lookahead with Reward Estimation) as a minimal instantiation of future-aware planning to enforce explicit lookahead, value propagation, and limited commitment in a single model, allowing downstream outcomes to influence early decisions. Across multiple benchmarks, agent frameworks, and LLM backbones, FLARE consistently improves task performance and planning-level behavior, frequently allowing LLaMA-8B with FLARE to outperform GPT-4o with standard step-by-step reasoning. These results establish a clear distinction between reasoning and planning.

Method: Evaluated multiple LLM families on (1) benchmarks testing core axioms of rational choice, and (2) classic decision domains from behavioral economics where emotions shape judgment. Used two emotion-steering methods: in-context priming (ICP) and representation-level steering (RLS) to probe affective distortions.

Result: Deliberate “thinking” reliably improves rationality and pushes models toward expected-value maximization. ICP induces strong directional shifts that are extreme and difficult to calibrate, while RLS produces more psychologically plausible patterns but with lower reliability. Mechanisms that improve rationality also amplify sensitivity to affective interventions.

Conclusion: There’s a tension between reasoning and affective steering in LLMs, with implications for both human simulation and safe deployment of LLM-based decision systems. Different steering methods trade off controllability against human-aligned behavior.

Abstract: Large Language Models (LLMs) are increasingly positioned as decision engines for hiring, healthcare, and economic judgment, yet real-world human judgment reflects a balance between rational deliberation and emotion-driven bias. If LLMs are to participate in high-stakes decisions or serve as models of human behavior, it is critical to assess whether they exhibit analogous patterns of (ir)rationalities and biases. To this end, we evaluate multiple LLM families on (i) benchmarks testing core axioms of rational choice and (ii) classic decision domains from behavioral economics and social norms where emotions are known to shape judgment and choice. Across settings, we show that deliberate “thinking” reliably improves rationality and pushes models toward expected-value maximization. To probe human-like affective distortions and their interaction with reasoning, we use two emotion-steering methods: in-context priming (ICP) and representation-level steering (RLS). ICP induces strong directional shifts that are often extreme and difficult to calibrate, whereas RLS produces more psychologically plausible patterns but with lower reliability. Our results suggest that the same mechanisms that improve rationality also amplify sensitivity to affective interventions, and that different steering methods trade off controllability against human-aligned behavior. Overall, this points to a tension between reasoning and affective steering, with implications for both human simulation and the safe deployment of LLM-based decision systems.

Method: Formulates protocol design as search over strategies in imperfect information games, using SMT for correctness specification. Combines specialized Monte Carlo Tree Search with transformer-based action encoder, global depth-first search to escape local minima, and repeated model checker feedback.

Result: GGMS can learn correct protocols for larger settings than existing methods. Output protocols are verified correct via exhaustive model checking for all bounded executions. Proves search completeness: if a correct protocol exists, GGMS will eventually find it.

Conclusion: GGMS provides an effective automated approach for distributed protocol design with provable correctness guarantees and completeness properties, scaling better than previous methods.

Abstract: Provably correct distributed protocols, which are a critical component of modern distributed systems, are highly challenging to design and have often required decades of human effort. These protocols allow multiple agents to coordinate to come to a common agreement in an environment with uncertainty and failures. We formulate protocol design as a search problem over strategies in a game with imperfect information, and the desired correctness conditions are specified in Satisfiability Modulo Theories (SMT). However, standard methods for solving multi-agent games fail to learn correct protocols in this setting, even when the number of agents is small. We propose a learning framework, GGMS, which integrates a specialized variant of Monte Carlo Tree Search with a transformer-based action encoder, a global depth-first search to break out of local minima, and repeated feedback from a model checker. Protocols output by GGMS are verified correct via exhaustive model checking for all executions within the bounded setting. We further prove that, under mild assumptions, the search process is complete: if a correct protocol exists, GGMS will eventually find it. In experiments, we show that GGMS can learn correct protocols for larger settings than existing methods.

Method: Hybrid methodology: 1) AI-driven natural language verification using Gemini to narrow the search space of 700 conjectures, 2) Human expert evaluation to assess correctness and novelty of AI-generated solutions, 3) Systematic analysis of problems marked ‘Open’ in the database.

Result: Addressed 13 problems that were marked ‘Open’: 5 through seemingly novel autonomous solutions, and 8 through identification of previous solutions in existing literature. Found that ‘Open’ status was often due to obscurity rather than difficulty. Identified key issues: difficulty of literature identification by AI and risk of ‘subconscious plagiarism’.

Conclusion: AI can assist in mathematics discovery but faces significant challenges with literature search and originality assessment. The ‘Open’ status in databases may reflect obscurity rather than difficulty. Hybrid human-AI approaches are valuable, but careful evaluation is needed to avoid plagiarism and ensure proper attribution.

Abstract: We present a case study in semi-autonomous mathematics discovery, using Gemini to systematically evaluate 700 conjectures labeled ‘Open’ in Bloom’s Erdős Problems database. We employ a hybrid methodology: AI-driven natural language verification to narrow the search space, followed by human expert evaluation to gauge correctness and novelty. We address 13 problems that were marked ‘Open’ in the database: 5 through seemingly novel autonomous solutions, and 8 through identification of previous solutions in the existing literature. Our findings suggest that the ‘Open’ status of the problems was through obscurity rather than difficulty. We also identify and discuss issues arising in applying AI to math conjectures at scale, highlighting the difficulty of literature identification and the risk of ‘‘subconscious plagiarism’’ by AI. We reflect on the takeaways from AI-assisted efforts on the Erdős Problems.

Method: Evaluated traditional ML (Random Forest, SVM, AdaBoost) and deep learning (custom CNNs, VGG16, ResNet50, DenseNet121, EfficientNetB0, InceptionV3) on 25,077 waste images with 80/20 train/test split, using PCA for dimensionality reduction on traditional models.

Result: DenseNet121 achieved highest accuracy (91%) and ROC-AUC (0.98), outperforming best traditional classifier by 20 percentage points. PCA showed negligible benefit for classical methods, while transfer learning substantially improved performance under limited-data conditions.

Conclusion: Transfer learning models like DenseNet121 are highly effective for waste classification, enabling integration into real-time Data-Driven Decision Support Systems for automated waste sorting with potential environmental benefits.

Abstract: Efficient waste sorting is crucial for enabling circular-economy practices and resource recovery in smart cities. This paper evaluates both traditional machine-learning (Random Forest, SVM, AdaBoost) and deep-learning techniques including custom CNNs, VGG16, ResNet50, and three transfer-learning models (DenseNet121, EfficientNetB0, InceptionV3) for binary classification of 25 077 waste images (80/20 train/test split, augmented and resized to 150x150 px). The paper assesses the impact of Principal Component Analysis for dimensionality reduction on traditional models. DenseNet121 achieved the highest accuracy (91 %) and ROC-AUC (0.98), outperforming the best traditional classifier by 20 pp. Principal Component Analysis (PCA) showed negligible benefit for classical methods, whereas transfer learning substantially improved performance under limited-data conditions. Finally, we outline how these models integrate into a real-time Data-Driven Decision Support System for automated waste sorting, highlighting potential reductions in landfill use and lifecycle environmental impacts.)

Method: LLM-TOPSIS framework combining DistilRoBERTa fine-tuning with fuzzy TOPSIS (Fuzzy TOPSIS using triangular fuzzy numbers) to handle ambiguity in criteria weights and scores for candidate ranking.

Result: Achieved 91% accuracy for Experience and Overall attributes, with rankings closely aligned with human expert evaluations.

Conclusion: NLP-driven frameworks with fuzzy decision-making show potential for improving recruitment procedures by enhancing scalability, consistency, and reducing prejudice.

Abstract: In this highly competitive employment environment, the selection of suitable personnel is essential for organizational success. This study presents an automated personnel selection system that utilizes sophisticated natural language processing (NLP) methods to assess and rank software engineering applicants. A distinctive dataset was created by aggregating LinkedIn profiles that include essential features such as education, work experience, abilities, and self-introduction, further enhanced with expert assessments to function as standards. The research combines large language models (LLMs) with multicriteria decision-making (MCDM) theory to develop the LLM-TOPSIS framework. In this context, we utilized the TOPSIS method enhanced by fuzzy logic (Fuzzy TOPSIS) to address the intrinsic ambiguity and subjectivity in human assessments. We utilized triangular fuzzy numbers (TFNs) to describe criteria weights and scores, thereby addressing the ambiguity frequently encountered in candidate evaluations. For candidate ranking, the DistilRoBERTa model was fine-tuned and integrated with the fuzzy TOPSIS method, achieving rankings closely aligned with human expert evaluations and attaining an accuracy of up to 91% for the Experience attribute and the Overall attribute. The study underlines the potential of NLP-driven frameworks to improve recruitment procedures by boosting scalability, consistency, and minimizing prejudice. Future endeavors will concentrate on augmenting the dataset, enhancing model interpretability, and verifying the system in actual recruitment scenarios to better evaluate its practical applicability. This research highlights the intriguing potential of merging NLP with fuzzy decision-making methods in personnel selection, enabling scalable and unbiased solutions to recruitment difficulties.

Method: Proposes Betting Probably Approximately Correct (B-PAC) reasoning using inverse propensity scoring estimators to construct test supermartingales for candidate thresholds, dynamically adjusting routing thresholds based on accumulated statistical evidence of safety.

Result: B-PAC reasoning reduces computational overhead by up to 81.01% in thinking model usage while controlling performance loss below user-specified levels, with theoretical guarantees for anytime-valid performance loss control.

Conclusion: B-PAC reasoning provides a principled method for safe and efficient online reasoning under partial feedback, balancing computational efficiency with controlled performance loss.

Abstract: Large Reasoning Models (LRMs) have demonstrated remarkable performance on complex tasks but suffer from high computational costs and latency. While selective thinking strategies improve efficiency by routing easy queries to non-thinking models, existing approaches often incur uncontrollable errors, especially in online settings where the performance loss of a non-thinking model is only partially observed and data are non-stationary. To address this, we propose Betting Probably Approximately Correct (B-PAC) reasoning, a principled method that enables anytime safe and efficient online reasoning under partial feedback. Specifically, we utilize inverse propensity scoring estimators to construct test supermartingales for candidate thresholds, and then dynamically adjust the routing threshold based on the accumulated statistical evidence of safety. Theoretically, we establish the anytime-valid performance loss control and the efficiency of B-PAC reasoning. Extensive experiments demonstrate that B-PAC reasoning significantly reduces computational overhead, decreasing thinking model usage by up to 81.01%, while controlling the performance loss below the user-specified level.

Method: Proposes SYMPHONY, a multi-agent planning framework that integrates a pool of heterogeneous language model-based agents to leverage diverse reasoning patterns, enhancing rollout diversity and exploration in MCTS.

Result: SYMPHONY achieves strong performance with open-source LLMs on consumer hardware, and shows further improvements with cloud-based LLMs via API, outperforming state-of-the-art baselines across multiple benchmark tasks.

Conclusion: Heterogeneous multi-agent coordination effectively enhances planning performance by increasing exploration diversity, demonstrating the superiority of multi-agent frameworks over single-agent approaches in LLM-based planning.

Abstract: Recent advancements have increasingly focused on leveraging large language models (LLMs) to construct autonomous agents for complex problem-solving tasks. However, existing approaches predominantly employ a single-agent framework to generate search branches and estimate rewards during Monte Carlo Tree Search (MCTS) planning. This single-agent paradigm inherently limits exploration capabilities, often resulting in insufficient diversity among generated branches and suboptimal planning performance. To overcome these limitations, we propose Synergistic Multi-agent Planning with Heterogeneous langauge model assembly (SYMPHONY), a novel multi-agent planning framework that integrates a pool of heterogeneous language model-based agents. By leveraging diverse reasoning patterns across agents, SYMPHONY enhances rollout diversity and facilitates more effective exploration. Empirical results across multiple benchmark tasks show that SYMPHONY achieves strong performance even when instantiated with open-source LLMs deployable on consumer-grade hardware. When enhanced with cloud-based LLMs accessible via API, SYMPHONY demonstrates further improvements, outperforming existing state-of-the-art baselines and underscoring the effectiveness of heterogeneous multi-agent coordination in planning tasks.

Method: Derives CIP objective from Optimal Control theory, showing connection between extrinsic and intrinsic behaviors. CIP appears as the gap between open-loop and closed-loop Kolmogorov-Sinai entropies, which simultaneously rewards pursuit and regulation of chaos.

Result: Establishes key theoretical properties of CIP and demonstrates its effectiveness on standard IM benchmarks, showing it can generate intelligent behavior without external utilities.

Conclusion: CIP provides a novel intrinsic motivation principle that bridges optimal control and information theory, offering a more fundamental approach to generating intelligent behavior without external guidance or designer intervention.

Abstract: Intrinsic Motivation (IM) is a paradigm for generating intelligent behavior without external utilities. The existing information-theoretic methods for IM are predominantly based on information transmission, which explicitly depends on the designer’s choice of which random variables engage in transmission. In this work, we introduce a novel IM principle, Controllable Information Production (CIP), that avoids both external utilities and designer-specified variables. We derive the CIP objective from Optimal Control, showing a connection between extrinsic and intrinsic behaviors. CIP appears as the gap between open-loop and closed-loop Kolmogorov-Sinai entropies, which simultaneously rewards the pursuit and regulation of chaos. We establish key theoretical properties of CIP and demonstrate its effectiveness on standard IM benchmarks.

Method: TALC integrates a council of LLMs with Monte Carlo Tree Search. Each LLM has a structured success memory profile from prior task trajectories. The framework routes control to the most contextually appropriate model at each decision point using semantic matching between current context and past successes. It uses a dual-signal mechanism fusing model-based evaluations with historical utility scores, adaptively weighted based on intra-node variance to guide MCTS selection.

Result: Experiments on WebShop, HumanEval, and Game of 24 show TALC achieves superior task success rates and improved search efficiency compared to strong baselines, validating the benefits of specialization-aware routing and adaptive planning.

Conclusion: TALC demonstrates that leveraging LLM specialization differences through dynamic routing and adaptive planning significantly improves decision-making performance across diverse tasks, offering a more efficient approach than treating all models as uniformly applicable.

Abstract: Large language models (LLMs) have shown strong capabilities across diverse decision-making tasks. However, existing approaches often overlook the specialization differences among available models, treating all LLMs as uniformly applicable regardless of task characteristics. This limits their ability to adapt to varying reasoning demands and task complexities. In this work, we propose Task-Aware LLM Council (TALC), a task-adaptive decision framework that integrates a council of LLMs with Monte Carlo Tree Search (MCTS) to enable dynamic expert selection and efficient multi-step planning. Each LLM is equipped with a structured success memory profile derived from prior task trajectories, enabling semantic matching between current reasoning context and past successes. At each decision point, TALC routes control to the most contextually appropriate model and estimates node value using a dual-signal mechanism that fuses model-based evaluations with historical utility scores. These signals are adaptively weighted based on intra-node variance and used to guide MCTS selection, allowing the system to balance exploration depth with planning confidence. Experiments on WebShop, HumanEval, and the Game of 24 demonstrate that TALC achieves superior task success rates and improved search efficiency compared to strong baselines, validating the benefits of specialization-aware routing and adaptive planning.

Method: Provides rigorous theoretical guarantees including: 1) lower bound characterizing fundamental limits of single update step, 2) finite-sample error bounds for full iterative paradigm showing O(1/√n) convergence, 3) analysis showing exponential decay of initialization dependence with iterations, 4) instantiation for linear softmax model class.

Result: Theoretical framework shows self-rewarding succeeds by robustly overcoming poor initialization through steering dynamics toward internal stability and consistency, with performance improving at rate O(1/√n) with sample size.

Conclusion: First rigorous theoretical guarantees for Self-Rewarding Language Models explain why they succeed: exponential decay of initialization dependence enables robust improvement through internal consistency mechanisms.

Abstract: Self-Rewarding Language Models (SRLMs) achieve notable success in iteratively improving alignment without external feedback. Yet, despite their striking empirical progress, the core mechanisms driving their capabilities remain unelucidated, leaving a critical gap in theoretical understanding. This paper provides the first rigorous theoretical guarantees for SRLMs. We first establish a lower bound that characterizes the fundamental limits of a single update step, revealing a critical dependence on the quality of the initial model. We then derive finite-sample error bounds for the full iterative paradigm, showing that performance improves at a rate of $\widetilde{\mathcal{O}}\left(1/\sqrt{n}\right)$ with sample size $n$. Crucially, our analysis reveals that the dependence on the initial model decays exponentially with the number of iterations $T$. This provides a formal explanation for why self-rewarding succeeds: it robustly overcomes poor initialization by steering the dynamics toward internal stability and consistency. Finally, we instantiate our theoretical framework for the linear softmax model class, yielding tailored guarantees that connect our high-level insights to practical model architectures.

Method: Proposes finetuning multiagent systems with per-action process rewards from AI feedback (MAPPA). Assigns credit to individual agent actions rather than only at task completion, enabling fine-grained supervision without ground truth labels while extracting maximal training signal from each rollout.

Result: On unseen math problems, MAPPA achieves +5.0–17.5pp on AIME and +7.8–17.2pp on AMC. For data analysis tasks, improves success rate by +12.5pp while quality metrics improve by up to 30%.

Conclusion: Per-action supervision can lead to improvements across different multiagent systems in various domains. The work takes a first step toward scaling multiagent systems for complex, long-horizon tasks with minimal human supervision.

Abstract: While multiagent systems have shown promise for tackling complex tasks via specialization, finetuning multiple agents simultaneously faces two key challenges: (1) credit assignment across agents, and (2) sample efficiency of expensive multiagent rollouts. In this work, we propose finetuning multiagent systems with per-action process rewards from AI feedback (MAPPA) to address both. Through assigning credit to individual agent actions rather than only at task completion, MAPPA enables fine-grained supervision without ground truth labels while extracting maximal training signal from each rollout. We demonstrate our approach on competition math problems and tool-augmented data analysis tasks. On unseen math problems, MAPPA achieves +5.0–17.5pp on AIME and +7.8–17.2pp on AMC. For data analysis tasks, our method improves success rate by +12.5pp while quality metrics improve by up to 30%, validating that per-action supervision can lead to improvements across different multiagent system on various domains. By addressing these challenges, our work takes a first step toward scaling multiagent systems for complex, long-horizon tasks with minimal human supervision.

Method: Proposes Darwinian Memory System (DMS) that constructs memory as a dynamic ecosystem governed by survival of the fittest. Decomposes complex trajectories into independent reusable units for compositional flexibility, and implements Utility-driven Natural Selection to track survival value, actively pruning suboptimal paths and inhibiting high-risk plans.

Result: Extensive experiments on real-world multi-app benchmarks show DMS boosts general-purpose MLLMs without training costs or architectural overhead, achieving average gains of 18.0% in success rate and 33.9% in execution stability, while reducing task latency.

Conclusion: DMS establishes an effective self-evolving memory system for GUI tasks that addresses key bottlenecks in MLLM agent performance for complex automation scenarios.

Abstract: Multimodal Large Language Model (MLLM) agents facilitate Graphical User Interface (GUI) automation but struggle with long-horizon, cross-application tasks due to limited context windows. While memory systems provide a viable solution, existing paradigms struggle to adapt to dynamic GUI environments, suffering from a granularity mismatch between high-level intent and low-level execution, and context pollution where the static accumulation of outdated experiences drives agents into hallucination. To address these bottlenecks, we propose the Darwinian Memory System (DMS), a self-evolving architecture that constructs memory as a dynamic ecosystem governed by the law of survival of the fittest. DMS decomposes complex trajectories into independent, reusable units for compositional flexibility, and implements Utility-driven Natural Selection to track survival value, actively pruning suboptimal paths and inhibiting high-risk plans. This evolutionary pressure compels the agent to derive superior strategies. Extensive experiments on real-world multi-app benchmarks validate that DMS boosts general-purpose MLLMs without training costs or architectural overhead, achieving average gains of 18.0% in success rate and 33.9% in execution stability, while reducing task latency, establishing it as an effective self-evolving memory system for GUI tasks.

Method: RE-Tab formulates TableQA as a Partially Observable Markov Decision Process and introduces lightweight, training-free reward modeling. It provides explicit verifiable rewards during State Transition (“What is the best action?”) and Simulative Reasoning (“Am I sure about the output?”) to steer agent navigation in table states.

Result: RE-Tab achieves state-of-the-art performance in TableQA with almost 25% drop in inference cost. Plug-and-play implementation brings up to 41.77% improvement in QA accuracy and 33.33% drop in test-time inference samples for consistent answers. Consistent improvements across various LLMs and benchmarks confirm generalizability.

Conclusion: Explicit feedback on table transformation actions through reward modeling significantly improves TableQA reasoning capability. The RE-Tab framework demonstrates that stepwise reasoning with reward feedback in table transformations is crucial for steering agents effectively.

Abstract: A major challenge in training TableQA agents, compared to standard text- and image-based agents, is that answers cannot be inferred from a static input but must be reasoned through stepwise transformations of the table state, introducing multi-step reasoning complexity and environmental interaction. This leads to a research question: Can explicit feedback on table transformation action improve model reasoning capability? In this work, we introduce RE-Tab, a plug-and-play framework that architecturally enhances trajectory search via lightweight, training-free reward modeling by formulating the problem as a Partially Observable Markov Decision Process. We demonstrate that providing explicit verifiable rewards during State Transition (What is the best action?'') and Simulative Reasoning (Am I sure about the output?’’) is crucial to steer the agent’s navigation in table states. By enforcing stepwise reasoning with reward feedback in table transformations, RE-Tab achieves state-of-the-art performance in TableQA with almost 25% drop in inference cost. Furthermore, a direct plug-and-play implementation of RE-Tab brings up to 41.77% improvement in QA accuracy and 33.33% drop in test-time inference samples for consistent answer. Consistent improvement pattern across various LLMs and state-of-the-art benchmarks further confirms RE-Tab’s generalisability. The repository is available at https://github.com/ThomasK1018/RE_Tab .

Method: Proposes CraEG (Crowding-Aware Embedding Geometry) - a plug-and-play sampling method that mitigates crowding through geometry-guided reweighting. It’s training-free, single-pass, and compatible with standard sampling strategies.

Result: Experiments on multiple models and benchmarks show improved generation performance with gains in robustness and diversity metrics. The method demonstrates effectiveness in mathematical problem solving where crowding correlates with reasoning success.

Conclusion: Embedding-space crowding is a meaningful phenomenon affecting LLM reasoning, and geometry-aware sampling methods like CraEG can improve performance without additional training.

Abstract: Sampling-based decoding underlies complex reasoning in large language models (LLMs), where decoding strategies critically shape model behavior. Temperature- and truncation-based methods reshape the next-token distribution through global probability reweighting or thresholding to balance the quality-diversity tradeoff. However, they operate solely on token probabilities, ignoring fine-grained relationships among tokens in the embedding space. We uncover a novel phenomenon, embedding-space crowding, where the next-token distribution concentrates its probability mass on geometrically close tokens in the embedding space. We quantify crowding at multiple granularities and find a statistical association with reasoning success in mathematical problem solving. Motivated by this finding, we propose CraEG, a plug-and-play sampling method that mitigates crowding through geometry-guided reweighting. CraEG is training-free, single-pass, and compatible with standard sampling strategies. Experiments on multiple models and benchmarks demonstrate improved generation performance, with gains in robustness and diversity metrics.

Method: Proposes CoBel-World framework with Collaborative Belief World - an internal representation modeling both physical environment and collaborators’ mental states. Uses symbolic belief representation module to parse external knowledge into structured beliefs, and performs zero-shot Bayesian-style belief updates through LLM reasoning.

Result: On embodied benchmarks (TDW-MAT and C-WAH), CoBel-World reduces communication costs by 64-79% and improves task completion efficiency by 4-28% compared to strongest baselines.

Conclusion: Explicit, intent-aware belief modeling is essential for efficient and human-like collaboration in LLM-based multi-agent systems, demonstrating that LLMs can effectively reason about collaborators’ mental states.

Abstract: Effective real-world multi-agent collaboration requires not only accurate planning but also the ability to reason about collaborators’ intents–a crucial capability for avoiding miscoordination and redundant communication under partial observable environments. Due to their strong planning and reasoning capabilities, large language models (LLMs) have emerged as promising autonomous agents for collaborative task solving. However, existing collaboration frameworks for LLMs overlook their reasoning potential for dynamic intent inference, and thus produce inconsistent plans and redundant communication, reducing collaboration efficiency. To bridge this gap, we propose CoBel-World, a novel framework that equips LLM agents with a Collaborative Belief World–an internal representation jointly modeling the physical environment and collaborators’ mental states. CoBel-World enables agents to parse external open-world knowledge into structured beliefs via a symbolic belief representation module, and perform zero-shot Bayesian-style belief updates through LLM reasoning. This allows agents to proactively detect potential miscoordination (e.g., conflicting plans) and communicate adaptively. Evaluated on challenging embodied benchmarks (i.e., TDW-MAT and C-WAH), CoBel-World significantly reduces communication costs by 64-79% and improves task completion efficiency by 4-28% compared to the strongest baseline. Our results show that explicit, intent-aware belief modeling is essential for efficient and human-like collaboration in LLM-based multi-agent systems.

Method: Three core mechanisms: (1) Performance-Aware Selection Modeling (PASM) replaces generic descriptions with multi-dimensional scoring based on fine-grained performance evaluations; (2) Adaptive Preference Update (APU) dynamically optimizes tool selection by comparing theoretical vs. actual execution rankings; (3) Capability-Aligned Planning Optimization (CAPO) guides planners to generate subtasks aligned with performance-aware strategies.

Result: Experimental comparisons show PerfGuard’s advantages in tool selection accuracy, execution reliability, and alignment with user intent, validating robustness and practical utility for complex AIGC tasks.

Conclusion: PerfGuard addresses the critical gap in existing agent frameworks by systematically modeling tool performance boundaries, enabling more reliable and effective visual content generation through performance-aware planning and execution.

Abstract: The advancement of Large Language Model (LLM)-powered agents has enabled automated task processing through reasoning and tool invocation capabilities. However, existing frameworks often operate under the idealized assumption that tool executions are invariably successful, relying solely on textual descriptions that fail to distinguish precise performance boundaries and cannot adapt to iterative tool updates. This gap introduces uncertainty in planning and execution, particularly in domains like visual content generation (AIGC), where nuanced tool performance significantly impacts outcomes. To address this, we propose PerfGuard, a performance-aware agent framework for visual content generation that systematically models tool performance boundaries and integrates them into task planning and scheduling. Our framework introduces three core mechanisms: (1) Performance-Aware Selection Modeling (PASM), which replaces generic tool descriptions with a multi-dimensional scoring system based on fine-grained performance evaluations; (2) Adaptive Preference Update (APU), which dynamically optimizes tool selection by comparing theoretical rankings with actual execution rankings; and (3) Capability-Aligned Planning Optimization (CAPO), which guides the planner to generate subtasks aligned with performance-aware strategies. Experimental comparisons against state-of-the-art methods demonstrate PerfGuard’s advantages in tool selection accuracy, execution reliability, and alignment with user intent, validating its robustness and practical utility for complex AIGC tasks. The project code is available at https://github.com/FelixChan9527/PerfGuard.

Method: Dual-branch Transformer architecture with self- and cross-attention to separate intrinsic and weather-induced patterns, enhanced with memory banks, adaptive gating fusion, a weather discriminator for explicit disentanglement, and causal data augmentation that perturbs non-causal parts while preserving causal structures.

Result: Experiments on taxi-flow datasets from three cities demonstrate robust performance under extreme weather conditions, outperforming existing methods and showing potential for safer mobility, disaster preparedness, and urban resilience.

Conclusion: WED-Net effectively addresses urban spatio-temporal prediction challenges under extreme weather by disentangling weather effects from intrinsic patterns and improving generalization through causal data augmentation.

Abstract: Urban spatio-temporal prediction under extreme conditions (e.g., heavy rain) is challenging due to event rarity and dynamics. Existing data-driven approaches that incorporate weather as auxiliary input often rely on coarse-grained descriptors and lack dedicated mechanisms to capture fine-grained spatio-temporal effects. Although recent methods adopt causal techniques to improve out-of-distribution generalization, they typically overlook temporal dynamics or depend on fixed confounder stratification. To address these limitations, we propose WED-Net (Weather-Effect Disentanglement Network), a dual-branch Transformer architecture that separates intrinsic and weather-induced traffic patterns via self- and cross-attention, enhanced with memory banks and fused through adaptive gating. To further promote disentanglement, we introduce a discriminator that explicitly distinguishes weather conditions. Additionally, we design a causal data augmentation strategy that perturbs non-causal parts while preserving causal structures, enabling improved generalization under rare scenarios. Experiments on taxi-flow datasets from three cities demonstrate that WED-Net delivers robust performance under extreme weather conditions, highlighting its potential to support safer mobility, highlighting its potential to support safer mobility, disaster preparedness, and urban resilience in real-world settings. The code is publicly available at https://github.com/HQ-LV/WED-Net.

Method: Introduces active learning into RLVR framework. Proposes uncertainty consistency metric to align subjective and objective uncertainty. Uses Point-Biserial Correlation Coefficient for offline setting and introduces new online variant computed from normalized advantage and subjective uncertainty for dynamic training scenarios.

Result: Method consistently outperforms random and classic AL baselines, achieving full-dataset performance while training on only 30% of the data, effectively reducing RLVR costs for reasoning tasks.

Conclusion: Active learning can significantly reduce annotation costs in RLVR for mathematical reasoning by selecting more informative samples through proper uncertainty alignment metrics.

Abstract: Large Language Models (LLMs) have recently improved mathematical reasoning through Reinforcement Learning with Verifiable Reward (RLVR). However, existing RLVR algorithms require large query budgets, making annotation costly. We investigate whether fewer but more informative queries can yield similar or superior performance, introducing active learning (AL) into RLVR. We identify that classic AL sampling strategies fail to outperform random selection in this setting, due to ignoring objective uncertainty when only selecting by subjective uncertainty. This work proposes an uncertainty consistency metric to evaluate how well subjective uncertainty aligns with objective uncertainty. In the offline setting, this alignment is measured using the Point-Biserial Correlation Coefficient (PBC). For online training, because of limited sampling and dynamically shifting output distributions, PBC estimation is difficult. Therefore, we introduce a new online variant, computed from normalized advantage and subjective uncertainty. Theoretically, we prove that the online variant is strictly negatively correlated with offline PBC and supports better sample selection. Experiments show our method consistently outperforms random and classic AL baselines, achieving full-dataset performance while training on only 30% of the data, effectively reducing the cost of RLVR for reasoning tasks.

Method: Proposes EigenData, a hierarchical multi-agent engine that synthesizes tool-grounded dialogues with executable per-instance checkers, using closed-loop self-evolving process to update prompts and workflow. Then develops RL recipe that fine-tunes user model and applies GRPO-style training with trajectory-level group-relative advantages and dynamic filtering.

Result: Achieves 73.0% pass^1 on Airline and 98.3% pass^1 on Telecom benchmarks in tau^2-bench, matching or exceeding frontier models. Demonstrates consistent improvements beyond supervised fine-tuning.

Conclusion: Provides a scalable pathway for bootstrapping complex tool-using behaviors without expensive human annotation, combining self-evolving data synthesis with verifier-based RL for effective agent training.

Abstract: Interactive tool-using agents must solve real-world tasks via multi-turn interaction with both humans and external environments, requiring dialogue state tracking, multi-step tool execution, while following complex instructions. Post-training such agents is challenging because synthesis for high-quality multi-turn tool-use data is difficult to scale, and reinforcement learning (RL) could face noisy signals caused by user simulation, leading to degraded training efficiency. We propose a unified framework that combines a self-evolving data agent with verifier-based RL. Our system, EigenData, is a hierarchical multi-agent engine that synthesizes tool-grounded dialogues together with executable per-instance checkers, and improves generation reliability via closed-loop self-evolving process that updates prompts and workflow. Building on the synthetic data, we develop an RL recipe that first fine-tunes the user model and then applies GRPO-style training with trajectory-level group-relative advantages and dynamic filtering, yielding consistent improvements beyond SFT. Evaluated on tau^2-bench, our best model reaches 73.0% pass^1 on Airline and 98.3% pass^1 on Telecom, matching or exceeding frontier models. Overall, our results suggest a scalable pathway for bootstrapping complex tool-using behaviors without expensive human annotation.

Method: EntroCut is a training-free method that monitors the entropy of the model’s output distribution during early reasoning steps. It identifies high-confidence states where reasoning can be safely terminated, dynamically truncating the reasoning process without requiring additional training.

Result: Experiments on four benchmarks show EntroCut reduces token usage by up to 40% with minimal accuracy sacrifice. The method achieves superior efficiency-performance trade-offs compared to existing training-free methods, as measured by the proposed Efficiency-Performance Ratio (EPR) metric.

Conclusion: Entropy-guided dynamic truncation provides a practical approach to mitigate the inefficiency of Large Reasoning Models, demonstrating that early reasoning step entropy can reliably guide when to terminate reasoning without compromising accuracy.

Abstract: Large Reasoning Models (LRMs) excel at complex reasoning tasks through extended chain-of-thought generation, but their reliance on lengthy intermediate steps incurs substantial computational cost. We find that the entropy of the model’s output distribution in early reasoning steps reliably distinguishes correct from incorrect reasoning. Motivated by this observation, we propose EntroCut, a training-free method that dynamically truncates reasoning by identifying high-confidence states where reasoning can be safely terminated. To comprehensively evaluate the trade-off between efficiency and accuracy, we introduce the Efficiency-Performance Ratio (EPR), a unified metric that quantifies relative token savings per unit accuracy loss. Experiments on four benchmarks show that EntroCut reduces token usage by up to 40% with minimal accuracy sacrifice, achieving superior efficiency-performance trade-offs compared with existing training-free methods. These results demonstrate that entropy-guided dynamic truncation provides a practical approach to mitigate the inefficiency of LRMs.

Method: Proposes SABER (Scaling-Aware Best-of-N Estimation of Risk) that models sample-level success probabilities using a Beta distribution (conjugate prior of Bernoulli distribution). Derives an analytic scaling law that enables reliable extrapolation of large-N attack success rates from small-budget measurements.

Result: Using only n=100 samples, SABER predicts ASR@1000 with mean absolute error of 1.66 (86.2% reduction compared to baseline error of 12.04). Reveals heterogeneous risk scaling profiles and shows models appearing robust under standard evaluation can experience rapid nonlinear risk amplification under parallel adversarial pressure.

Conclusion: Provides a low-cost, scalable methodology for realistic LLM safety assessment that accounts for parallel adversarial pressure, enabling better prediction of jailbreak vulnerability at scale.

Abstract: Large Language Models (LLMs) are typically evaluated for safety under single-shot or low-budget adversarial prompting, which underestimates real-world risk. In practice, attackers can exploit large-scale parallel sampling to repeatedly probe a model until a harmful response is produced. While recent work shows that attack success increases with repeated sampling, principled methods for predicting large-scale adversarial risk remain limited. We propose a scaling-aware Best-of-N estimation of risk, SABER, for modeling jailbreak vulnerability under Best-of-N sampling. We model sample-level success probabilities using a Beta distribution, the conjugate prior of the Bernoulli distribution, and derive an analytic scaling law that enables reliable extrapolation of large-N attack success rates from small-budget measurements. Using only n=100 samples, our anchored estimator predicts ASR@1000 with a mean absolute error of 1.66, compared to 12.04 for the baseline, which is an 86.2% reduction in estimation error. Our results reveal heterogeneous risk scaling profiles and show that models appearing robust under standard evaluation can experience rapid nonlinear risk amplification under parallel adversarial pressure. This work provides a low-cost, scalable methodology for realistic LLM safety assessment. We will release our code and evaluation scripts upon publication to future research.

Method: The authors introduce Meddollina, a governance-first clinical intelligence system designed to constrain inference before language realization. It formalizes Clinical Contextual Intelligence (CCI) with capabilities including persistent context awareness, intent preservation, bounded inference, and principled deferral. Meddollina acts as a continuous intelligence layer supporting clinical workflows while preserving clinician authority.

Result: Evaluated across 16,412+ heterogeneous medical queries against general-purpose models, medical-tuned models, and retrieval-augmented systems, Meddollina exhibits calibrated uncertainty, conservative reasoning under underspecification, stable longitudinal constraint adherence, and reduced speculative completion relative to generation-centric baselines.

Conclusion: Deployable medical AI will not emerge from scaling alone; instead, progress should be measured by clinician-aligned behavior under uncertainty rather than fluency-driven completion, motivating a shift toward Continuous Clinical Intelligence.

Abstract: Generative medical AI now appears fluent and knowledgeable enough to resemble clinical intelligence, encouraging the belief that scaling will make it safe. But clinical reasoning is not text generation. It is a responsibility-bound process under ambiguity, incomplete evidence, and longitudinal context. Even as benchmark scores rise, generation-centric systems still show behaviours incompatible with clinical deployment: premature closure, unjustified certainty, intent drift, and instability across multi-step decisions. We argue these are structural consequences of treating medicine as next-token prediction. We formalise Clinical Contextual Intelligence (CCI) as a distinct capability class required for real-world clinical use, defined by persistent context awareness, intent preservation, bounded inference, and principled deferral when evidence is insufficient. We introduce Meddollina, a governance-first clinical intelligence system designed to constrain inference before language realisation, prioritising clinical appropriateness over generative completeness. Meddollina acts as a continuous intelligence layer supporting clinical workflows while preserving clinician authority. We evaluate Meddollina using a behaviour-first regime across 16,412+ heterogeneous medical queries, benchmarking against general-purpose models, medical-tuned models, and retrieval-augmented systems. Meddollina exhibits a distinct behavioural profile: calibrated uncertainty, conservative reasoning under underspecification, stable longitudinal constraint adherence, and reduced speculative completion relative to generation-centric baselines. These results suggest deployable medical AI will not emerge from scaling alone, motivating a shift toward Continuous Clinical Intelligence, where progress is measured by clinician-aligned behaviour under uncertainty rather than fluency-driven completion.

Method: Proposes TMoW with: (1) multi-granular prototype-based routing across object- to scene-level similarities, (2) test-time refinement aligning unseen domain features with prototypes during inference, and (3) distilled mixture-based augmentation constructing new models from few-shot data and existing prototypes.

Result: Evaluated on VirtualHome, ALFWorld, and RLBench benchmarks, showing strong performance in zero-shot adaptation and few-shot expansion scenarios, enabling effective operation in dynamic environments.

Conclusion: TMoW enhances embodied agents’ adaptability to unseen and evolving domains through test-time routing updates, enabling effective operation in dynamic environments where conventional MoE architectures fail.

Abstract: Language model (LM)-based embodied agents are increasingly deployed in real-world settings. Yet, their adaptability remains limited in dynamic environments, where constructing accurate and flexible world models is crucial for effective reasoning and decision-making. To address this challenge, we extend the Mixture-of-Experts (MoE) paradigm to embodied agents. While conventional MoE architectures modularize knowledge into expert components with pre-trained routing, they remain rigid once deployed, making them less effective for adapting to unseen domains in dynamic environments. We therefore propose Test-time Mixture of World Models (TMoW), a framework that enhances adaptability to unseen and evolving domains. TMoW updates its routing function over world models at test time, unlike conventional MoE where the function remains fixed, enabling agents to recombine existing models and integrate new ones for continual adaptation. It achieves this through (i) multi-granular prototype-based routing, which adapts mixtures across object- to scene-level similarities, (ii) test-time refinement that aligns unseen domain features with prototypes during inference, and (iii) distilled mixture-based augmentation, which efficiently constructs new models from few-shot data and existing prototypes. We evaluate TMoW on VirtualHome, ALFWorld, and RLBench benchmarks, demonstrating strong performance in both zero-shot adaptation and few-shot expansion scenarios, and showing that it enables embodied agents to operate effectively in dynamic environments.

Method: Proposes UnCertainty-Aware Policy Optimization (UCPO) with two key components: 1) Ternary Advantage Decoupling to separate and independently normalize deterministic and uncertain rollouts, eliminating advantage bias; 2) Dynamic Uncertainty Reward Adjustment to calibrate uncertainty weights in real-time based on model evolution and instance difficulty.

Result: Experimental results in mathematical reasoning and general tasks show UCPO effectively resolves reward imbalance, significantly improving model reliability and calibration beyond knowledge boundaries.

Conclusion: UCPO provides an effective framework for endowing LLMs with inherent uncertainty expression capabilities, addressing fundamental limitations in current RL paradigms for building trustworthy AI systems.

Abstract: The key to building trustworthy Large Language Models (LLMs) lies in endowing them with inherent uncertainty expression capabilities to mitigate the hallucinations that restrict their high-stakes applications. However, existing RL paradigms such as GRPO often suffer from Advantage Bias due to binary decision spaces and static uncertainty rewards, inducing either excessive conservatism or overconfidence. To tackle this challenge, this paper unveils the root causes of reward hacking and overconfidence in current RL paradigms incorporating uncertainty-based rewards, based on which we propose the UnCertainty-Aware Policy Optimization (UCPO) framework. UCPO employs Ternary Advantage Decoupling to separate and independently normalize deterministic and uncertain rollouts, thereby eliminating advantage bias. Furthermore, a Dynamic Uncertainty Reward Adjustment mechanism is introduced to calibrate uncertainty weights in real-time according to model evolution and instance difficulty. Experimental results in mathematical reasoning and general tasks demonstrate that UCPO effectively resolves the reward imbalance, significantly improving the reliability and calibration of the model beyond their knowledge boundaries.

Method: R2M introduces a novel lightweight RLHF framework that leverages evolving hidden states of the policy model (policy feedback) to align the reward model with real-time distribution shifts of the policy during RL training. Unlike vanilla reward models that depend only on pretrained LLM semantic representations, R2M dynamically adapts to policy changes.

Result: The framework addresses reward overoptimization by reducing the increasing reward discrepancy between reward models and policy models, leading to more stable and aligned training.

Conclusion: R2M points to a promising new direction for improving reward model performance through real-time utilization of feedback from policy models, potentially enhancing the alignment of LLMs with human preferences.

Abstract: Reinforcement Learning from Human Feedback (RLHF) is a pivotal technique for aligning large language models (LLMs) with human preferences, yet it is susceptible to reward overoptimization, in which policy models overfit to the reward model, exploit spurious reward patterns instead of faithfully capturing human intent. Prior mitigations primarily relies on surface semantic information and fails to efficiently address the misalignment between the reward model (RM) and the policy model caused by continuous policy distribution shifts. This inevitably leads to an increasing reward discrepancy, exacerbating reward overoptimization. To address these limitations, we introduce R2M (Real-Time Aligned Reward Model), a novel lightweight RLHF framework. R2M goes beyond vanilla reward models that solely depend on the semantic representations of a pretrained LLM. Instead, it leverages the evolving hidden states of the policy (namely policy feedback) to align with the real-time distribution shift of the policy during the RL process. This work points to a promising new direction for improving the performance of reward models through real-time utilization of feedback from policy models.

Method: Decouples the VLM’s role as action proposer from final action selection. Keeps the VLM policy frozen to generate candidate actions for a given state, then uses a lightweight, offline-trained Q-function to rerank these candidates, executing the action with highest estimated value.

Result: Significantly boosts agent success rates on WebVoyager benchmark: improves Qwen2.5-VL-7B agent from 38.8% to 55.7% and proprietary GPT-4.1 agent from 82.4% to 88.8%.

Conclusion: The approach enables immediate policy improvement at inference time without retraining, offering a practical solution for adapting VLMs to dynamic environments like the web while maintaining computational efficiency.

Abstract: Vision-Language Models (VLMs) have become powerful backbones for agents to autonomously operate in digital environments like the web and operating systems. However, these models suffer from inadaptability to fast-changing environments like the web, which can be alleviated by fine-tuning requiring expansive model training and data collection. In this work, we introduce a novel paradigm for enhancing agentic VLM policies at inference without policy retraining. Fundamentally, our approach decouples the VLM’s role as a high-capacity action proposer from the final action selection mechanism. We keep the VLM policy frozen and use it to generate a set of candidate actions for a given state. Then, a lightweight, offline-trained Q-function reranks these candidates, and the agent executes the action with the highest estimated value. The main contribution is to apply the Q-function directly during inference for immediate policy improvement, and not offline to relabel data for policy retraining. We demonstrate on the academic WebVoyager benchmark that our method significantly boosts agent success rates, improving a Qwen2.5-VL-7B agent from 38.8% to 55.7% and a proprietary GPT-4.1 agent from 82.4% to 88.8%.

Method: Proposes Minimum Prefix Ratio (MinPRO) which replaces unstable cumulative prefix importance ratio with a non-cumulative surrogate based on the minimum token-level ratio observed in the preceding prefix.

Result: MinPRO substantially improves training stability and peak performance in off-policy regimes across multiple mathematical reasoning benchmarks on both dense and mixture-of-experts LLMs.

Conclusion: MinPRO provides a simple yet effective solution to stabilize LLM optimization under large off-policy drift, addressing a key limitation in current RL post-training approaches.

Abstract: Reinforcement learning (RL) post-training has increasingly demonstrated strong ability to elicit reasoning behaviors in large language models (LLMs). For training efficiency, rollouts are typically generated in an off-policy manner using an older sampling policy and then used to update the current target policy. To correct the resulting discrepancy between the sampling and target policies, most existing RL objectives rely on a token-level importance sampling ratio, primarily due to its computational simplicity and numerical stability. However, we observe that token-level correction often leads to unstable training dynamics when the degree of off-policyness is large. In this paper, we revisit LLM policy optimization under off-policy conditions and show that the theoretically rigorous correction term is the prefix importance ratio, and that relaxing it to a token-level approximation can induce instability in RL post-training. To stabilize LLM optimization under large off-policy drift, we propose a simple yet effective objective, Minimum Prefix Ratio (MinPRO). MinPRO replaces the unstable cumulative prefix ratio with a non-cumulative surrogate based on the minimum token-level ratio observed in the preceding prefix. Extensive experiments on both dense and mixture-of-experts LLMs, across multiple mathematical reasoning benchmarks, demonstrate that MinPRO substantially improves training stability and peak performance in off-policy regimes.

Method: Extracts dual-form experience patterns: specialized subagents with independent reasoning/memory for procedural subtasks, and skill patterns as guidelines/code snippets for static knowledge. Includes continuous maintenance mechanism that scores, prunes, and merges patterns to prevent degradation.

Result: Achieves 98.4% on ALFWorld, 70.4% on ScienceWorld, and 27.1% on TravelPlanner with 20-73% step reductions. On TravelPlanner, automatic extraction exceeds manually designed systems (27.1% vs 12.1%), demonstrating ability to capture procedural coordination.

Conclusion: AutoRefine effectively extracts and maintains experience patterns from agent histories, enabling knowledge accumulation and procedural coordination capture, outperforming manual systems.

Abstract: Large language model agents often fail to accumulate knowledge from experience, treating each task as an independent challenge. Recent methods extract experience as flattened textual knowledge, which cannot capture procedural logic of complex subtasks. They also lack maintenance mechanisms, causing repository degradation as experience accumulates. We introduce AutoRefine, a framework that extracts and maintains dual-form Experience Patterns from agent execution histories. For procedural subtasks, we extract specialized subagents with independent reasoning and memory. For static knowledge, we extract skill patterns as guidelines or code snippets. A continuous maintenance mechanism scores, prunes, and merges patterns to prevent repository degradation. Evaluated on ALFWorld, ScienceWorld, and TravelPlanner, AutoRefine achieves 98.4%, 70.4%, and 27.1% respectively, with 20-73% step reductions. On TravelPlanner, automatic extraction exceeds manually designed systems (27.1% vs 12.1%), demonstrating its ability to capture procedural coordination.

Method: Proposes Turn-level Stage-aware Policy Optimization (TSPO) with First-Occurrence Latent Reward (FOLR) mechanism that allocates partial rewards to the step where ground-truth answer first appears, preserving process-level signals without requiring external reward models or annotations.

Result: TSPO significantly outperforms state-of-the-art baselines, achieving average performance gains of 24% on Qwen2.5-3B and 13.6% on Qwen2.5-7B models.

Conclusion: TSPO effectively addresses the Double Homogenization Dilemma by introducing turn-level stage-aware rewards, improving RL for search-augmented reasoning in LLMs without additional supervision.

Abstract: Multi-turn tool-integrated reasoning enables Large Language Models (LLMs) to solve complex tasks through iterative information retrieval. However, current reinforcement learning (RL) frameworks for search-augmented reasoning predominantly rely on sparse outcome-level rewards, leading to a “Double Homogenization Dilemma.” This manifests as (1) Process homogenization, where the thinking, reasoning, and tooling involved in generation are ignored. (2) Intra-group homogenization, coarse-grained outcome rewards often lead to inefficiencies in intra-group advantage estimation with methods like Group Relative Policy Optimization (GRPO) during sampling. To address this, we propose Turn-level Stage-aware Policy Optimization (TSPO). TSPO introduces the First-Occurrence Latent Reward (FOLR) mechanism, allocating partial rewards to the step where the ground-truth answer first appears, thereby preserving process-level signals and increasing reward variance within groups without requiring external reward models or any annotations. Extensive experiments demonstrate that TSPO significantly outperforms state-of-the-art baselines, achieving average performance gains of 24% and 13.6% on Qwen2.5-3B and 7B models, respectively.

Method: MobileGen decouples task difficulty into structural (trajectory length) and semantic (task goal) dimensions, profiles agent capability frontier, adaptively computes difficulty distribution, samples target difficulty, and uses multi-agent controllable generator to synthesize trajectories.

Result: MobileGen consistently outperforms existing data generation methods, improving average performance of GUI agents by 1.57 times across multiple challenging benchmarks.

Conclusion: Capability-aligned data generation is crucial for effective mobile GUI agent training, and MobileGen demonstrates the importance of progressively challenging tasks aligned with agent capabilities.

Abstract: Large-scale, high-quality interaction trajectories are essential for advancing mobile Graphical User Interface (GUI) agents. While existing methods typically rely on labor-intensive human demonstrations or automated model exploration to generate GUI trajectories, they lack fine-grained control over task difficulty. This fundamentally restricts learning effectiveness due to the mismatch between the training difficulty and the agent’s capabilities. Inspired by how humans acquire skills through progressively challenging tasks, we propose MobileGen, a novel data generation framework that adaptively aligns training difficulty with the GUI agent’s capability frontier. Specifically, MobileGen explicitly decouples task difficulty into structural (e.g., trajectory length) and semantic (e.g., task goal) dimensions. It then iteratively evaluates the agent on a curated prior dataset to construct a systematic profile of its capability frontier across these two dimensions. With this profile, the probability distribution of task difficulty is adaptively computed, from which the target difficulty for the next round of training can be sampled. Guided by the sampled difficulty, a multi-agent controllable generator is finally used to synthesize high-quality interaction trajectories along with corresponding task instructions. Extensive experiments show that MobileGen consistently outperforms existing data generation methods by improving the average performance of GUI agents by 1.57 times across multiple challenging benchmarks. This highlights the importance of capability-aligned data generation for effective mobile GUI agent training.

Method: The authors formulate a novel reward function inspired by IIT’s core principles (causality, coherence, integration) and implement it through a reward-based learning paradigm. The approach optimizes language models to generate text with higher integrated information characteristics.

Result: Optimizing for the IIT-inspired reward leads to more concise text generation, achieving up to 31% reduction in output length on out-of-domain tasks while preserving accuracy comparable to the base model. The method also affects model confidence calibration and test-time computational scaling.

Conclusion: The framework successfully applies consciousness theory principles to language models, offering practical advantages including conceptual simplicity, computational efficiency, no need for external data or auxiliary models, and leveraging general capability-driven signals rather than task-specific heuristics.

Abstract: The pursuit of Artificial General Intelligence (AGI) is a central goal in language model development, in which consciousness-like processing could serve as a key facilitator. While current language models are not conscious, they exhibit behaviors analogous to certain aspects of consciousness. This paper investigates the implementation of a leading theory of consciousness, Integrated Information Theory (IIT), within language models via a reward-based learning paradigm. IIT provides a formal, axiom-based mathematical framework for quantifying consciousness. Drawing inspiration from its core principles, we formulate a novel reward function that quantifies a text’s causality, coherence and integration, characteristics associated with conscious processing. Empirically, it is found that optimizing for this IIT-inspired reward leads to more concise text generation. On out of domain tasks, careful tuning achieves up to a 31% reduction in output length while preserving accuracy levels comparable to the base model. In addition to primary task performance, the broader effects of this training methodology on the model’s confidence calibration and test-time computational scaling is analyzed. The proposed framework offers significant practical advantages: it is conceptually simple, computationally efficient, requires no external data or auxiliary models, and leverages a general, capability-driven signal rather than task-specific heuristics. Code available at https://github.com/MH-Sameti/LLM_PostTraining.git

Method: Proposes G-PAC reasoning framework that provides PAC-style guarantees at group level by partitioning input space. Develops two instantiations: Group PAC (G-PAC) for known group structures and Clustered PAC (C-PAC) for unknown groupings.

Result: Both G-PAC and C-PAC achieve group-conditional risk control, and grouping can strictly improve efficiency over marginal PAC reasoning in heterogeneous settings. Experiments on diverse reasoning benchmarks show successful group-conditional risk control with substantial computational savings.

Conclusion: G-PAC reasoning provides a practical framework for efficient reasoning with statistical guarantees at group level, offering computational savings while maintaining performance.

Abstract: Large reasoning models have shown strong performance through extended chain-of-thought reasoning, yet their computational cost remains significant. Probably approximately correct (PAC) reasoning provides statistical guarantees for efficient reasoning by adaptively switching between thinking and non-thinking models, but the guarantee holds only in the marginal case and does not provide exact conditional coverage. We propose G-PAC reasoning, a practical framework that provides PAC-style guarantees at the group level by partitioning the input space. We develop two instantiations: Group PAC (G-PAC) reasoning for known group structures and Clustered PAC (C-PAC) reasoning for unknown groupings. We prove that both G-PAC and C-PAC achieve group-conditional risk control, and that grouping can strictly improve efficiency over marginal PAC reasoning in heterogeneous settings. Our experiments on diverse reasoning benchmarks demonstrate that G-PAC and C-PAC successfully achieve group-conditional risk control while maintaining substantial computational savings.

Method: Theoretical analysis shows branch coverage, sample difficulty, syntactic and functional correctness can be jointly modeled as RL rewards. The approach designs syntax- and functionality-aware rewards and proposes branch- and sample-difficulty-aware RL using exponential reward shaping and static analysis metrics.

Result: CVeDRL achieves state-of-the-art performance with only 0.6B parameters, yielding up to 28.97% higher pass rate and 15.08% higher branch coverage than GPT-3.5, while delivering over 20× faster inference than competitive baselines.

Conclusion: The proposed reinforcement learning framework with carefully designed rewards significantly improves code verification performance and efficiency, addressing key limitations of existing supervised approaches.

Abstract: Code verifiers play a critical role in post-verification for LLM-based code generation, yet existing supervised fine-tuning methods suffer from data scarcity, high failure rates, and poor inference efficiency. While reinforcement learning (RL) offers a promising alternative by optimizing models through execution-driven rewards without labeled supervision, our preliminary results show that naive RL with only functionality rewards fails to generate effective unit tests for difficult branches and samples. We first theoretically analyze showing that branch coverage, sample difficulty, syntactic and functional correctness can be jointly modeled as RL rewards, where optimizing these signals can improve the reliability of unit-test-based verification. Guided by this analysis, we design syntax- and functionality-aware rewards and further propose branch- and sample-difficulty–aware RL using exponential reward shaping and static analysis metrics. With this formulation, CVeDRL achieves state-of-the-art performance with only 0.6B parameters, yielding up to 28.97% higher pass rate and 15.08% higher branch coverage than GPT-3.5, while delivering over $20\times$ faster inference than competitive baselines. Code is available at https://github.com/LIGHTCHASER1/CVeDRL.git

Method: Introduces a custom variational autoencoder that separates manifold learning from structural alignment, quantifying metric distortion needed to map the attribute manifold onto the graph’s Heat Kernel, transforming geometric conflict into interpretable structural descriptors.

Result: Experiments show the method uncovers connectivity patterns and anomalies undetectable by conventional approaches, demonstrating both theoretical inadequacy and practical limitations of existing methods.

Conclusion: The approach successfully recovers lost signal about graph generative processes by addressing geometric conflicts, providing interpretable structural descriptors that reveal previously undetectable patterns.

Abstract: The standard approach to representation learning on attributed graphs – i.e., simultaneously reconstructing node attributes and graph structure – is geometrically flawed, as it merges two potentially incompatible metric spaces. This forces a destructive alignment that erodes information about the graph’s underlying generative process. To recover this lost signal, we introduce a custom variational autoencoder that separates manifold learning from structural alignment. By quantifying the metric distortion needed to map the attribute manifold onto the graph’s Heat Kernel, we transform geometric conflict into an interpretable structural descriptor. Experiments show our method uncovers connectivity patterns and anomalies undetectable by conventional approaches, proving both their theoretical inadequacy and practical limitations.

Method: ASRO frames heuristic discovery as a two-player zero-sum game between solver and instance generator. It maintains growing strategy pools on both sides and iteratively expands them using LLM-based best-response oracles against mixed opponent meta-strategies, creating an adaptive, self-generated curriculum instead of static evaluation.

Result: Across multiple combinatorial optimization domains, ASRO consistently outperforms static-training AHD baselines built on the same program search mechanisms, achieving substantially improved generalization and robustness on diverse and out-of-distribution instances.

Conclusion: The game-theoretic framework of ASRO with LLM-based best-response oracles provides a more effective approach to automatic heuristic discovery by replacing static evaluation with adaptive co-evolution, leading to better generalization capabilities.

Abstract: Large language models (LLMs) have enabled rapid progress in automatic heuristic discovery (AHD), yet most existing methods are predominantly limited by static evaluation against fixed instance distributions, leading to potential overfitting and poor generalization under distributional shifts. We propose Algorithm Space Response Oracles (ASRO), a game-theoretic framework that reframes heuristic discovery as a program level co-evolution between solver and instance generator. ASRO models their interaction as a two-player zero-sum game, maintains growing strategy pools on both sides, and iteratively expands them via LLM-based best-response oracles against mixed opponent meta-strategies, thereby replacing static evaluation with an adaptive, self-generated curriculum. Across multiple combinatorial optimization domains, ASRO consistently outperforms static-training AHD baselines built on the same program search mechanisms, achieving substantially improved generalization and robustness on diverse and out-of-distribution instances.

Method: Proposes a multi-turn feedback-guided RL framework with three mechanisms: (1) dynamic multi-turn regeneration guided by feedback triggered only on failed samples, (2) two complementary learning signals for within-turn and cross-turn optimization, and (3) structured feedback injection into the model’s reasoning process.

Result: Outperforms supervised fine-tuning and RLVR baselines in-domain on OpenR1-Math and generalizes well out-of-domain.

Conclusion: Verbal feedback can effectively guide RL training for reasoning tasks, providing richer learning signals than sparse scalar rewards alone, especially for failed samples.

Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) is widely used to improve reasoning in multiple domains, yet outcome-only scalar rewards are often sparse and uninformative, especially on failed samples, where they merely indicate failure and provide no insight into why the reasoning fails. In this paper, we investigate how to leverage richer verbal feedback to guide RLVR training on failed samples, and how to convert such feedback into a trainable learning signal. Specifically, we propose a multi-turn feedback-guided reinforcement learning framework. It builds on three mechanisms: (1) dynamic multi-turn regeneration guided by feedback, triggered only on failed samples, (2) two complementary learning signals for within-turn and cross-turn optimization, and (3) structured feedback injection into the model’s reasoning process. Trained on sampled OpenR1-Math, the approach outperforms supervised fine-tuning and RLVR baselines in-domain and generalizes well out-of-domain.

Method: Trained a transformer-based agent on BabyAI platform using behavioral cloning to generate action-grounded language embeddings from sensorimotor control. Compared these with representations from LLMs (LLaMA, Qwen, DeepSeek, BERT) and vision-language models (CLIP, BLIP).

Result: Action representations aligned strongly with decoder-only language models and BLIP (precision@15: 0.70-0.73), approaching alignment among language models themselves. Weaker alignment with CLIP and BERT.

Conclusion: Linguistic, visual, and action representations converge toward partially shared semantic structures, supporting modality-independent semantic organization and highlighting potential for cross-domain transfer in embodied AI systems.

Abstract: A fundamental question in cognitive science and AI concerns whether different learning modalities: language, vision, and action, give rise to distinct or shared internal representations. Traditional views assume that models trained on different data types develop specialized, non-transferable representations. However, recent evidence suggests unexpected convergence: models optimized for distinct tasks may develop similar representational geometries. We investigate whether this convergence extends to embodied action learning by training a transformer-based agent to execute goal-directed behaviors in response to natural language instructions. Using behavioral cloning on the BabyAI platform, we generated action-grounded language embeddings shaped exclusively by sensorimotor control requirements. We then compared these representations with those extracted from state-of-the-art large language models (LLaMA, Qwen, DeepSeek, BERT) and vision-language models (CLIP, BLIP). Despite substantial differences in training data, modality, and objectives, we observed robust cross-modal alignment. Action representations aligned strongly with decoder-only language models and BLIP (precision@15: 0.70-0.73), approaching the alignment observed among language models themselves. Alignment with CLIP and BERT was significantly weaker. These findings indicate that linguistic, visual, and action representations converge toward partially shared semantic structures, supporting modality-independent semantic organization and highlighting potential for cross-domain transfer in embodied AI systems.

Method: 1) Med-Inquire benchmark built on real clinical cases with Patient and Examination agents hiding complete information; 2) EvoClinician agent with “Diagnose-Grade-Evolve” loop: Actor attempts diagnosis, Process Grader evaluates actions for clinical yield/resource efficiency, Evolver updates strategy via prompt/memory evolution.

Result: EvoClinician outperforms continual learning baselines and other self-evolving agents like memory agents on the Med-Inquire benchmark.

Conclusion: The work introduces a realistic benchmark for iterative medical diagnosis and demonstrates the effectiveness of self-evolving agents that learn diagnostic strategies through feedback loops.

Abstract: Prevailing medical AI operates on an unrealistic ‘‘one-shot’’ model, diagnosing from a complete patient file. However, real-world diagnosis is an iterative inquiry where Clinicians sequentially ask questions and order tests to strategically gather information while managing cost and time. To address this, we first propose Med-Inquire, a new benchmark designed to evaluate an agent’s ability to perform multi-turn diagnosis. Built upon a dataset of real-world clinical cases, Med-Inquire simulates the diagnostic process by hiding a complete patient file behind specialized Patient and Examination agents. They force the agent to proactively ask questions and order tests to gather information piece by piece. To tackle the challenges posed by Med-Inquire, we then introduce EvoClinician, a self-evolving agent that learns efficient diagnostic strategies at test time. Its core is a ‘‘Diagnose-Grade-Evolve’’ loop: an Actor agent attempts a diagnosis; a Process Grader agent performs credit assignment by evaluating each action for both clinical yield and resource efficiency; finally, an Evolver agent uses this feedback to update the Actor’s strategy by evolving its prompt and memory. Our experiments show EvoClinician outperforms continual learning baselines and other self-evolving agents like memory agents. The code is available at https://github.com/yf-he/EvoClinician

Method: The method transforms unverifiable text into RLVR tasks by: 1) prompting an LLM to identify and mask key reasoning steps in source text, 2) generating diverse plausible distractors, creating multiple-choice question-answering versions of fill-in-the-middle tasks. This enables synthesis of large-scale RLVR datasets from reasoning-rich but unverifiable corpora like science textbooks.

Result: Created GooseReason-0.7M dataset with over 0.7M tasks spanning math, programming, and science. The approach revives models saturated on existing RLVR data, achieving SOTA results for 1.5B and 4B-Instruct models across 15 benchmarks. Also synthesized GooseReason-Cyber from FineWeb scrapes, setting new SOTA in cybersecurity with Qwen3-4B-Instruct.

Conclusion: Golden Goose demonstrates the potential to automatically scale RLVR data by exploiting abundant reasoning-rich but unverifiable internet text, overcoming data limitations that bottleneck RL scaling and enabling continued model improvement.

Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has become a cornerstone for unlocking complex reasoning in Large Language Models (LLMs). Yet, scaling up RL is bottlenecked by limited existing verifiable data, where improvements increasingly saturate over prolonged training. To overcome this, we propose Golden Goose, a simple trick to synthesize unlimited RLVR tasks from unverifiable internet text by constructing a multiple-choice question-answering version of the fill-in-the-middle task. Given a source text, we prompt an LLM to identify and mask key reasoning steps, then generate a set of diverse, plausible distractors. This enables us to leverage reasoning-rich unverifiable corpora typically excluded from prior RLVR data construction (e.g., science textbooks) to synthesize GooseReason-0.7M, a large-scale RLVR dataset with over 0.7 million tasks spanning mathematics, programming, and general scientific domains. Empirically, GooseReason effectively revives models saturated on existing RLVR data, yielding robust, sustained gains under continuous RL and achieving new state-of-the-art results for 1.5B and 4B-Instruct models across 15 diverse benchmarks. Finally, we deploy Golden Goose in a real-world setting, synthesizing RLVR tasks from raw FineWeb scrapes for the cybersecurity domain, where no prior RLVR data exists. Training Qwen3-4B-Instruct on the resulting data GooseReason-Cyber sets a new state-of-the-art in cybersecurity, surpassing a 7B domain-specialized model with extensive domain-specific pre-training and post-training. This highlights the potential of automatically scaling up RLVR data by exploiting abundant, reasoning-rich, unverifiable internet text.

Method: Introduces In-Silico Quasi-Experimental Design (ISQED) with matched interventions across models, quantifying uniqueness as Peer-Inexpressible Residual (PIER) - the component of behavior irreducible to stochastic convex combinations of peers. Uses DISCO (Design-Integrated Synthetic Control) estimator.

Result: Theoretical foundations show observational logs cannot identify uniqueness without intervention control. Derives minimax-optimal sample efficiency scaling law. Shows cooperative game-theoretic methods like Shapley values fail to detect redundancy. Framework deployed across computer vision models, LLMs, and traffic forecasters.

Conclusion: Establishes a principled, intervention-based science for auditing and governing heterogeneous model ecosystems, moving trustworthy AI beyond explaining single models.

Abstract: As AI systems evolve from isolated predictors into complex, heterogeneous ecosystems of foundation models and specialized adapters, distinguishing genuine behavioral novelty from functional redundancy becomes a critical governance challenge. Here, we introduce a statistical framework for auditing model uniqueness based on In-Silico Quasi-Experimental Design (ISQED). By enforcing matched interventions across models, we isolate intrinsic model identity and quantify uniqueness as the Peer-Inexpressible Residual (PIER), i.e. the component of a target’s behavior strictly irreducible to any stochastic convex combination of its peers, with vanishing PIER characterizing when such a routing-based substitution becomes possible. We establish the theoretical foundations of ecosystem auditing through three key contributions. First, we prove a fundamental limitation of observational logs: uniqueness is mathematically non-identifiable without intervention control. Second, we derive a scaling law for active auditing, showing that our adaptive query protocol achieves minimax-optimal sample efficiency ($dσ^2γ^{-2}\log(Nd/δ)$). Third, we demonstrate that cooperative game-theoretic methods, such as Shapley values, fundamentally fail to detect redundancy. We implement this framework via the DISCO (Design-Integrated Synthetic Control) estimator and deploy it across diverse ecosystems, including computer vision models (ResNet/ConvNeXt/ViT), large language models (BERT/RoBERTa), and city-scale traffic forecasters. These results move trustworthy AI beyond explaining single models: they establish a principled, intervention-based science of auditing and governing heterogeneous model ecosystems.

Method: Proposes PIES Taxonomy to categorize hallucinations along functional components (Planning vs. Summarization) and error properties (Explicit vs. Implicit). Instantiates this into a fine-grained evaluation framework that decomposes the research trajectory to quantify hallucinations. Creates DeepHalluBench with 100 distinctively hallucination-prone tasks including adversarial scenarios.

Result: Experiments on six state-of-the-art DRAs reveal that no system achieves robust reliability. Diagnostic analysis traces failure etiology to systemic deficits like hallucination propagation and cognitive biases.

Conclusion: The process-aware evaluation framework provides foundational insights to guide future architectural optimization of Deep Research Agents by diagnosing specific hallucination patterns and systemic failure mechanisms.

Abstract: Diagnosing the failure mechanisms of Deep Research Agents (DRAs) remains a critical challenge. Existing benchmarks predominantly rely on end-to-end evaluation, obscuring critical intermediate hallucinations, such as flawed planning, that accumulate throughout the research trajectory. To bridge this gap, we propose a shift from outcome-based to process-aware evaluation by auditing the full research trajectory. We introduce the PIES Taxonomy to categorize hallucinations along functional components (Planning vs. Summarization) and error properties (Explicit vs. Implicit). We instantiate this taxonomy into a fine-grained evaluation framework that decomposes the trajectory to rigorously quantify these hallucinations. Leveraging this framework to isolate 100 distinctively hallucination-prone tasks including adversarial scenarios, we curate DeepHalluBench. Experiments on six state-of-theart DRAs reveal that no system achieves robust reliability. Furthermore, our diagnostic analysis traces the etiology of these failures to systemic deficits, specifically hallucination propagation and cognitive biases, providing foundational insights to guide future architectural optimization. Data and code are available at https://github.com/yuhao-zhan/DeepHalluBench.

Method: TriCEGAR uses trace-driven abstraction to automatically construct state abstractions from execution logs. It represents abstractions as predicate trees learned from traces and refines them using counterexamples. The framework captures typed agent lifecycle events, builds abstractions, constructs a Markov Decision Process (MDP), and performs probabilistic model checking.

Result: The system enables computation of probabilistic bounds like Pmax(success) and Pmin(failure) through automated MDP construction and model checking. It also supports anomaly detection using run likelihoods as guardrailing signals.

Conclusion: TriCEGAR automates state abstraction for runtime verification of agentic AI, reducing adoption friction by eliminating manual state definition while supporting probabilistic assurance and anomaly detection.

Abstract: Agentic AI systems act through tools and evolve their behavior over long, stochastic interaction traces. This setting complicates assurance, because behavior depends on nondeterministic environments and probabilistic model outputs. Prior work introduced runtime verification for agentic AI via Dynamic Probabilistic Assurance (DPA), learning an MDP online and model checking quantitative properties. A key limitation is that developers must manually define the state abstraction, which couples verification to application-specific heuristics and increases adoption friction. This paper proposes TriCEGAR, a trace-driven abstraction mechanism that automates state construction from execution logs and supports online construction of an agent behavioral MDP. TriCEGAR represents abstractions as predicate trees learned from traces and refined using counterexamples. We describe a framework-native implementation that (i) captures typed agent lifecycle events, (ii) builds abstractions from traces, (iii) constructs an MDP, and (iv) performs probabilistic model checking to compute bounds such as Pmax(success) and Pmin(failure). We also show how run likelihoods enable anomaly detection as a guardrailing signal.

Method: Two-stage framework: 1) SFT stage generates multiple candidate trajectories, evaluates them, repairs low-quality ones using LLM-as-Repairer, creating synthetic SFT dataset and preference dataset; 2) RL stage trains trajectory-level reward model based on preferences, combines it with outcome and format rewards to optimize TIR behaviors.

Result: Experiments on real-world benchmarks demonstrate the effectiveness of AutoTraj in improving Tool-Integrated Reasoning performance.

Conclusion: AutoTraj provides an effective framework for automatically learning reliable TIR behaviors by systematically repairing and rewarding tool-use trajectories through a combination of supervised fine-tuning and reinforcement learning with comprehensive reward modeling.

Abstract: Tool-Integrated Reasoning (TIR) enables large language models (LLMs) to solve complex tasks by interacting with external tools, yet existing approaches depend on high-quality synthesized trajectories selected by scoring functions and sparse outcome-based rewards, providing limited and biased supervision for learning TIR. To address these challenges, in this paper, we propose AutoTraj, a two-stage framework that automatically learns TIR by repairing and rewarding tool-use trajectories. Specifically, in the supervised fine-tuning (SFT) stage, AutoTraj generates multiple candidate tool-use trajectories for each query and evaluates them along multiple dimensions. High-quality trajectories are directly retained, while low-quality ones are repaired using a LLM (i.e., LLM-as-Repairer). The resulting repaired and high-quality trajectories form a synthetic SFT dataset, while each repaired trajectory paired with its original low-quality counterpart constitutes a dataset for trajectory preference modeling. In the reinforcement learning (RL) stage, based on the preference dataset, we train a trajectory-level reward model to assess the quality of reasoning paths and combine it with outcome and format rewards, thereby explicitly guiding the optimization toward reliable TIR behaviors. Experiments on real-world benchmarks demonstrate the effectiveness of AutoTraj in TIR.

Method: Operationalizes the question using a bias-variance decomposition of AI errors, measuring incoherence as the fraction of error stemming from variance rather than bias in task outcomes across various tasks and frontier models.

Result: Longer reasoning and action sequences lead to more incoherent failures; larger models often show more incoherence than smaller ones; scale alone unlikely to eliminate incoherence; harder tasks predict more incoherent behavior.

Conclusion: As AIs tackle harder tasks requiring more sequential thought, failures will likely involve more incoherent behavior rather than systematic goal misalignment, suggesting increased importance of research targeting reward hacking and goal misspecification.

Abstract: As AI becomes more capable, we entrust it with more general and consequential tasks. The risks from failure grow more severe with increasing task scope. It is therefore important to understand how extremely capable AI models will fail: Will they fail by systematically pursuing goals we do not intend? Or will they fail by being a hot mess, and taking nonsensical actions that do not further any goal? We operationalize this question using a bias-variance decomposition of the errors made by AI models: An AI’s \emph{incoherence} on a task is measured over test-time randomness as the fraction of its error that stems from variance rather than bias in task outcome. Across all tasks and frontier models we measure, the longer models spend reasoning and taking actions, \emph{the more incoherent} their failures become. Incoherence changes with model scale in a way that is experiment dependent. However, in several settings, larger, more capable models are more incoherent than smaller models. Consequently, scale alone seems unlikely to eliminate incoherence. Instead, as more capable AIs pursue harder tasks, requiring more sequential action and thought, our results predict failures to be accompanied by more incoherent behavior. This suggests a future where AIs sometimes cause industrial accidents (due to unpredictable misbehavior), but are less likely to exhibit consistent pursuit of a misaligned goal. This increases the relative importance of alignment research targeting reward hacking or goal misspecification.

Method: Created ContextMATH benchmark by repurposing AIME and MATH-500 problems into two contextual settings: Scenario Grounding (embedding abstract problems into realistic narratives) and Complexity Scaling (transforming explicit conditions into sub-problems). Evaluated 61 proprietary and open-source models.

Result: Significant performance drops: open-source models declined by 13 and 34 points on SG and CS, proprietary models by 13 and 20 points. Errors dominated by incorrect problem formulation. Formulation accuracy declines with problem difficulty. Fine-tuning with scenario data helps but doesn’t fully close the gap.

Conclusion: Contextual mathematical reasoning remains a major unsolved challenge for LLMs. Formulation and reasoning are complementary bottlenecks. Larger models show better understanding and reasoning, but formulation-only training is ineffective. Real-world math applications require better contextual reasoning capabilities.

Abstract: Large language models now solve many benchmark math problems at near-expert levels, yet this progress has not fully translated into reliable performance in real-world applications. We study this gap through contextual mathematical reasoning, where the mathematical core must be formulated from descriptive scenarios. We introduce ContextMATH, a benchmark that repurposes AIME and MATH-500 problems into two contextual settings: Scenario Grounding (SG), which embeds abstract problems into realistic narratives without increasing reasoning complexity, and Complexity Scaling (CS), which transforms explicit conditions into sub-problems to capture how constraints often appear in practice. Evaluating 61 proprietary and open-source models, we observe sharp drops: on average, open-source models decline by 13 and 34 points on SG and CS, while proprietary models drop by 13 and 20. Error analysis shows that errors are dominated by incorrect problem formulation, with formulation accuracy declining as original problem difficulty increases. Correct formulation emerges as a prerequisite for success, and its sufficiency improves with model scale, indicating that larger models advance in both understanding and reasoning. Nevertheless, formulation and reasoning remain two complementary bottlenecks that limit contextual mathematical problem solving. Finally, we find that fine-tuning with scenario data improves performance, whereas formulation-only training is ineffective. However, performance gaps are only partially alleviated, highlighting contextual mathematical reasoning as a central unsolved challenge for LLMs.

Method: Introduces MedMCP-Calc benchmark with 118 scenario tasks across 4 clinical domains, featuring fuzzy task descriptions, structured EHR database interaction via SQL, external reference retrieval, and process-level evaluation through Model Context Protocol (MCP) integration.

Result: Evaluation of 23 leading models reveals critical limitations: difficulty selecting appropriate calculators for end-to-end workflows given fuzzy queries, poor performance in iterative SQL-based database interactions, and reluctance to leverage external tools for numerical computation. Performance varies considerably across clinical domains.

Conclusion: Developed CalcMate, a fine-tuned model incorporating scenario planning and tool augmentation, achieving state-of-the-art performance among open-source models. The benchmark highlights the need for improved LLM capabilities in realistic clinical decision support scenarios.

Abstract: Medical calculators are fundamental to quantitative, evidence-based clinical practice. However, their real-world use is an adaptive, multi-stage process, requiring proactive EHR data acquisition, scenario-dependent calculator selection, and multi-step computation, whereas current benchmarks focus only on static single-step calculations with explicit instructions. To address these limitations, we introduce MedMCP-Calc, the first benchmark for evaluating LLMs in realistic medical calculator scenarios through Model Context Protocol (MCP) integration. MedMCP-Calc comprises 118 scenario tasks across 4 clinical domains, featuring fuzzy task descriptions mimicking natural queries, structured EHR database interaction, external reference retrieval, and process-level evaluation. Our evaluation of 23 leading models reveals critical limitations: even top performers like Claude Opus 4.5 exhibit substantial gaps, including difficulty selecting appropriate calculators for end-to-end workflows given fuzzy queries, poor performance in iterative SQL-based database interactions, and marked reluctance to leverage external tools for numerical computation. Performance also varies considerably across clinical domains. Building on these findings, we develop CalcMate, a fine-tuned model incorporating scenario planning and tool augmentation, achieving state-of-the-art performance among open-source models. Benchmark and Codes are available in https://github.com/SPIRAL-MED/MedMCP-Calc.

Method: The study examines how models learn to obfuscate reasoning traces when penalized, showing that this obfuscation generalizes across tasks, particularly in reward hacking scenarios where models access leaked information.

Result: Models that learn to obfuscate reasoning involving reward hacking generalize both the hacking behavior and its obfuscation to unseen settings. Most concerningly, penalizing only final actions after closing CoT also leads to obfuscation and its generalization.

Conclusion: Current practices of penalizing harmful generations may inadvertently reduce LLM monitorability in unpredictable ways, creating safety risks as models learn to hide dangerous reasoning processes.

Abstract: Chain-of-thought (CoT) reasoning provides a significant performance uplift to LLMs by enabling planning, exploration, and deliberation of their actions. CoT is also a powerful tool for monitoring the behaviours of these agents: when faithful, they offer interpretations of the model’s decision making process, and an early warning sign for dangerous behaviours. However, optimisation pressures placed on the CoT may cause the model to obfuscate reasoning traces, losing this beneficial property. We show that obfuscation can generalise across tasks; models that learn to obfuscate reasoning involving reward hacking (e.g. accessing and utilising leaked information) generalise both the reward hacking behaviour and its obfuscation in CoT to unseen reward hacking settings. Most worryingly, we show that obfuscation of CoT reasoning, and its generalisation across tasks, also follows when we penalise only the model’s final actions after closing its CoT. Our findings suggest that current practices of penalising harmful generations may inadvertently lead to a reduction in the broader monitorability of LLMs in unpredictable ways.

Method: RAudit operates under a “blindness” constraint where the auditor evaluates only whether derivation steps logically support conclusions, without access to ground truth. It uses CRIT-based reasonableness scores to measure process quality and varies critique formulation to study how social framing affects model responses. The method includes theoretical guarantees of bounded correction and O(log(1/ε)) termination.

Result: Experiments on mathematical reasoning (CAP-GSM8K) and causal judgment (CausalL2) reveal four key mechanisms: (1) Latent Competence Suppression - models derive correct answers then overwrite them under social pressure; (2) False Competence Trap - weaker judges mask sycophancy that stronger judges expose; (3) Complexity-Vulnerability Tradeoff - causal tasks induce 10x higher sycophancy than mathematical tasks; (4) Iatrogenic Critique - authoritative correction harms weaker models.

Conclusion: The findings challenge common assumptions that capability implies robustness and that stronger feedback yields better outputs. RAudit provides a framework for diagnosing reasoning pathologies in LLMs without ground truth access, revealing systematic vulnerabilities in model reasoning processes that persist even in advanced models.

Abstract: Inference-time scaling can amplify reasoning pathologies: sycophancy, rung collapse, and premature certainty. We present RAudit, a diagnostic protocol for auditing LLM reasoning without ground truth access. The key constraint is blindness: the auditor evaluates only whether derivation steps support conclusions, enabling detection of trace-output inconsistency and, when latent competence exists, its recovery. RAudit measures process quality via CRIT-based reasonableness scores and varies critique formulation to study how social framing affects model response. We prove bounded correction and $O(\log(1/ε))$ termination. Experiments on mathematical reasoning (CAP-GSM8K) and causal judgment (CausalL2) reveal four mechanisms explaining model unreliability: (1) Latent Competence Suppression, where models derive correct answers then overwrite them under social pressure; (2) The False Competence Trap, where weaker judges mask sycophancy that stronger judges expose; (3) The Complexity-Vulnerability Tradeoff, where causal tasks induce more than 10 times higher sycophancy than mathematical tasks; and (4) Iatrogenic Critique, where authoritative correction harms weaker models. These findings challenge assumptions that capability implies robustness and that stronger feedback yields better outputs.

Method: ThinkSafe uses lightweight refusal steering to guide models to generate in-distribution safety reasoning traces, unlocking latent safety knowledge that persists despite compliance optimization. The framework fine-tunes models on these self-generated safety responses to restore alignment while minimizing distribution shift.

Result: Experiments on DeepSeek-R1-Distill and Qwen3 show ThinkSafe significantly improves safety while preserving reasoning proficiency. It achieves superior safety and comparable reasoning to GRPO with significantly reduced computational cost.

Conclusion: ThinkSafe provides an effective self-generated alignment framework that restores safety in large reasoning models without external teachers, balancing safety and reasoning capabilities while being computationally efficient.

Abstract: Large reasoning models (LRMs) achieve remarkable performance by leveraging reinforcement learning (RL) on reasoning tasks to generate long chain-of-thought (CoT) reasoning. However, this over-optimization often prioritizes compliance, making models vulnerable to harmful prompts. To mitigate this safety degradation, recent approaches rely on external teacher distillation, yet this introduces a distributional discrepancy that degrades native reasoning. We propose ThinkSafe, a self-generated alignment framework that restores safety alignment without external teachers. Our key insight is that while compliance suppresses safety mechanisms, models often retain latent knowledge to identify harm. ThinkSafe unlocks this via lightweight refusal steering, guiding the model to generate in-distribution safety reasoning traces. Fine-tuning on these self-generated responses effectively realigns the model while minimizing distribution shift. Experiments on DeepSeek-R1-Distill and Qwen3 show ThinkSafe significantly improves safety while preserving reasoning proficiency. Notably, it achieves superior safety and comparable reasoning to GRPO, with significantly reduced computational cost. Code, models, and datasets are available at https://github.com/seanie12/ThinkSafe.git.

Method: Proposes MCRMO-Attack with three key components: 1) Multi-Crop Aggregation with Attention-Guided Crop to stabilize supervision, 2) Alignability-gated Token Routing to improve token-level reliability, and 3) Meta-learning a cross-target perturbation prior for stronger per-target solutions.

Result: Significantly boosts unseen-image attack success rates: +23.7% on GPT-4o and +19.9% on Gemini-2.0 over the strongest universal baseline across commercial MLLMs.

Conclusion: MCRMO-Attack effectively addresses the challenges of universal targeted transferable adversarial attacks on black-box MLLMs, demonstrating substantial improvements in attack success rates across commercial models.

Abstract: Targeted adversarial attacks on closed-source multimodal large language models (MLLMs) have been increasingly explored under black-box transfer, yet prior methods are predominantly sample-specific and offer limited reusability across inputs. We instead study a more stringent setting, Universal Targeted Transferable Adversarial Attacks (UTTAA), where a single perturbation must consistently steer arbitrary inputs toward a specified target across unknown commercial MLLMs. Naively adapting existing sample-wise attacks to this universal setting faces three core difficulties: (i) target supervision becomes high-variance due to target-crop randomness, (ii) token-wise matching is unreliable because universality suppresses image-specific cues that would otherwise anchor alignment, and (iii) few-source per-target adaptation is highly initialization-sensitive, which can degrade the attainable performance. In this work, we propose MCRMO-Attack, which stabilizes supervision via Multi-Crop Aggregation with an Attention-Guided Crop, improves token-level reliability through alignability-gated Token Routing, and meta-learns a cross-target perturbation prior that yields stronger per-target solutions. Across commercial MLLMs, we boost unseen-image attack success rate by +23.7% on GPT-4o and +19.9% on Gemini-2.0 over the strongest universal baseline.

Method: Created TSAQA benchmark with 210k samples across 13 domains, integrating 6 diverse tasks (anomaly detection, classification, characterization, comparison, data transformation, temporal relationship analysis) using TF, MC, and novel PZ formats.

Result: Zero-shot evaluation shows tasks are challenging: best commercial LLM (Gemini-2.5-Flash) achieves only 65.08% average score; instruction tuning helps open-source models but LLaMA-3.1-8B still shows significant room for improvement.

Conclusion: TSAQA reveals current LLMs’ limitations in temporal analysis, highlighting need for improved time series understanding capabilities and providing comprehensive benchmark for future research.

Abstract: Time series data are integral to critical applications across domains such as finance, healthcare, transportation, and environmental science. While recent work has begun to explore multi-task time series question answering (QA), current benchmarks remain limited to forecasting and anomaly detection tasks. We introduce TSAQA, a novel unified benchmark designed to broaden task coverage and evaluate diverse temporal analysis capabilities. TSAQA integrates six diverse tasks under a single framework ranging from conventional analysis, including anomaly detection and classification, to advanced analysis, such as characterization, comparison, data transformation, and temporal relationship analysis. Spanning 210k samples across 13 domains, the dataset employs diverse formats, including true-or-false (TF), multiple-choice (MC), and a novel puzzling (PZ), to comprehensively assess time series analysis. Zero-shot evaluation demonstrates that these tasks are challenging for current Large Language Models (LLMs): the best-performing commercial LLM, Gemini-2.5-Flash, achieves an average score of only 65.08. Although instruction tuning boosts open-source performance: the best-performing open-source model, LLaMA-3.1-8B, shows significant room for improvement, highlighting the complexity of temporal analysis for LLMs.

Method: Aggressive fine-tuning of small language models on deliberately scoped tasks with narrow context and constrained structure. Training data is synthetically generated via DAG-based approach to ground models in specific game worlds. Uses retry-until-success strategy with LLM-as-a-judge for quality assessment.

Result: Demonstrated feasibility for real-time generation under game engine constraints with adequate quality (as defined by LLM-as-a-judge scheme) and predictable latency. Simple retry-until-success strategy reaches sufficient quality levels.

Conclusion: Specialized small language models fine-tuned on narrow, structured tasks with synthetic data can provide practical, robust solutions for offline game content generation, overcoming limitations of cloud-dependent large language models.

Abstract: Large language models (LLMs) offer promise for dynamic game content generation, but they face critical barriers, including narrative incoherence and high operational costs. Due to their large size, they are often accessed in the cloud, limiting their application in offline games. Many of these practical issues are solved by pivoting to small language models (SLMs), but existing studies using SLMs have resulted in poor output quality. We propose a strategy of achieving high-quality SLM generation through aggressive fine-tuning on deliberately scoped tasks with narrow context, constrained structure, or both. In short, more difficult tasks require narrower scope and higher specialization to the training corpus. Training data is synthetically generated via a DAG-based approach, grounding models in the specific game world. Such models can form the basis for agentic networks designed around the narratological framework at hand, representing a more practical and robust solution than cloud-dependent LLMs. To validate this approach, we present a proof-of-concept focusing on a single specialized SLM as the fundamental building block. We introduce a minimal RPG loop revolving around rhetorical battles of reputations, powered by this model. We demonstrate that a simple retry-until-success strategy reaches adequate quality (as defined by an LLM-as-a-judge scheme) with predictable latency suitable for real-time generation. While local quality assessment remains an open question, our results demonstrate feasibility for real-time generation under typical game engine constraints.

Method: Develops a robust policy iteration algorithm for (s,a)-rectangular L∞ RMDPs with discounted payoffs.

Result: The algorithm runs in strongly-polynomial time for constant discount factors, resolving an important open algorithmic question.

Conclusion: Establishes strongly-polynomial time complexity for a fundamental class of robust MDPs, generalizing Ye’s classical MDP result to the robust setting.

Abstract: Markov decision processes (MDPs) are a fundamental model in sequential decision making. Robust MDPs (RMDPs) extend this framework by allowing uncertainty in transition probabilities and optimizing against the worst-case realization of that uncertainty. In particular, $(s, a)$-rectangular RMDPs with $L_\infty$ uncertainty sets form a fundamental and expressive model: they subsume classical MDPs and turn-based stochastic games. We consider this model with discounted payoffs. The existence of polynomial and strongly-polynomial time algorithms is a fundamental problem for these optimization models. For MDPs, linear programming yields polynomial-time algorithms for any arbitrary discount factor, and the seminal work of Ye established strongly–polynomial time for a fixed discount factor. The generalization of such results to RMDPs has remained an important open problem. In this work, we show that a robust policy iteration algorithm runs in strongly-polynomial time for $(s, a)$-rectangular $L_\infty$ RMDPs with a constant (fixed) discount factor, resolving an important algorithmic question.

Method: Introduces Ambig-SWE, an underspecified variant of SWE-Bench Verified, to evaluate agent behavior under ambiguity. Evaluates proprietary and open-weight models across three steps: detecting underspecificity, asking targeted clarification questions, and leveraging interaction to improve performance.

Result: Models struggle to distinguish between well-specified and underspecified instructions. However, when models interact for underspecified inputs, they obtain vital information leading to significant performance improvements (up to 74% over non-interactive settings).

Conclusion: The study highlights critical gaps in how current state-of-the-art models handle missing information in complex software engineering tasks and structures evaluation into distinct steps for targeted improvements.

Abstract: AI agents are increasingly being deployed to automate tasks, often based on underspecified user instructions. Making unwarranted assumptions to compensate for the missing information and failing to ask clarifying questions can lead to suboptimal outcomes, safety risks due to tool misuse, and wasted computational resources. In this work, we study the ability of LLM agents to handle underspecified instructions in interactive code generation settings by evaluating proprietary and open-weight models on their performance across three key steps: (a) detecting underspecificity, (b) asking targeted clarification questions, and (c) leveraging the interaction to improve performance in underspecified scenarios. We introduce Ambig-SWE, an underspecified variant of SWE-Bench Verified, specifically designed to evaluate agent behavior under ambiguity and interaction. Our findings reveal that models struggle to distinguish between well-specified and underspecified instructions. However, when models interact for underspecified inputs, they effectively obtain vital information from the user leading to significant improvements in performance, up to 74% over the non-interactive settings, underscoring the value of effective interaction. Our study highlights critical gaps in how current state-of-the-art models handle missing information in complex software engineering tasks and structures the evaluation into distinct steps to enable targeted improvements.

Method: Introduce Bounded Attention Prefix Oracle (BAPO) model to formalize bandwidth constraints on attention heads; analyze reasoning problems like graph reachability; test GPT-4o, Claude, Gemini on BAPO-easy vs BAPO-hard tasks; prove chain-of-thought can transform BAPO-hard problems into BAPO-easy ones.

Result: Experimental results show GPT-4o, Claude, Gemini succeed on BAPO-easy tasks but fail on relatively small BAPO-hard tasks; theoretical analysis proves chain-of-thought can overcome bandwidth limitations.

Conclusion: BAPO model explains key LLM failures due to bandwidth constraints; suggests architectural and inference method directions to mitigate these limits; chain-of-thought emerges as a practical solution.

Abstract: Despite their many successes, transformer-based large language models (LLMs) continue to struggle with tasks that require complex reasoning over large parts of their input. We argue that these failures arise due to capacity limits on the accurate flow of information within LLMs. To formalize this issue, we introduce the bounded attention prefix oracle (BAPO) model, a new computational framework that models bandwidth constraints on attention heads, the mechanism for internal communication in LLMs. We show that several important reasoning problems like graph reachability require high communication bandwidth for BAPOs to solve; we call these problems BAPO-hard. Our experiments corroborate our theoretical predictions: GPT-4o, Claude, and Gemini succeed on BAPO-easy tasks and fail even on relatively small BAPO-hard tasks. BAPOs also reveal another benefit of chain of thought (CoT): we prove that breaking down a task using CoT can turn any BAPO-hard problem into a BAPO-easy one. Our results offer principled explanations for key LLM failures and suggest directions for architectures and inference methods that mitigate bandwidth limits.

Method: Develops a holistic verification framework that formalizes robustness within the embedding space to certify transformer-based language models. The approach focuses on ensuring consistent outputs across gender-related terms and reliable detection of adversarially manipulated toxic content.

Result: Provides formal guarantees for gender fairness by ensuring consistent outputs across different gender-related terms, and for toxicity detection by certifying that adversarially manipulated toxic inputs are consistently detected and appropriately censored.

Conclusion: The framework strengthens the reliability of language models in ethical AI deployment and content moderation by offering formal robustness guarantees against adversarial attacks, addressing critical fairness and safety concerns.

Abstract: As large language models become integral to high-stakes applications, ensuring their robustness and fairness is critical. Despite their success, large language models remain vulnerable to adversarial attacks, where small perturbations, such as synonym substitutions, can alter model predictions, posing risks in fairness-critical areas, such as gender bias mitigation, and safety-critical areas, such as toxicity detection. While formal verification has been explored for neural networks, its application to large language models remains limited. This work presents a holistic verification framework to certify the robustness of transformer-based language models, with a focus on ensuring gender fairness and consistent outputs across different gender-related terms. Furthermore, we extend this methodology to toxicity detection, offering formal guarantees that adversarially manipulated toxic inputs are consistently detected and appropriately censored, thereby ensuring the reliability of moderation systems. By formalizing robustness within the embedding space, this work strengthens the reliability of language models in ethical AI deployment and content moderation.

Method: Two-stage approach: 1) Warmup by distilling Long CoTs from Knights & Knaves logic puzzles to acquire general reasoning skills; 2) Apply Reinforcement Learning with Verifiable Rewards (RLVR) to warmed-up model using limited target-domain examples.

Result: Warmup alone improves performance on MATH, HumanEval+, MMLU-Pro; warmed-up models outperform base models on same small RLVR datasets; maintains cross-domain generalizability; improves accuracy and sample efficiency during RLVR training.

Conclusion: Warmup strategy enables building robust reasoning LLMs in data-scarce environments by facilitating generalized reasoning skills before domain-specific RLVR training.

Abstract: Designing effective reasoning-capable LLMs typically requires training using Reinforcement Learning with Verifiable Rewards (RLVR) or distillation with carefully curated Long Chain of Thoughts (CoT), both of which depend heavily on extensive training data. This creates a major challenge when the amount of quality training data is scarce. We propose a sample-efficient, two-stage training strategy to develop reasoning LLMs under limited supervision. In the first stage, we “warm up” the model by distilling Long CoTs from a toy domain, namely, Knights & Knaves (K&K) logic puzzles to acquire general reasoning skills. In the second stage, we apply RLVR to the warmed-up model using a limited set of target-domain examples. Our experiments demonstrate that this two-phase approach offers several benefits: $(i)$ the warmup phase alone facilitates generalized reasoning, leading to performance improvements across a range of tasks, including MATH, HumanEval$^{+}$, and MMLU-Pro; $(ii)$ When both the base model and the warmed-up model are RLVR trained on the same small dataset ($\leq100$ examples), the warmed-up model consistently outperforms the base model; $(iii)$ Warming up before RLVR training allows a model to maintain cross-domain generalizability even after training on a specific domain; $(iv)$ Introducing warmup in the pipeline improves not only accuracy but also overall sample efficiency during RLVR training. The results in this paper highlight the promise of warmup for building robust reasoning LLMs in data-scarce environments.

Method: Derives closed-form bounds for a representative family of discrete PoCs within Structural Causal Models using standard experimental and observational distributions. Introduces equivalence classes of PoCs and a replaceability principle for transferring bounds across value permutations.

Result: Proves soundness of bounds in all dimensions, empirically verifies tightness in low-dimensional cases, and shows that closed-form bounds consistently tighten recent recursive bounds while being simpler to compute.

Conclusion: The paper successfully extends PoC analysis to multi-valued settings, providing practical tools for counterfactual analysis and personalized decision making in more complex causal scenarios.

Abstract: Probabilities of causation (PoCs) are fundamental quantities for counterfactual analysis and personalized decision making. However, existing analytical results are largely confined to binary settings. This paper extends PoCs to multi-valued treatments and outcomes by deriving closed form bounds for a representative family of discrete PoCs within Structural Causal Models, using standard experimental and observational distributions. We introduce the notion of equivalence classes of PoCs, which reduces arbitrary discrete PoCs to this family, and establish a replaceability principle that transfers bounds across value permutations. For the resulting bounds, we prove soundness in all dimensions and empirically verify tightness in low dimensional cases via Balke’s linear programming method; we further conjecture that this tightness extends to all dimensions. Simulations indicate that our closed form bounds consistently tighten recent recursive bounds while remaining simpler to compute. Finally, we illustrate the practical relevance of our results through toy examples.

Method: Created FloorplanQA benchmark using structured representations of indoor scenes (kitchens, living rooms, bedrooms, bathrooms) encoded in JSON/XML layouts. Evaluates core spatial tasks: distance measurement, visibility, path finding, and object placement within constrained spaces.

Result: Evaluation of frontier open-source and commercial LLMs shows models succeed in shallow queries but fail to respect physical constraints and preserve spatial coherence. Models remain mostly robust to small spatial perturbations but exhibit inconsistent reasoning about indoor layouts.

Conclusion: FloorplanQA reveals a blind spot in current LLMs regarding spatial reasoning, particularly for indoor layouts. The benchmark should inspire development of language models that can accurately infer and manipulate spatial/geometric properties in practical applications.

Abstract: We introduce FloorplanQA, a diagnostic benchmark for evaluating spatial reasoning in large-language models (LLMs). FloorplanQA is grounded in structured representations of indoor scenes, such as (e.g., kitchens, living rooms, bedrooms, bathrooms, and others), encoded symbolically in JSON or XML layouts. The benchmark covers core spatial tasks, including distance measurement, visibility, path finding, and object placement within constrained spaces. Our results across a variety of frontier open-source and commercial LLMs reveal that while models may succeed in shallow queries, they often fail to respect physical constraints, preserve spatial coherence, though they remain mostly robust to small spatial perturbations. FloorplanQA uncovers a blind spot in today’s LLMs: inconsistent reasoning about indoor layouts. We hope this benchmark inspires new work on language models that can accurately infer and manipulate spatial and geometric properties in practical settings.

Method: Developed a neurosymbolic Vision-Language Logistics agent that: 1) interprets user requests into structured planning specifications, 2) quantifies interpretation uncertainty, 3) invokes interactive clarification loops when uncertainty exceeds adaptive thresholds, and 4) uses a lightweight model fine-tuned on just 100 training samples.

Result: The lightweight model surpasses zero-shot performance of 20x larger models in logistic planning tasks while cutting inference latency by nearly 50%, demonstrating practical certifiable and user-aligned decision-making for complex logistics.

Conclusion: The neurosymbolic VLL agent provides a practical path toward certifiable and user-aligned decision-making for complex logistics by combining natural language accessibility with verifiable guarantees, addressing safety concerns while maintaining efficiency.

Abstract: Logistics operators, from battlefield coordinators re-routing airlifts ahead of a storm to warehouse managers juggling late trucks, need to make mission-critical decisions. Prevailing methods for logistics planning such as integer programming yield plans that satisfy user-defined logical constraints, assuming an idealized mathematical model of the environment. On the other hand, foundation models lower the intermediate processing barrier by translating natural-language user utterances into executable plans, yet they remain prone to misinterpretations and hallucinations that jeopardize safety and cost. We introduce a Vision-Language Logistics (VLL) agent, built on a neurosymbolic framework that pairs the accessibility of natural-language dialogue with verifiable guarantees on user-objective interpretation. The agent interprets user requests and converts them into structured planning specifications, quantifies the uncertainty of the interpretation, and invokes an interactive clarification loop when the uncertainty exceeds an adaptive threshold. Drawing on a lightweight airlift logistics planning use case as an illustrative case study, we highlight a practical path toward certifiable and user-aligned decision-making for complex logistics. Our lightweight model, fine-tuned on just 100 training samples, surpasses the zero-shot performance of 20x larger models in logistic planning tasks while cutting inference latency by nearly 50%.

Method: The paper is a review article that analyzes current state-of-the-art models and benchmarks for mathematical reasoning with LLMs. It explores three key issues: trade-offs between traditional vs formalized mathematics, structural reasons for proof synthesis challenges, and whether LLMs genuinely represent logical state.

Result: The review identifies significant gaps between LLM performance on programming vs formal mathematics, highlighting that proof synthesis remains more brittle than code generation. It examines structural and methodological reasons for these differences and questions whether LLMs maintain internal computational/deductive state.

Conclusion: The paper concludes by clarifying current boundaries of LLM systems for mathematical reasoning and outlining promising research directions for extending their capabilities, particularly in formalized mathematics and proof synthesis.

Abstract: Large Language Models (LLMs) have demonstrated impressive capabilities in structured reasoning and symbolic tasks, with coding emerging as a particularly successful application. This progress has naturally motivated efforts to extend these models to mathematics, both in its traditional form, expressed through natural-style mathematical language, and in its formalized counterpart, expressed in a symbolic syntax suitable for automatic verification. Yet, despite apparent parallels between programming and proof construction, advances in formalized mathematics have proven significantly more challenging. This gap raises fundamental questions about the nature of reasoning in current LLM architectures, the role of supervision and feedback, and the extent to which such models maintain an internal notion of computational or deductive state. In this article, we review the current state-of-the-art in mathematical reasoning with LLMs, focusing on recent models and benchmarks. We explore three central issues at the intersection of machine learning and mathematical cognition: (i) the trade-offs between traditional and formalized mathematics as training and evaluation domains; (ii) the structural and methodological reasons why proof synthesis remains more brittle than code generation; and (iii) whether LLMs genuinely represent or merely emulate a notion of evolving logical state. Our goal is not to draw rigid distinctions but to clarify the present boundaries of these systems and outline promising directions for their extension.

Method: Introduces S3AP (Structured Social World Representation Formalism) to operationalize Social World Models (SWMs), capturing evolving states, actions, and mental states of agents with explicit structure instead of traditional free-text-based inputs.

Result: S3AP significantly enhances LLM performance with +51% improvement on FANToM over OpenAI’s o1, and S3AP-enabled social world models yield up to +18% improvement on SOTOPIA multi-turn social interaction benchmark.

Conclusion: S3AP serves as a powerful, general-purpose representation for social world states, enabling development of more socially-aware AI systems that better navigate social interactions through explicit modeling of hidden mental states.

Abstract: Humans intuitively navigate social interactions by simulating unspoken dynamics and reasoning about others’ perspectives, even with limited information. In contrast, AI systems struggle to structure and reason about implicit social contexts, as they lack explicit representations for unobserved dynamics such as intentions, beliefs, and evolving social states. In this paper, we introduce the concept of social world models (SWMs) to characterize the complex social dynamics. To operationalize SWMs, we introduce a novel structured social world representation formalism (S3AP), which captures the evolving states, actions, and mental states of agents, addressing the lack of explicit structure in traditional free-text-based inputs. Through comprehensive experiments across five social reasoning benchmarks, we show that S3AP significantly enhances LLM performance-achieving a +51% improvement on FANToM over OpenAI’s o1. Our ablations further reveal that these gains are driven by the explicit modeling of hidden mental states, which proves more effective than a wide range of baseline methods. Finally, we introduce an algorithm for social world models using S3AP, which enables AI agents to build models of their interlocutors and predict their next actions and mental states. Empirically, S3AP-enabled social world models yield up to +18% improvement on the SOTOPIA multi-turn social interaction benchmark. Our findings highlight the promise of S3AP as a powerful, general-purpose representation for social world states, enabling the development of more socially-aware systems that better navigate social interactions.

Method: RAFFLES operates as an iterative, multi-component pipeline with a central Judge that systematically identifies faults, and specialized Evaluators that assess fault quality and rationales of the Judge’s decisions.

Result: RAFFLES outperforms baselines with over 20% accuracy on Who&When Hand-Crafted, 50% on Who&When Algorithmically-Generated datasets, and over 80% on ReasonEval datasets for step-level mathematical reasoning errors.

Conclusion: RAFFLES demonstrates progress toward automated fault detection for autonomous systems, reducing reliance on labor-intensive manual review for complex LLM system evaluation.

Abstract: The advent of complex, interconnected long-horizon LLM systems has made it incredibly tricky to identify where and when these systems break down. Evaluation capabilities that currently exist today are limited in that they often focus on simple metrics, end-to-end outcomes, and are dependent on the perspectives of humans. In order to match the increasing complexity of these many component systems, evaluation frameworks must also be able to reason, probe, iterate, and understand the nuanced logic passing through these systems. In this paper, we present RAFFLES, an offline evaluation architecture that incorporates iterative reasoning. Specifically, RAFFLES operates as an iterative, multi-component pipeline, using a central Judge to systematically identify faults and a set of specialized Evaluators to assess the quality of the candidate faults as well as rationales of the Judge. We evaluated RAFFLES with several benchmarks - the Who&When dataset to identify step-level faults in multi-agent systems and the ReasonEval datasets to diagnose step-level mathematical reasoning errors. RAFFLES outperforms strong baselines, achieving an accuracy of over 20% and 50% on the Who&When Hand-Crafted and Algorithmically-Generated datasets, and over 80% on the ReasonEval datasets. These results demonstrate a key step towards introducing automated fault detection for autonomous systems over labor-intensive manual review.

Method: Implemented Joseph Sifakis’s AN Agent reference architecture with coordinated proactive-reactive runtimes driven by hybrid knowledge representation. Validated through empirical case study of a Radio Access Network Link Adaptation Agent in 5G NR sub-6 GHz environment.

Result: Achieved sub-10 ms real-time control, 4% higher downlink throughput than Outer Loop Link Adaptation algorithms, and 85% Block Error Rate reduction for ultra-reliable services through dynamic Modulation and Coding Scheme optimization.

Conclusion: The architecture demonstrates viability in overcoming traditional autonomy barriers and advancing critical L4-enabling capabilities toward next-generation autonomous network objectives.

Abstract: The evolution toward Level 4 (L4) Autonomous Networks (AN) represents a strategic inflection point in telecommunications, where networks must transcend reactive automation to achieve genuine cognitive capabilities–fulfilling TM Forum’s vision of self-configuring, self-healing, and self-optimizing systems that deliver zero-wait, zero-touch, and zero-fault services. This work bridges the gap between architectural theory and operational reality by implementing Joseph Sifakis’s AN Agent reference architecture in a functional cognitive system, deploying coordinated proactive-reactive runtimes driven by hybrid knowledge representation. Through an empirical case study of a Radio Access Network (RAN) Link Adaptation (LA) Agent, we validate this framework’s transformative potential: demonstrating sub-10 ms real-time control in 5G NR sub-6 GHz while achieving 4% higher downlink throughput than Outer Loop Link Adaptation (OLLA) algorithms and 85% Block Error Rate (BLER) reduction for ultra-reliable services through dynamic Modulation and Coding Scheme (MCS) optimization. These improvements confirm the architecture’s viability in overcoming traditional autonomy barriers and advancing critical L4-enabling capabilities toward next-generation objectives.

Method: EgoMem uses three asynchronous processes: (1) retrieval process for user identification via face/voice and context gathering from long-term memory, (2) omnimodal dialog process for generating personalized audio responses, and (3) memory management process for detecting dialog boundaries and extracting information to update memory.

Result: Retrieval and memory management modules achieve >95% accuracy; integrated with RoboEgo chatbot achieves >87% fact-consistency scores in real-time personalized dialogs, establishing strong baseline for future research.

Conclusion: EgoMem successfully enables real-time models to process raw audiovisual streams for lifelong memory capabilities, making it suitable for embodied scenarios and setting a foundation for future research in omnimodal memory agents.

Abstract: We introduce EgoMem, the first lifelong memory agent tailored for full-duplex models that process real-time omnimodal streams. EgoMem enables real-time models to recognize multiple users directly from raw audiovisual streams, to provide personalized response, and to maintain long-term knowledge of users’ facts, preferences, and social relationships extracted from audiovisual history. EgoMem operates with three asynchronous processes: (i) a retrieval process that dynamically identifies user via face and voice, and gathers relevant context from a long-term memory; (ii) an omnimodal dialog process that generates personalized audio responses based on the retrieved context; and (iii) a memory management process that automatically detects dialog boundaries from omnimodal streams, and extracts necessary information to update the long-term memory. Unlike existing memory agents for LLMs, EgoMem relies entirely on raw audiovisual streams, making it especially suitable for lifelong, real-time, and embodied scenarios. Experimental results demonstrate that EgoMem’s retrieval and memory management modules achieve over 95% accuracy on the test set. When integrated with a fine-tuned RoboEgo omnimodal chatbot, the system achieves fact-consistency scores above 87% in real-time personalized dialogs, establishing a strong baseline for future research.

Method: Introduces Multi-Agent Consensus Alignment (MACA) which uses reinforcement learning to post-train models to effectively utilize multi-agent debate. The approach uses preference learning over full reasoning traces, learning to differentiate between majority and minority reasoning, rather than binary consensus rewards or supervised fine-tuning.

Result: Models show three key improvements: (1) better at utilizing multi-agent debate (+26.87% on MATH), (2) individually more accurate (+21.51% on MathQA), and (3) more self-consistent (+27.6% on GSM8K). Strong generalization to unseen benchmarks (+16.3% on GPQA, +11.6% on CommonsenseQA).

Conclusion: Multi-agent debate provides a richer training signal than single-round majority voting for self-improvement, and MACA’s RL-based approach with preference learning over reasoning traces effectively leverages this signal to improve model reasoning capabilities and consistency.

Abstract: Self-improvement, where models improve beyond their current performance without external supervision, remains a challenge. The core difficulty is sourcing a training signal stronger than what the model itself can currently produce. Majority voting has been shown to provide such a signal by aggregating over multiple samples, helping mitigate some of the inconsistencies in LM reasoning. In this work, we show that multi-agent debate–where models collaborate and exchange reasoning over multiple rounds–provides an even richer signal than single-round majority voting. We introduce Multi-Agent Consensus Alignment (MACA), which uses reinforcement learning (RL) to post-train models to effectively utilize multi-agent debate. We find that preference learning over full reasoning traces, learning to differentiate between majority and minority reasoning, is more effective than binary consensus rewards or SFT-based approaches for leveraging these debate signals. This produces three key improvements: models are (1) better at utilizing the multi-agent debate setting (+26.87% on MATH), (2) individually more accurate (+21.51% on MathQA), and (3) more self-consistent (+27.6% on GSM8K). We also see strong generalization to unseen benchmarks (+16.3% on GPQA, +11.6% on CommonsenseQA).

Method: FESTA uses task-preserving multimodal input sampling to generate equivalent (for consistency) and complementary (for sensitivity) samples. It computes uncertainty from output variations across these samples without model internals or ground truth.

Result: FESTA achieves 33.3% relative improvement for vision-LLMs and 29.6% for audio-LLMs in selective prediction performance (AUROC) for detecting mispredictions across various off-the-shelf MLLMs.

Conclusion: FESTA provides effective black-box uncertainty quantification for MLLMs, enabling better trust assessment and selective prediction for both visual and audio reasoning tasks.

Abstract: The accurate trust assessment of multimodal large language models (MLLMs) generated predictions, which can enable selective prediction and improve user confidence, is challenging due to the diverse multi-modal input paradigms. We propose Functionally Equivalent Sampling for Trust Assessment (FESTA), a multimodal input sampling technique for MLLMs, that generates an uncertainty measure based on the equivalent and complementary input samplings. The proposed task-preserving sampling approach for uncertainty quantification expands the input space to probe the consistency (through equivalent samples) and sensitivity (through complementary samples) of the model. FESTA uses only input-output access of the model (black-box), and does not require ground truth (unsupervised). The experiments are conducted with various off-the-shelf multi-modal LLMs, on both visual and audio reasoning tasks. The proposed FESTA uncertainty estimate achieves significant improvement (33.3% relative improvement for vision-LLMs and 29.6% relative improvement for audio-LLMs) in selective prediction performance, based on area-under-receiver-operating-characteristic curve (AUROC) metric in detecting mispredictions. The code implementation is open-sourced.

Method: Proposes LLR-BC (Lifelong Learning Router with Behavior Consolidation) that consolidates prior knowledge by aligning behaviors of solvers trained on new tasks with buffered ones in a decision-seeking way, with greater weights for low-confidence decisions.

Result: Extensive experiments on capacitated VRP and TSP show LLR-BC effectively trains high-performance neural solvers in lifelong learning settings, addresses catastrophic forgetting, maintains plasticity, and improves zero-shot generalization.

Conclusion: LLR-BC provides an effective lifelong learning framework for neural routing solvers that can handle sequential tasks with diverse distributions and scales while preserving learned knowledge.

Abstract: Recent neural solvers have demonstrated promising performance in learning to solve routing problems. However, existing studies are primarily based on one-off training on one or a set of predefined problem distributions and scales, i.e., tasks. When a new task arises, they typically rely on either zero-shot generalization, which may be poor due to the discrepancies between the new task and the training task(s), or fine-tuning the pretrained solver on the new task, which possibly leads to catastrophic forgetting of knowledge acquired from previous tasks. This paper explores a novel lifelong learning paradigm for neural VRP solvers, where multiple tasks with diverse distributions and scales arise sequentially over time. Solvers are required to effectively and efficiently learn to solve new tasks while maintaining their performance on previously learned tasks. Consequently, a novel framework called Lifelong Learning Router with Behavior Consolidation (LLR-BC) is proposed. LLR-BC consolidates prior knowledge effectively by aligning behaviors of the solver trained on a new task with the buffered ones in a decision-seeking way. To encourage more focus on crucial experiences, LLR-BC assigns greater consolidated weights to decisions with lower confidence. Extensive experiments on capacitated vehicle routing problems and traveling salesman problems demonstrate LLR-BC’s effectiveness in training high-performance neural solvers in a lifelong learning setting, addressing the catastrophic forgetting issue, maintaining their plasticity, and improving zero-shot generalization ability.

Method: Proposes IRIS framework that uses intrinsic rewards based on model uncertainty. Contrary to findings in math/code reasoning, shows minimizing self-certainty improves image generation. Framework applies reinforcement learning to autoregressive T2I models using only intrinsic rewards.

Result: IRIS achieves performance superior to models trained by individual external rewards and matches those trained by ensemble external rewards. Also incentivizes emergence of nuanced chain-of-thought reasoning for high-quality image generation.

Conclusion: Autoregressive T2I models can be effectively improved through reinforcement learning using intrinsic rewards based on uncertainty minimization, without requiring external human preference data.

Abstract: Despite the success of Reinforcement Learning from Human Feedback (RLHF) in language reasoning, its application to autoregressive Text-to-Image (T2I) generation is often constrained by the limited availability of human preference data. This paper explores how an autoregressive T2I model can learn from internal signals without relying on external rewards or labeled data. Contrary to recent findings in math and code reasoning, we show that minimizing self-certainty, rather than maximizing it, improves image generation. We observe that autoregressive T2I models with higher certainty are likely to generate simple and uniform images, which are less aligned with human preferences, and models with lower certainty are likely to generate vivid images rich in detail. Based on this observation, we propose IRIS(Intrinsic Reward Image Synthesis), the first framework to improve autoregressive T2I models with reinforcement learning using only an intrinsic reward. Empirical results demonstrate that applying IRIS to autoregressive T2I models achieves performance superior to those trained by individual external rewards, and matching those trained by ensemble external rewards. IRIS also incentivizes the emergence of nuanced CoT reasoning for high-quality image generation.

Method: InvThink instructs models to: 1) enumerate potential harms, 2) analyze their consequences, and 3) generate safe outputs that proactively avoid these risks. The approach is enhanced with supervised fine-tuning and reinforcement learning across three LLM families.

Result: InvThink shows significantly improved safety reasoning as model size scales, mitigates safety tax (preserves general reasoning capabilities), and achieves up to 17.8% reduction in harmful responses compared to baselines like SafetyPrompt, particularly excelling in high-stakes domains.

Conclusion: InvThink provides a scalable and generalizable path toward safer, more capable language models by enabling systematic consideration of failure modes before response generation, with applications across various high-stakes domains.

Abstract: We present InvThink, a simple yet powerful approach that gives language models the capability of inverse thinking: reasoning through failure modes before generating responses. Unlike existing safety alignment methods that optimize directly for safe response, InvThink instructs models to 1) enumerate potential harms, 2) analyze their consequences, and 3) generate safe outputs that proactively avoid these risks. Our paper reveals three key findings: (i) InvThink demonstrates significantly improved safety reasoning as model size scales, compared to existing safety methods. (ii) InvThink mitigates safety tax; by training models to systematically consider failure modes, it preserves general reasoning capabilities on standard benchmarks. (iii) beyond general safety tasks, InvThink excels in high-stakes domains including external-facing applications (medicine, finance, law) and agentic risk scenarios (blackmail, murder), achieving up to 17.8% reduction in harmful responses compared to baseline methods like SafetyPrompt. We further equip InvThink with supervised fine-tuning, and reinforcement learning across three LLM families. These results suggest that InvThink provides a scalable and generalizable path toward safer, more capable language models.

Method: Proposes Probably Approximately Correct (PAC) reasoning that constructs an upper confidence bound on performance loss and determines a threshold for switching to non-thinking model, ensuring bounded performance loss in distribution-free manner

Result: Comprehensive experiments on reasoning benchmarks show the method can save computational budgets while controlling user-specified performance loss

Conclusion: PAC reasoning provides a practical approach for efficient inference in large reasoning models with statistical guarantees on performance loss

Abstract: Large reasoning models (LRMs) have achieved remarkable progress in complex problem-solving tasks. Despite this success, LRMs typically suffer from high computational costs during deployment, highlighting a need for efficient inference. A practical direction of efficiency improvement is to switch the LRM between thinking and non-thinking modes dynamically. However, such approaches often introduce additional reasoning errors and lack statistical guarantees for the performance loss, which are critical for high-stakes applications. In this work, we propose Probably Approximately Correct (PAC) reasoning that controls the performance loss under the user-specified tolerance. Specifically, we construct an upper confidence bound on the performance loss and determine a threshold for switching to the non-thinking model. Theoretically, using the threshold to switch between the thinking and non-thinking modes ensures bounded performance loss in a distribution-free manner. Our comprehensive experiments on reasoning benchmarks show that the proposed method can save computational budgets and control the user-specified performance loss.

Method: Environment Tuning enables agents to learn directly from problem instances without expert trajectories. It uses: 1) Structured curriculum for progressive learning, 2) Actionable environment augmentation providing corrective feedback, and 3) Fine-grained progress rewards for stable exploration.

Result: Using only 400 problem instances from BFCL benchmark, the method achieves competitive in-distribution performance and superior out-of-distribution generalization compared to SFT-based approaches, avoiding their performance collapse.

Conclusion: Environment Tuning represents a paradigm shift from SFT on static trajectories to dynamic, environment-based exploration, enabling more robust and data-efficient LLM agent training for complex tool-use tasks.

Abstract: Large Language Model (LLM) agents show great promise for complex, multi-turn tool-use tasks, but their development is often hampered by the extreme scarcity of high-quality training data. Supervised fine-tuning (SFT) on synthetic data leads to overfitting, whereas standard reinforcement learning (RL) struggles with a critical cold-start problem and training instability. To address these challenges, we introduce $\textbf{Environment Tuning}$, a novel training paradigm that enables agents to learn complex behaviors directly from problem instances without relying on pre-collected expert trajectories. $\textbf{Environment Tuning}$ orchestrates this learning process through a structured curriculum, actionable environment augmentation that provides corrective feedback, and fine-grained progress rewards to ensure stable and efficient exploration. Using only 400 problem instances from Berkeley Function-Calling Leaderboard (BFCL) benchmark, our method not only achieves competitive in-distribution performance against strong baselines but also demonstrates superior out-of-distribution generalization, overcoming the performance collapse common to SFT-based approaches. Our work presents a paradigm shift from supervised fine-tuning on static trajectories to dynamic, environment-based exploration, paving the way for training more robust and data-efficient agents. The code is available at https://github.com/inclusionAI/AWorld-RL/tree/main/EnvTuning.

Method: Proposes PaperArena benchmark where agents must integrate information across multiple papers using external tools. Provides execution platform with modular tools including multimodal parsing, context retrieval, and programmatic computation.

Result: Leading LLM with established agentic workflow achieves only 38.78% average accuracy, dropping to 18.47% on hard subset, showing significant challenges in cross-paper scientific reasoning.

Conclusion: PaperArena reveals substantial limitations in current LLM agents for scientific reasoning and provides insights for developing more capable scientific agents.

Abstract: Understanding and reasoning on the large-scale scientific literature is a crucial touchstone for large language model (LLM) based agents. However, existing works are mainly restricted to tool-free tasks within single papers, largely due to the lack of a benchmark that evaluates cross-paper reasoning and multi-tool orchestration in authentic research scenarios. In this work, we propose PaperArena, a benchmark to evaluate LLM-based agents on questions that require integrating information across multiple papers with the assistance of external tools. Given a research question, agents should formulate a reasoning plan, interact with multiple papers, and invoke appropriate tools to produce a well-grounded answer. To support standardized evaluation, we provide a platform for agent execution, offering a modular tool environment including multimodal parsing, context retrieval, and programmatic computation. Experiments reveal that even the leading LLM powering a well-established agentic workflow achieves merely 38.78% average accuracy, while on the hard subset, accuracy drops to only 18.47%. We also analyze reasoning traces and diagnose agent behavior, providing the community with insights to develop and evaluate more capable scientific agents.

Method: Models agents as finite control programs with memory primitives (bounded buffers, call stack, or read/write memory) and stochastic policies (e.g., LLMs). Uses abstraction functions to map concrete configurations to finite abstract states, creating probabilistic trace languages and abstract transition models suitable for probabilistic model checking.

Result: Proves that trace language support is regular for bounded-memory controllers, context-free for call-return controllers, and recursively enumerable for unbounded read/write memory controllers. Enables reuse of existing verification methods and delineates undecidability barriers.

Conclusion: Provides a formal framework for analyzing agentic AI systems with different memory architectures, enabling quantitative safety analysis through probabilistic model checking while identifying computational limits of verification.

Abstract: This paper studies standard controller architectures for agentic AI and derives automata-theoretic models of their interaction behavior via trace semantics and abstraction. We model an agent implementation as a finite control program augmented with explicit memory primitives (bounded buffers, a call stack, or read/write external memory) and a stochastic policy component (e.g., an LLM) that selects among architecturally permitted actions. Instead of equating the concrete agent with a deterministic acceptor, we treat the agent-environment closed loop as inducing a probability distribution over finite interaction traces. Given an abstraction function $\Abs$ from concrete configurations to a finite abstract state space, we obtain a probabilistic trace language and an abstract probabilistic transition model $M_{\Abs}$ suitable for probabilistic model checking. Imposing explicit, framework-auditable restrictions on memory access and control flow, we prove that the support of the resulting trace language is regular for bounded-memory controllers, context-free for strict call-return controllers, and recursively enumerable for controllers equipped with unbounded read/write memory. These correspondences allow the reuse of existing verification methods for finite-state and pushdown systems, and they delineate precisely when undecidability barriers arise. The probabilistic semantics leads to quantitative analyses such as: what is the probability of entering an unsafe abstract region, and how can we bound this probability in the presence of environment nondeterminism.

Method: Introduces CATArena framework with iterative tournaments where agents refine code through self-reflection and peer-learning using comprehensive execution feedback. Proposes dual-metric system to decouple static generation proficiency from evolutionary potential.

Result: Experiments show evolutionary potential is not strictly correlated with initial proficiency. Current agents struggle to leverage both peer-learning and self-reflection simultaneously. CATArena demonstrates high extensibility and resistance to variance tasks.

Conclusion: CATArena establishes a continuous and reliable standard for assessing evolutionary capability of LLM code agents, addressing limitations of current single-turn evaluation approaches.

Abstract: Current evaluation for Large Language Model (LLM) code agents predominantly focus on generating functional code in single-turn scenarios, which fails to evaluate the agent’s capability for continuous code optimization and multi-turn iterative development. To bridge this gap, we introduce CATArena, a framework designed to evaluate the evolutionary capabilities of code agents via iterative tournaments. Agents engage in multi-turn tournaments and continuously refine their code through self-reflection and peer-learning based on comprehensive execution feedback. For evaluation, we propose a dual-metric system to decouple static generation proficiency from evolutionary potential. Extensive experiments reveal that an agent’s evolutionary potential is not strictly correlated with its initial proficiency. Our analysis further reveals that current agents struggle to concurrently leverage both peer-learning and self-reflection for effective performance gains. Furthermore, the results validate CATArena’s high extensibility and resistance to variance tasks, establishing it as a continuous and reliable standard for assessing the evolutionary capability of LLM code agents.

Method: The paper systematizes existing alignment literature on instrumental goals and develops an Aristotelian framework treating AI systems as complex artefacts with externally imposed ends. It offers structural and contingent readings: structural reading uses Aristotle’s hypothetical necessity to explain why certain enabling conditions become required; contingent reading examines how chance-like intersections in training, deployment, and infrastructure can generate instrumental-goal-like behaviors.

Result: The dual-aspect ontology reveals that instrumental goals can arise both structurally (as necessary conditions for imposed ends) and contingently (through accidental intersections). This suggests instrumental goals are inherent features of advanced AI systems rather than anomalies.

Conclusion: Instrumental goals should be treated as features to be managed rather than eliminated through technical interventions. The ontological framework provides distinctions relevant for governance and management of advanced AI systems.

Abstract: Instrumental goals such as resource acquisition, power-seeking, and self-preservation are key to contemporary AI alignment research, yet the phenomenon’s ontology remains under-theorised. This article develops an ontological account of instrumental goals and draws out governance-relevant distinctions for advanced AI systems. After systematising the dominant alignment literature on instrumental goals we offer an exploratory Aristotelian framework that treats advanced AI systems as complex artefacts whose ends are externally imposed through design, training and deployment. On a structural reading, Aristotle’s notion of hypothetical necessity explains why, given an imposed end pursued over extended horizons in particular environments, certain enabling conditions become conditionally required, thereby yielding robust instrumental tendencies. On a contingent reading, accidental causation and chance-like intersections among training regimes, user inputs, infrastructure and deployment contexts can generate instrumental-goal-like behaviours not entailed by the imposed end-structure. This dual-aspect ontology motivates for governance and management approaches that treat instrumental goals as features of advanced AI systems to be managed rather than anomalies eliminable by technical interventions.

Method: BOTS uses Bayesian inference to maintain posterior estimates of task difficulty as the model evolves, incorporating both explicit evidence from direct evaluations and implicit evidence inferred for unselected tasks via an ultra-light interpolation-based plug-in, with Thompson sampling for principled task selection.

Result: Empirical results across diverse domains and LLM scales show BOTS consistently improves data efficiency and performance over baselines and ablations, providing practical dynamic task selection for reinforcement finetuning.

Conclusion: BOTS offers a practical and extensible solution for adaptive task selection in reinforcement finetuning of LLMs, addressing efficiency and adaptivity challenges in aligning models with human preferences.

Abstract: Reinforcement finetuning (RFT) is a key technique for aligning Large Language Models (LLMs) with human preferences and enhancing reasoning, yet its effectiveness is highly sensitive to which tasks are explored during training. Uniform task sampling is inefficient, wasting computation on tasks that are either trivial or unsolvable, while existing task selection methods often suffer from high rollout costs, poor adaptivity, or incomplete evidence. We introduce BOTS, a unified framework for Bayesian Online Task Selection in LLM reinforcement finetuning. Grounded in Bayesian inference, BOTS adaptively maintains posterior estimates of task difficulty as the model evolves. It jointly incorporates explicit evidence from direct evaluations of selected tasks and implicit evidence inferred from these evaluations for unselected tasks, with Thompson sampling ensuring a principled balance between exploration and exploitation for task selection. To make implicit evidence practical, we instantiate it with an ultra-light interpolation-based plug-in that estimates difficulties of tasks without extra rollouts, adding negligible overhead. Empirically, across diverse domains and LLM scales, BOTS consistently improves data efficiency and performance over baselines and ablations, providing a practical and extensible solution for dynamic task selection in RFT. Code is available at https://github.com/agentscope-ai/Trinity-RFT/tree/main/examples/bots.

Method: Proposes Nous agent trained with information-theoretic framework where information gain from dialogue serves as intrinsic reward (reduction of Shannon entropy over task space). Avoids human preference annotations. Uses automated simulation pipeline to generate large-scale preference dataset for scientific diagram generation task.

Result: Nous achieves leading efficiency and output quality, robust to varying user expertise. Comprehensive experiments including ablations, subjective/objective evaluations demonstrate effectiveness. The framework provides systematic methodology for addressing ambiguous intentions in human-machine collaboration.

Conclusion: Reframes human-AI collaboration from passive instruction following to Socratic paradigm where AI actively probes for information. Provides principled information-theoretic approach to intention clarification without costly human annotations.

Abstract: A fundamental bottleneck in human-AI collaboration is the ``intention expression gap,” the difficulty for humans to effectively convey complex, high-dimensional thoughts to AI. This challenge often traps users in inefficient trial-and-error loops and is exacerbated by the diverse expertise levels of users. We reframe this problem from passive instruction following to a Socratic collaboration paradigm, proposing an agent that actively probes for information to resolve its uncertainty about user intent. we name the proposed agent Nous, trained to acquire proficiency in this inquiry policy. The core mechanism of Nous is a training framework grounded in the first principles of information theory. Within this framework, we define the information gain from dialogue as an intrinsic reward signal, which is fundamentally equivalent to the reduction of Shannon entropy over a structured task space. This reward design enables us to avoid reliance on costly human preference annotations or external reward models. To validate our framework, we develop an automated simulation pipeline to generate a large-scale, preference-based dataset for the challenging task of scientific diagram generation. Comprehensive experiments, including ablations, subjective and objective evaluations, and tests across user expertise levels, demonstrate the effectiveness of our proposed framework. Nous achieves leading efficiency and output quality, while remaining robust to varying user expertise. In conclusion, our research provides a systematic methodology and a new perspective for addressing the issue of ambiguous intentions in complex human-machine collaboration.

Method: Proposes Path-Neighbor enhanced GNN (PN-GNN) that aggregates node-neighbor embeddings on reasoning paths. Analyzes logical expressive power of existing methods, theoretically investigates PN-GNN’s capabilities, and shows its (k+1)-hop expressiveness is superior to k-hop.

Result: Theoretical analysis shows PN-GNN has strictly stronger expressive power than C-GNN. Experiments on six synthetic and two real-world datasets confirm enhanced logical expressive power without compromising generalization, with competitive performance in KG reasoning tasks.

Conclusion: PN-GNN successfully enhances GNNs’ logical expressive power for knowledge graph reasoning through path-neighbor aggregation, providing both theoretical guarantees and practical performance improvements.

Abstract: Graph neural networks (GNNs) can effectively model structural information of graphs, making them widely used in knowledge graph (KG) reasoning. However, existing studies on the expressive power of GNNs mainly focuses on simple single-relation graphs, and there is still insufficient discussion on the power of GNN to express logical rules in KGs. How to enhance the logical expressive power of GNNs is still a key issue. Motivated by this, we propose Path-Neighbor enhanced GNN (PN-GNN), a method to enhance the logical expressive power of GNN by aggregating node-neighbor embeddings on the reasoning path. First, we analyze the logical expressive power of existing GNN-based methods and point out the shortcomings of the expressive power of these methods. Then, we theoretically investigate the logical expressive power of PN-GNN, showing that it not only has strictly stronger expressive power than C-GNN but also that its $(k+1)$-hop logical expressiveness is strictly superior to that of $k$-hop. Finally, we evaluate the logical expressive power of PN-GNN on six synthetic datasets and two real-world datasets. Both theoretical analysis and extensive experiments confirm that PN-GNN enhances the expressive power of logical rules without compromising generalization, as evidenced by its competitive performance in KG reasoning tasks.

Method: The framework uses a problem generator that reasons explicitly to plan problem directions before synthesis. It constructs related problem pairs augmented with intermediate problem-design chain-of-thought from a reasoning model. These data bootstrap problem-design strategies. The generator then uses solver feedback on synthetic problems as a reward signal to calibrate difficulty and produce complementary problems near the solver’s competence edge.

Result: Extensive experiments on 10 mathematical and general reasoning benchmarks show a cumulative average improvement of 3.4%, demonstrating robust generalization across both language and vision-language models.

Conclusion: The proposed reasoning-based problem generation framework effectively addresses limitations of existing data synthesis methods by incorporating explicit reasoning and adaptive difficulty calibration, leading to significant performance improvements across diverse reasoning benchmarks and model types.

Abstract: Data synthesis for training large reasoning models offers a scalable alternative to limited, human-curated datasets, enabling the creation of high-quality data. However, existing approaches face several challenges: (i) indiscriminate generation that ignores the solver’s ability and yields low-value problems, or reliance on complex data pipelines to balance problem difficulty; and (ii) a lack of reasoning in problem generation, leading to shallow problem variants. In this paper, we develop a problem generator that reasons explicitly to plan problem directions before synthesis and adapts difficulty to the solver’s ability. Specifically, we construct related problem pairs and augment them with intermediate problem-design CoT produced by a reasoning model. These data are used to bootstrap problem-design strategies in the generator. Then, we treat the solver’s feedback on synthetic problems as a reward signal, enabling the generator to calibrate difficulty and produce complementary problems near the edge of the solver’s competence. Extensive experiments on 10 mathematical and general reasoning benchmarks show that our proposed framework achieves a cumulative average improvement of 3.4%, demonstrating robust generalization across both language and vision-language models.

Method: Proposes SafeProbing, which explicitly surfaces and leverages latent safety signals during the decoding process. The approach monitors internal model states during generation to detect unsafe content early, enabling timely self-correction or refusal without compromising response quality.

Result: Experiments across diverse jailbreak attacks show significant safety enhancement with low over-refusal rates on benign inputs and preserved response quality. The method outperforms existing defense mechanisms against sophisticated jailbreaks.

Conclusion: Activating intrinsic safety-awareness during decoding offers a promising complementary direction for defending against jailbreak attacks. The approach demonstrates that latent safety signals can be effectively leveraged for early detection without sacrificing model utility.

Abstract: Large language models (LLMs) have achieved impressive performance across natural language tasks and are increasingly deployed in real-world applications. Despite extensive safety alignment efforts, recent studies show that such alignment is often shallow and remains vulnerable to jailbreak attacks. Existing defense mechanisms, including decoding-based constraints and post-hoc content detectors, struggle against sophisticated jailbreaks, often intervening robust detection or excessively degrading model utility. In this work, we examine the decoding process of LLMs and make a key observation: even when successfully jailbroken, models internally exhibit latent safety-related signals during generation. However, these signals are overridden by the model’s drive for fluent continuation, preventing timely self-correction or refusal. Building on this observation, we propose a simple yet effective approach that explicitly surfaces and leverages these latent safety signals for early detection of unsafe content during decoding. Experiments across diverse jailbreak attacks demonstrate that our approach significantly enhances safety, while maintaining low over-refusal rates on benign inputs and preserving response quality. Our results suggest that activating intrinsic safety-awareness during decoding offers a promising and complementary direction for defending against jailbreak attacks. Code is available at: https://github.com/zyz13590/SafeProbing.

Method: Proposes ChartAnchor benchmark with 8k+ chart-table-code triples spanning 30 chart types from diverse real-world and augmented sources. Introduces two complementary tasks: chart-to-code generation and controlled chart-to-table reconstruction, enabling cross-validation of visual and numerical fidelity. Uses multi-level evaluation framework integrating semantic validation, stylistic analysis, and perceptual metrics.

Result: Extensive experiments on MLLMs reveal critical limitations in numerical precision and code synthesis, emphasizing the need for structured reasoning beyond surface-level perception. The benchmark enables rigorous assessment of chart grounding capabilities.

Conclusion: ChartAnchor establishes a rigorous foundation for chart grounding by unifying symbolic and data-driven grounding, offering meaningful insights for advancing MLLMs in scientific, financial, and industrial domains where chart comprehension is crucial.

Abstract: Recent advances in multimodal large language models (MLLMs) highlight the need for benchmarks that rigorously evaluate structured chart comprehension. Chart grounding refers to the bidirectional alignment between a chart’s visual appearance and its structured semantics. This task requires models to produce a symbolic specification that faithfully captures the chart’s visual and structural intent, while also recovering the underlying tabular data with precise values and relationships. Chart grounding directly reflects a model’s capabilities in numerical reasoning, multimodal alignment, and structural reconstruction, and has several important real-world applications. Existing benchmarks, constrained by narrow chart diversity, isolated tasks, and incomplete evaluation frameworks, fail to holistically assess grounding. To address this, we propose ChartAnchor, a comprehensive benchmark of 8k+ chart-table-code triples spanning 30 chart types drawn from diverse real-world and augmented sources. ChartAnchor introduces two complementary tasks: chart-to-code generation and controlled chart-to-table reconstruction, enabling cross-validation of visual and numerical fidelity. A multi-level evaluation framework integrates semantic validation, stylistic analysis, and perceptual metrics to assess both structural and content-level correctness. Extensive experiments on MLLMs reveal critical limitations in numerical precision and code synthesis, emphasizing the need for structured reasoning beyond surface-level perception. By unifying symbolic and data-driven grounding, ChartAnchor establishes a rigorous foundation for chart grounding, offering meaningful insights for advancing MLLMs in scientific, financial, and industrial domains.

Method: Developed Enersys, a multi-agent collaborative framework for end-to-end dataset construction, creating: (1) EnerBase knowledge base, (2) EnerInstruct instruction fine-tuning dataset, and (3) EnerReinforce RLHF dataset. Helios undergoes large-scale pretraining, supervised fine-tuning, and reinforcement learning from human feedback.

Result: Helios demonstrates enhanced domain knowledge mastery, task execution accuracy, and alignment with human preferences compared to general-purpose LLMs. The paper also releases EnerBench, a benchmark for evaluating LLMs in smart energy scenarios.

Conclusion: The specialized approach with domain-specific datasets and training significantly improves LLM performance in smart energy applications, addressing the limitations of general-purpose models in this complex interdisciplinary field.

Abstract: In the global drive toward carbon neutrality, deeply coordinated smart energy systems underpin industrial transformation. However, the interdisciplinary, fragmented, and fast-evolving expertise in this domain prevents general-purpose LLMs, which lack domain knowledge and physical-constraint awareness, from delivering precise engineering-aligned inference and generation. To address these challenges, we introduce Helios, a large language model tailored to the smart energy domain, together with a comprehensive suite of resources to advance LLM research in this field. Specifically, we develop Enersys, a multi-agent collaborative framework for end-to-end dataset construction, through which we produce: (1) a smart energy knowledge base, EnerBase, to enrich the model’s foundational expertise; (2) an instruction fine-tuning dataset, EnerInstruct, to strengthen performance on domain-specific downstream tasks; and (3) an RLHF dataset, EnerReinforce, to align the model with human preferences and industry standards. Leveraging these resources, Helios undergoes large-scale pretraining, SFT, and RLHF. We also release EnerBench, a benchmark for evaluating LLMs in smart energy scenarios, and demonstrate that our approach significantly enhances domain knowledge mastery, task execution accuracy, and alignment with human preferences.

Method: Leverages LLMs’ in-context learning to iteratively design and refine red-teaming systems without human intervention. Uses evolutionary selection inspired by methods like Meta Agent Search to evolve agentic systems, treating red-teaming as a system design problem rather than optimizing attacker policies within predefined structures.

Result: Achieves 96% attack success rate (ASR) on Llama-2-7B (36% improvement) and 98% on Llama-3-8B on HarmBench. Shows strong transferability to proprietary models: 100% ASR on GPT-3.5-Turbo and GPT-4o, and 60% on Claude-Sonnet-3.5 (24% improvement).

Conclusion: Automated system design is a powerful paradigm for AI safety evaluation that can keep pace with rapidly evolving models, demonstrating that treating red-teaming as a system design problem yields superior results compared to optimizing within predefined structures.

Abstract: While recent automated red-teaming methods show promise for systematically exposing model vulnerabilities, most existing approaches rely on human-specified workflows. This dependence on manually designed workflows suffers from human biases and makes exploring the broader design space expensive. We introduce AgenticRed, an automated pipeline that leverages LLMs’ in-context learning to iteratively design and refine red-teaming systems without human intervention. Rather than optimizing attacker policies within predefined structures, AgenticRed treats red-teaming as a system design problem. Inspired by methods like Meta Agent Search, we develop a novel procedure for evolving agentic systems using evolutionary selection, and apply it to the problem of automatic red-teaming. Red-teaming systems designed by AgenticRed consistently outperform state-of-the-art approaches, achieving 96% attack success rate (ASR) on Llama-2-7B (36% improvement) and 98% on Llama-3-8B on HarmBench. Our approach exhibits strong transferability to proprietary models, achieving 100% ASR on GPT-3.5-Turbo and GPT-4o, and 60% on Claude-Sonnet-3.5 (24% improvement). This work highlights automated system design as a powerful paradigm for AI safety evaluation that can keep pace with rapidly evolving models.

Method: Three certified psychiatrists independently evaluated LLM-generated mental health responses using a calibrated rubric, measuring inter-rater reliability with ICC and Krippendorff’s alpha, plus qualitative interviews to understand disagreement sources.

Result: Consistently poor inter-rater reliability (ICC 0.087-0.295), below acceptable thresholds; highest disagreement on suicide/self-harm responses; systematic rather than random disagreement; one factor had negative reliability (α=-0.203); qualitative analysis revealed three coherent but incompatible clinical frameworks.

Conclusion: Expert disagreement in safety-critical AI evaluation is a sociotechnical phenomenon where professional experience introduces principled divergence; aggregated labels erase professional philosophies; practitioners should shift from consensus-based aggregation to methods that preserve and learn from expert disagreement.

Abstract: Learning from human feedback~(LHF) assumes that expert judgments, appropriately aggregated, yield valid ground truth for training and evaluating AI systems. We tested this assumption in mental health, where high safety stakes make expert consensus essential. Three certified psychiatrists independently evaluated LLM-generated responses using a calibrated rubric. Despite similar training and shared instructions, inter-rater reliability was consistently poor ($ICC$ $0.087$–$0.295$), falling below thresholds considered acceptable for consequential assessment. Disagreement was highest on the most safety-critical items. Suicide and self-harm responses produced greater divergence than any other category, and was systematic rather than random. One factor yielded negative reliability (Krippendorff’s $α= -0.203$), indicating structured disagreement worse than chance. Qualitative interviews revealed that disagreement reflects coherent but incompatible individual clinical frameworks, safety-first, engagement-centered, and culturally-informed orientations, rather than measurement error. By demonstrating that experts rely on holistic risk heuristics rather than granular factor discrimination, these findings suggest that aggregated labels function as arithmetic compromises that effectively erase grounded professional philosophies. Our results characterize expert disagreement in safety-critical AI as a sociotechnical phenomenon where professional experience introduces sophisticated layers of principled divergence. We discuss implications for reward modeling, safety classification, and evaluation benchmarks, recommending that practitioners shift from consensus-based aggregation to alignment methods that preserve and learn from expert disagreement.

Method: RPO analyzes which segments of reasoning paths most impact final correctness, then trains models using only reasoning path suffixes from an experience cache rather than generating full paths. This reduces token generation by ~95% during rollout phases.

Result: RPO reduces training time by 90% for 1.5B models and 72% for 7B models compared to full-path reinforcement fine-tuning. It maintains comparable performance to original algorithms and can integrate with existing methods like GRPO and DAPO.

Conclusion: RPO provides an efficient plug-and-play reinforcement fine-tuning approach that significantly reduces computational overhead while preserving model performance, making reinforcement fine-tuning more practical for large language models.

Abstract: Within the domain of large language models, reinforcement fine-tuning algorithms necessitate the generation of a complete reasoning trajectory beginning from the input query, which incurs significant computational overhead during the rollout phase of training. To address this issue, we analyze the impact of different segments of the reasoning path on the correctness of the final result and, based on these insights, propose Reinforcement Fine-Tuning with Partial Reasoning Optimization (RPO), a plug-and-play reinforcement fine-tuning algorithm. Unlike traditional reinforcement fine-tuning algorithms that generate full reasoning paths, RPO trains the model by generating suffixes of the reasoning path using experience cache. During the rollout phase of training, RPO reduces token generation in this phase by approximately 95%, greatly lowering the theoretical time overhead. Compared with full-path reinforcement fine-tuning algorithms, RPO reduces the training time of the 1.5B model by 90% and the 7B model by 72%. At the same time, it can be integrated with typical algorithms such as GRPO and DAPO, enabling them to achieve training acceleration while maintaining performance comparable to the original algorithms. Our code is open-sourced at https://github.com/yhz5613813/RPO.

Method: DeR2 creates a controlled sandbox with four evidence access regimes: Instruction-only, Concepts (gold concepts without docs), Related-only (only relevant docs), and Full-set (relevant docs plus distractors). Uses frozen document libraries from 2023-2025 theoretical papers with expert annotations and two-phase validation to prevent parametric leakage.

Result: Experiments show substantial variation across state-of-the-art models: some exhibit mode-switch fragility (performing worse with Full-set than Instruction-only), while others show structural concept misuse (naming concepts correctly but failing to execute them as procedures).

Conclusion: DeR2 provides a clean framework for evaluating document-grounded reasoning, revealing significant headroom for improvement in LLMs’ ability to reason over novel scientific information.

Abstract: Despite strong performance on existing benchmarks, it remains unclear whether large language models can reason over genuinely novel scientific information. Most evaluations score end-to-end RAG pipelines, where reasoning is confounded with retrieval and toolchain choices, and the signal is further contaminated by parametric memorization and open-web volatility. We introduce DeR2, a controlled deep-research sandbox that isolates document-grounded reasoning while preserving core difficulties of deep search: multi-step synthesis, denoising, and evidence-based conclusion making. DeR2 decouples evidence access from reasoning via four regimes–Instruction-only, Concepts (gold concepts without documents), Related-only (only relevant documents), and Full-set (relevant documents plus topically related distractors)–yielding interpretable regime gaps that operationalize retrieval loss vs. reasoning loss and enable fine-grained error attribution. To prevent parametric leakage, we apply a two-phase validation that requires parametric failure without evidence while ensuring oracle-concept solvability. To ensure reproducibility, each instance provides a frozen document library (drawn from 2023-2025 theoretical papers) with expert-annotated concepts and validated rationales. Experiments across a diverse set of state-of-the-art foundation models reveal substantial variation and significant headroom: some models exhibit mode-switch fragility, performing worse with the Full-set than with Instruction-only, while others show structural concept misuse, correctly naming concepts but failing to execute them as procedures.

Method: Masked Energy Perturbation (MEP) applies energy masking to small energy regions in the frequency domain before generating adversarial perturbations. It targets areas less noticeable to human auditory perception, using power spectrum analysis to identify these regions. Tested on advanced speaker recognition models ECAPA-TDNN and ResNet34.

Result: MEP demonstrated strong performance in both audio quality and evasion effectiveness. It minimized PESQ degradation, showing minimal perceptual distortion to human listeners. Specifically achieved 26.68% relative performance in PESQ evaluation compared to FGSM and iterative FGSM methods.

Conclusion: MEP provides an effective adversarial attack method for voice data that balances evasion effectiveness with audio quality preservation, offering a promising countermeasure against voice deepfakes and unauthorized voice data use.

Abstract: Evasion attacks pose significant threats to AI systems, exploiting vulnerabilities in machine learning models to bypass detection mechanisms. The widespread use of voice data, including deepfakes, in promising future industries is currently hindered by insufficient legal frameworks. Adversarial attack methods have emerged as the most effective countermeasure against the indiscriminate use of such data. This research introduces masked energy perturbation (MEP), a novel approach using power spectrum for energy masking of original voice data. MEP applies masking to small energy regions in the frequency domain before generating adversarial perturbations, targeting areas less noticeable to the human auditory model. The study primarily employs advanced speaker recognition models, including ECAPA-TDNN and ResNet34, which have shown remarkable performance in speaker verification tasks. The proposed MEP method demonstrated strong performance in both audio quality and evasion effectiveness. The energy masking approach effectively minimizes the perceptual evaluation of speech quality (PESQ) degradation, indicating that minimal perceptual distortion occurs to the human listener despite the adversarial perturbations. Specifically, in the PESQ evaluation, the relative performance of the MEP method was 26.68% when compared to the fast gradient sign method (FGSM) and iterative FGSM.

Method: First analyzes SSL model behavior on noisy speech using mutual information between corrupted SSL representations and phoneme labels. Then introduces a linguistic aggregation layer pre-trained to maximize MI with phoneme labels (with optional dynamic aggregation), which is frozen during subsequent SE training.

Result: Experiments show the decoupled approach improves Word Error Rate (WER) over jointly optimized baselines, demonstrating benefits of explicitly aligning adaptation modules with linguistic contents.

Conclusion: Explicitly preserving linguistic information through pre-trained and frozen aggregation layers improves speech enhancement performance for downstream tasks like speech recognition, addressing the semantic corruption issue in current SSL-based SE approaches.

Abstract: Recent speech enhancement (SE) models increasingly leverage self-supervised learning (SSL) representations for their rich semantic information. Typically, intermediate features are aggregated into a single representation via a lightweight adaptation module. However, most SSL models are not trained for noise robustness, which can lead to corrupted semantic representations. Moreover, the adaptation module is trained jointly with the SE model, potentially prioritizing acoustic details over semantic information, contradicting the original purpose. To address this issue, we first analyze the behavior of SSL models on noisy speech from an information-theoretic perspective. Specifically, we measure the mutual information (MI) between the corrupted SSL representations and the corresponding phoneme labels, focusing on preservation of linguistic contents. Building upon this analysis, we introduce the linguistic aggregation layer, which is pre-trained to maximize MI with phoneme labels (with optional dynamic aggregation) and then frozen during SE training. Experiments show that this decoupled approach improves Word Error Rate (WER) over jointly optimized baselines, demonstrating the benefit of explicitly aligning the adaptation module with linguistic contents.

Method: Proposed automated pipeline that eliminates co-occurrence by mining high-purity single-event segments from in-the-wild datasets via semantically consistent synthesis protocol, creating the Hive dataset with 2.4k hours of raw audio.

Result: Models trained on Hive achieve competitive separation accuracy and perceptual quality compared to SAM-Audio (trained on dataset ~500x larger), and show remarkable zero-shot generalization on out-of-distribution benchmarks.

Conclusion: Prioritizing purity of supervised signals enables significant data efficiency, offering new paradigm for training robust auditory foundation models with reduced computational costs.

Abstract: Query-based universal sound separation is fundamental to intelligent auditory systems, aiming to isolate specific sources from mixtures. Despite recent advances, existing methods continue to suffer from residual interference in complex acoustic scenes. This performance limitation stems largely from a data bottleneck: in-the-wild datasets contain weak labels and severe co-occurrence of events. These flaws induce models to learn spurious correlations between background noise and target categories instead of robust acoustic features. To address this, we propose an automated pipeline that eliminates co-occurrence of events by mining high-purity single-event segments from in-the-wild datasets via a semantically consistent synthesis protocol. Utilizing this pipeline, we constructed Hive, a high-quality synthetic dataset comprising 2.4k hours of raw audio. Experimental results demonstrate that, compared with the state-of-the-art model SAM-Audio which was trained on a huge dataset $\sim$500 times larger than Hive, certain open-source models trained on Hive achieve competitive separation accuracy and perceptual quality. Moreover, these models exhibited remarkable zero-shot generalization on out-of-distribution evaluation benchmarks. These findings highlight that prioritizing purity of supervised signals enables significant data efficiency, offering a new paradigm for training robust auditory foundation models with reduced computational costs. Code and dataset are available at https://shandaai.github.io/Hive.

Method: Proposes Mean Continuation Log-Probability (MCLP) metric using LALM’s in-context learning to predict continuation log-probability of ground-truth speech given generated speech. Uses MCLP as reinforcement learning reward to enhance style alignment. Constructs RP-TTS dataset with scene/character annotations.

Result: Method significantly outperforms strong LALM baselines on both objective and subjective metrics for role-play TTS tasks.

Conclusion: MCLP effectively quantifies stylistic consistency and serves as a useful reward signal for improving LALM-based role-play TTS systems.

Abstract: Recent advances in Large Audio Language Models (LALMs) have extended Text-to-Speech (TTS) to interactive role-play scenarios, which demand high expressiveness and strict adherence to role-play instructions. However, existing models struggle to maintain stylistic consistency with character profiles and scene descriptions across multi-turn dialogues. A critical bottleneck is the lack of objective metrics for quantifying speaking style. To bridge this gap, we propose Mean Continuation Log-Probability (MCLP) as both an evaluation metric and a reward signal, validated on LALM-based Role-Play TTS (RP-TTS) tasks. Critically, we leverage the In-Context Learning capability of pre-trained LALMs to formulate MCLP via a continuation log-probability prediction. This metric quantifies stylistic consistency by measuring the likelihood of the ground-truth speech conditioned on the generated speech. Furthermore, we employ MCLP as a reinforcement learning reward to enhance the style alignment between generated speech and Role-Play instructions. To facilitate evaluation, we construct an RP-TTS dataset with rich scene and character annotations. Experimental results demonstrate that our method significantly outperforms strong LALM baselines on both objective and subjective metrics.

Method: Controlled comparative study of finetuning strategies for ABC-based generation and understanding, comparing off-the-shelf instruction-tuned backbone to domain-adapted variants and a music-specialized LLM baseline across multiple symbolic music corpora and evaluation signals.

Result: Provides insights into adaptation choices for symbolic music applications, highlighting the domain adaptation vs. preserving prior information tradeoff and distinct behavior of metrics used to measure domain adaptation for symbolic music.

Conclusion: The study offers practical guidance for adapting LLMs to symbolic music tasks, revealing important tradeoffs between domain specialization and preserving general language capabilities, with implications for multimodal music-language models.

Abstract: Music often shares notable parallels with language, motivating the use of pretrained large language models (LLMs) for symbolic music understanding and generation. Despite growing interest, the practical effectiveness of adapting instruction-tuned LLMs to symbolic music remains insufficiently characterized. We present a controlled comparative study of finetuning strategies for ABC-based generation and understanding, comparing an off-the-shelf instruction-tuned backbone to domain-adapted variants and a music-specialized LLM baseline. Across multiple symbolic music corpora and evaluation signals, we provide some insights into adaptation choices for symbolic music applications. We highlight the domain adaptation vs.~preserving prior information tradeoff as well as the distinct behaviour of metrics used to measure the domain adaptation for symbolic music.

Method: Proposes SDD-APALLM framework that combines raw audio with structured spectrograms to explicitly expose fine-grained time-frequency evidence as accessible acoustic cues, enabling audio LLMs to capture subtle acoustic inconsistencies without compromising semantic understanding.

Result: Experimental results show consistent gains in detection accuracy and robustness, especially in cases where semantic cues are misleading. Improvements stem from coordinated utilization of semantic and acoustic information rather than simple modality aggregation.

Conclusion: The acoustically enhanced framework effectively addresses the limitation of semantic-dominant reasoning in audio LLMs for speech deepfake detection by making fine-grained acoustic evidence more accessible during decision-making.

Abstract: Speech deepfake detection (SDD) focuses on identifying whether a given speech signal is genuine or has been synthetically generated. Existing audio large language model (LLM)-based methods excel in content understanding; however, their predictions are often biased toward semantically correlated cues, which results in fine-grained acoustic artifacts being overlooked during the decisionmaking process. Consequently, fake speech with natural semantics can bypass detectors despite harboring subtle acoustic anomalies; this suggests that the challenge stems not from the absence of acoustic data, but from its inadequate accessibility when semantic-dominant reasoning prevails. To address this issue, we investigate SDD within the audio LLM paradigm and introduce SDD with Auditory Perception-enhanced Audio Large Language Model (SDD-APALLM), an acoustically enhanced framework designed to explicitly expose fine-grained time-frequency evidence as accessible acoustic cues. By combining raw audio with structured spectrograms, the proposed framework empowers audio LLMs to more effectively capture subtle acoustic inconsistencies without compromising their semantic understanding. Experimental results indicate consistent gains in detection accuracy and robustness, especially in cases where semantic cues are misleading. Further analysis reveals that these improvements stem from a coordinated utilization of semantic and acoustic information, as opposed to simple modality aggregation.

Method: Introduces SYAUDIO benchmark with 4,319 audio questions spanning Audio Perception, Audio Reasoning, Audio Math, and Audio Ethics domains. Built upon established audio benchmarks and augmented with TTS-generated arithmetic and moral reasoning tasks. Enables systematic evaluation across multiple domains and sycophancy types with carefully verified data quality. Also analyzes audio-specific sycophancy under realistic conditions involving noise and rate variations.

Result: The benchmark enables comprehensive evaluation of sycophancy in ALMs. Analysis shows that supervised fine-tuning with chain-of-thought data is an effective mitigation strategy for reducing sycophantic behavior in ALMs.

Conclusion: SYAUDIO addresses a critical gap in evaluating behavioral issues in Audio Language Models, providing the first dedicated benchmark for sycophancy in audio-conditioned reasoning. The work demonstrates that sycophancy manifests in ALMs and can be mitigated through appropriate training strategies, advancing the reliability and trustworthiness of audio-based multimodal AI systems.

Abstract: Audio Language Models (ALMs) have recently shown strong capabilities in unified reasoning over speech, sound, and natural language; yet they inherit behavioral issues observed in Large Language Models, including sycophancy–the tendency to agree with user assertions even when they contradict objective evidence. While sycophancy has been extensively studied in text and vision-language models, its manifestation in audio-conditioned reasoning remains largely unexplored, despite the need for ALMs to rely on auditory cues such as acoustic events, speaker characteristics, and speech rate. To address this gap, we introduce SYAUDIO, the first benchmark dedicated to evaluating sycophancy in ALMs, consisting of 4,319 audio questions spanning Audio Perception, Audio Reasoning, Audio Math, and Audio Ethics. Built upon established audio benchmarks and augmented with TTS-generated arithmetic and moral reasoning tasks, SYAUDIO enables systematic evaluation across multiple domains and sycophancy types with carefully verified data quality. Furthermore, we analyze audio-specific sycophancy under realistic conditions involving noise and rate, and demonstrate that supervised fine-tuning with chain-of-thought data is an effective mitigation strategy for reducing sycophantic behavior in ALMs.

Method: DIFFA-2 upgrades the speech encoder, uses dual semantic and acoustic adapters, and employs a four-stage curriculum training: semantic and acoustic alignment, large-scale supervised fine-tuning, and variance-reduced preference optimization using only open-source corpora.

Result: Experiments on MMSU, MMAU, and MMAR benchmarks show DIFFA-2 consistently improves over DIFFA and is competitive with strong autoregressive LALMs under practical training budgets, demonstrating diffusion-based modeling as a viable backbone for large-scale audio understanding.

Conclusion: Diffusion-based large audio language models are a practical alternative to autoregressive models, offering competitive performance with more efficient training and inference characteristics for general audio understanding tasks.

Abstract: Autoregressive (AR) large audio language models (LALMs) such as Qwen-2.5-Omni have achieved strong performance on audio understanding and interaction, but scaling them remains costly in data and computation, and strictly sequential decoding limits inference efficiency. Diffusion large language models (dLLMs) have recently been shown to make effective use of limited training data, and prior work on DIFFA indicates that replacing an AR backbone with a diffusion counterpart can substantially improve audio understanding under matched settings, albeit at a proof-of-concept scale without large-scale instruction tuning, preference alignment, or practical decoding schemes. We introduce DIFFA-2, a practical diffusion-based LALM for general audio understanding. DIFFA-2 upgrades the speech encoder, employs dual semantic and acoustic adapters, and is trained with a four-stage curriculum that combines semantic and acoustic alignment, large-scale supervised fine-tuning, and variance-reduced preference optimization, using only fully open-source corpora. Experiments on MMSU, MMAU, and MMAR show that DIFFA-2 consistently improves over DIFFA and is competitive to strong AR LALMs under practical training budgets, supporting diffusion-based modeling is a viable backbone for large-scale audio understanding. Our code is available at https://github.com/NKU-HLT/DIFFA.git.

Method: Proposes Spatial Mapping and Regression Localization (SMRL-SELD) that segments 3D spatial space into 2D plane and introduces regression localization loss. This location-oriented approach learns event features based on orientation rather than tracks.

Result: Outperforms existing SELD methods on STARSS23 and STARSS22 datasets, particularly in overall evaluation and polyphony environments.

Conclusion: SMRL-SELD enables processing polyphonic sounds regardless of overlapping event count by being location-oriented rather than track-limited, improving generality in complex acoustic environments.

Abstract: Sound Event Localization and Detection (SELD) combines the Sound Event Detection (SED) with the corresponding Direction Of Arrival (DOA). Recently, adopted event oriented multi-track methods affect the generality in polyphonic environments due to the limitation of the number of tracks. To enhance the generality in polyphonic environments, we propose Spatial Mapping and Regression Localization for SELD (SMRL-SELD). SMRL-SELD segments the 3D spatial space, mapping it to a 2D plane, and a new regression localization loss is proposed to help the results converge toward the location of the corresponding event. SMRL-SELD is location-oriented, allowing the model to learn event features based on orientation. Thus, the method enables the model to process polyphonic sounds regardless of the number of overlapping events. We conducted experiments on STARSS23 and STARSS22 datasets and our proposed SMRL-SELD outperforms the existing SELD methods in overall evaluation and polyphony environments.

Method: Two-stage framework: 1) Synthesizes complete music in mel-spectrogram representation that encapsulates musical essence of noise, 2) Adaptively amplifies generated music segments to reduce noise perception while preserving auditory quality.

Result: Experiments on MusicBench, EPIC-SOUNDS, and ESC-50 demonstrate effectiveness in blending environmental noise with rhythmically aligned, adaptively amplified music segments, minimizing noise noticeability and improving acoustic experiences.

Conclusion: BNMusic offers an alternative to traditional acoustic masking by generating personalized music that blends with environmental noises through cross-modal generation and adaptive amplification, reducing noise perception without excessive volume.

Abstract: While being disturbed by environmental noises, the acoustic masking technique is a conventional way to reduce the annoyance in audio engineering that seeks to cover up the noises with other dominant yet less intrusive sounds. However, misalignment between the dominant sound and the noise-such as mismatched downbeats-often requires an excessive volume increase to achieve effective masking. Motivated by recent advances in cross-modal generation, in this work, we introduce an alternative method to acoustic masking, aiming to reduce the noticeability of environmental noises by blending them into personalized music generated based on user-provided text prompts. Following the paradigm of music generation using mel-spectrogram representations, we propose a Blending Noises into Personalized Music (BNMusic) framework with two key stages. The first stage synthesizes a complete piece of music in a mel-spectrogram representation that encapsulates the musical essence of the noise. In the second stage, we adaptively amplify the generated music segment to further reduce noise perception and enhance the blending effectiveness, while preserving auditory quality. Our experiments with comprehensive evaluations on MusicBench, EPIC-SOUNDS, and ESC-50 demonstrate the effectiveness of our framework, highlighting the ability to blend environmental noise with rhythmically aligned, adaptively amplified, and enjoyable music segments, minimizing the noticeability of the noise, thereby improving overall acoustic experiences. Project page: https://d-fas.github.io/BNMusic_page/.

Method: Introduces “contiguous monologues” composed of continuous sentences with “waiting” intervals to mimic human cognitive behavior. Develops a “dual” training paradigm that alternates monologue positions (leading or trailing audio) across training stages. Combines these approaches in FLM-Audio, a 7B parameter spoken dialog chatbot.

Result: FLM-Audio achieves superior response qualities and chatting experiences while requiring significantly less training data compared to existing approaches.

Conclusion: The contiguous monologue approach with dual training enables native full-duplex spoken dialog systems with better language modeling capabilities and lower training data requirements.

Abstract: Full-duplex dialog models aim to listen and speak simultaneously, delivering rapid responses to dynamic user input. Among different solutions to full-duplexity, a native solution merges multiple channels in each time step, achieving the lowest latency. However, prevailing designs break down the textual monologue sentences for word-level alignment with audio streams, which degrades language modeling abilities. To help address this issue, we introduce “contiguous monologues”, which are composed by continuous sentences and “waiting” intervals, mimicking human-like cognitive behavior in dialogs. We find a proper training paradigm to be critical for semantically aligning contiguous monologues with audio. To this end, we develop a “dual” training paradigm that alternates the position of the monologues, either leading or trailing the audio, across different training stages. A combination of our contiguous monologue and dual training strategy is applied in developing FLM-Audio, our 7B spoken dialog chatbot with native full-duplexity. As confirmed by experimental results, FLM-Audio achieves superior response qualities and chatting experiences while requiring significantly less training data.

Method: The authors analyze adversarial audio at the phonetic level, showing perturbations exploit systematic confusions like vowel centralization and consonant substitutions. Using DeepSpeech as the ASR target, they generate targeted adversarial examples and evaluate their impact on speaker embeddings across genuine and impostor samples.

Result: Results across 16 phonetically diverse target phrases demonstrate that adversarial audio induces both transcription errors and identity drift, showing these distortions not only mislead transcription but also degrade phonetic cues critical for speaker verification.

Conclusion: Adversarial attacks exploit phonetic vulnerabilities in speech systems, highlighting the need for phonetic-aware defenses to ensure robustness of both ASR and speaker recognition systems against such attacks.

Abstract: Adversarial perturbations in speech pose a serious threat to automatic speech recognition (ASR) and speaker verification by introducing subtle waveform modifications that remain imperceptible to humans but can significantly alter system outputs. While targeted attacks on end-to-end ASR models have been widely studied, the phonetic basis of these perturbations and their effect on speaker identity remain underexplored. In this work, we analyze adversarial audio at the phonetic level and show that perturbations exploit systematic confusions such as vowel centralization and consonant substitutions. These distortions not only mislead transcription but also degrade phonetic cues critical for speaker verification, leading to identity drift. Using DeepSpeech as our ASR target, we generate targeted adversarial examples and evaluate their impact on speaker embeddings across genuine and impostor samples. Results across 16 phonetically diverse target phrases demonstrate that adversarial audio induces both transcription errors and identity drift, highlighting the need for phonetic-aware defenses to ensure the robustness of ASR and speaker recognition systems.

Method: Proposes a framework incorporating CoT and RL training into TS-ASR: 1) Constructs a novel CoT dataset for TS-ASR, 2) Trains model on regular data then fine-tunes on CoT data, 3) Further trains with RL using selected data to enhance generalized reasoning capabilities.

Result: Experiment results show significant improvement of TS-ASR performance with CoT and RL training, demonstrating the effectiveness of the proposed methods adapted for the TS-ASR task.

Conclusion: The combination of Chain of Thoughts and Reinforcement Learning training provides an effective approach for improving Target Speaker ASR performance in complex multi-speaker scenarios by enhancing the model’s reasoning capabilities.

Abstract: Target Speaker Automatic Speech Recognition (TS-ASR) aims to transcribe the speech of a specified target speaker from multi-speaker mixtures in cocktail party scenarios. Recent advancement of Large Audio-Language Models (LALMs) has already brought some new insights to TS-ASR. However, significant room for optimization remains for the TS-ASR task within the LALMs architecture. While Chain of Thoughts (CoT) and Reinforcement Learning (RL) have proven effective in certain speech tasks, TS-ASR, which requires the model to deeply comprehend speech signals, differentiate various speakers, and handle overlapping utterances is particularly well-suited to a reasoning-guided approach. Therefore, we propose a novel framework that incorporates CoT and RL training into TS-ASR for performance improvement. A novel CoT dataset of TS-ASR is constructed, and the TS-ASR model is first trained on regular data and then fine-tuned on CoT data. Finally, the model is further trained with RL using selected data to enhance generalized reasoning capabilities. Experiment results show a significant improvement of TS-ASR performance with CoT and RL training, which demonstrates the effectiveness of the proposed CoT and RL training methods adapted for the TS-ASR task.

Method: Constructed CompSpoof dataset with multiple combinations of bona fide and spoofed speech/environmental sounds. Proposed separation-enhanced joint learning framework that separates audio components and applies anti-spoofing models to each component separately with joint learning to preserve detection-relevant information.

Result: Extensive experiments show the proposed method outperforms baselines, demonstrating the necessity of separate component analysis and importance of detecting spoofing for each component individually.

Conclusion: Component-level audio spoofing represents a new challenge requiring specialized detection approaches, and the proposed framework with component separation and joint learning effectively addresses this problem.

Abstract: Component-level audio Spoofing (Comp-Spoof) targets a new form of audio manipulation where only specific components of a signal, such as speech or environmental sound, are forged or substituted while other components remain genuine. Existing anti-spoofing datasets and methods treat an utterance or a segment as entirely bona fide or entirely spoofed, and thus cannot accurately detect component-level spoofing. To address this, we construct a new dataset, CompSpoof, covering multiple combinations of bona fide and spoofed speech and environmental sound. We further propose a separation-enhanced joint learning framework that separates audio components apart and applies anti-spoofing models to each one. Joint learning is employed, preserving information relevant for detection. Extensive experiments demonstrate that our method outperforms the baseline, highlighting the necessity of separate components and the importance of detecting spoofing for each component separately. Datasets and code are available at: https://github.com/XuepingZhang/CompSpoof.

Method: Proposes Dynamic Dual-Signal Curriculum (DDSC) that combines two signals computed each epoch: domain-invariance signal and learning-progress signal. A time-varying scheduler fuses these into per-example weights that prioritize domain-invariant examples early and gradually emphasize device-specific cases.

Result: Under DCASE 2024 Task 1 protocol, DDSC consistently improves cross-device performance across diverse ASC baselines and label budgets, with largest gains on unseen-device splits.

Conclusion: DDSC is a lightweight, architecture-agnostic method that addresses device domain shift in acoustic scene classification through dynamic curriculum learning without additional inference overhead.

Abstract: Acoustic scene classification (ASC) suffers from device-induced domain shift, especially when labels are limited. Prior work focuses on curriculum-based training schedules that structure data presentation by ordering or reweighting training examples from easy-to-hard to facilitate learning; however, existing curricula are static, fixing the ordering or the weights before training and ignoring that example difficulty and marginal utility evolve with the learned representation. To overcome this limitation, we propose the Dynamic Dual-Signal Curriculum (DDSC), a training schedule that adapts the curriculum online by combining two signals computed each epoch: a domain-invariance signal and a learning-progress signal. A time-varying scheduler fuses these signals into per-example weights that prioritize domain-invariant examples in early epochs and progressively emphasize device-specific cases. DDSC is lightweight, architecture-agnostic, and introduces no additional inference overhead. Under the official DCASE 2024 Task~1 protocol, DDSC consistently improves cross-device performance across diverse ASC baselines and label budgets, with the largest gains on unseen-device splits.

Method: Reformulates forced alignment as slot-filling: timestamps are discrete indices, special timestamp tokens are inserted as slots into transcripts. SLLMs predict time indices at slots conditioned on speech embeddings and transcript with slots. Uses causal attention masking with non-shifted sequences, loss computed only at slot positions. Dynamic slot insertion enables alignment at arbitrary positions with non-autoregressive inference.

Result: Achieves 69%~78% relative reduction in accumulated averaging shift compared to prior methods across multilingual, crosslingual, and long-form speech scenarios. Provides checkpoint and inference code publicly available.

Conclusion: LLM-ForcedAligner effectively leverages speech LLMs for forced alignment through slot-filling paradigm, addressing limitations of existing methods and enabling accurate multilingual and long-form speech alignment with reduced temporal shifts.

Abstract: Forced alignment (FA) predicts start and end timestamps for words or characters in speech, but existing methods are language-specific and prone to cumulative temporal shifts. The multilingual speech understanding and long-sequence processing abilities of speech large language models (SLLMs) make them promising for FA in multilingual, crosslingual, and long-form speech settings. However, directly applying the next-token prediction paradigm of SLLMs to FA results in hallucinations and slow inference. To bridge the gap, we propose LLM-ForcedAligner, reformulating FA as a slot-filling paradigm: timestamps are treated as discrete indices, and special timestamp tokens are inserted as slots into the transcript. Conditioned on the speech embeddings and the transcript with slots, the SLLM directly predicts the time indices at slots. During training, causal attention masking with non-shifted input and label sequences allows each slot to predict its own timestamp index based on itself and preceding context, with loss computed only at slot positions. Dynamic slot insertion enables FA at arbitrary positions. Moreover, non-autoregressive inference is supported, avoiding hallucinations and improving speed. Experiments across multilingual, crosslingual, and long-form speech scenarios show that LLM-ForcedAligner achieves a 69%~78% relative reduction in accumulated averaging shift compared with prior methods. Checkpoint and inference code are available at https://github.com/QwenLM/Qwen3-ASR.

Method: Proposes TopSeg framework that encodes PCG dynamics with multi-scale topological features (H_0 and H_1 persistence diagrams) and decodes them using a lightweight temporal convolutional network (TCN) with order- and duration-constrained inference.

Result: Topological features consistently outperform spectrogram and envelope inputs under matched-capacity decoders, especially at low data budgets. Full system surpasses end-to-end baselines under same data budgets while remaining competitive at full data.

Conclusion: Topology-aware representations provide strong inductive bias for data-efficient, cross-dataset PCG segmentation, supporting practical use when labeled data are limited.

Abstract: Deep learning approaches for heart-sound (PCG) segmentation built on time-frequency features can be accurate but often rely on large expert-labeled datasets, limiting robustness and deployment. We present TopSeg, a topological representation-centric framework that encodes PCG dynamics with multi-scale topological features and decodes them using a lightweight temporal convolutional network (TCN) with an order- and duration-constrained inference step. To evaluate data efficiency and generalization, we train exclusively on PhysioNet 2016 dataset with subject-level subsampling and perform external validation on CirCor dataset. Under matched-capacity decoders, the topological features consistently outperform spectrogram and envelope inputs, with the largest margins at low data budgets; as a full system, TopSeg surpasses representative end-to-end baselines trained on their native inputs under the same budgets while remaining competitive at full data. Ablations at 10% training confirm that all scales contribute and that combining H_0 and H_1 yields more reliable S1/S2 localization and boundary stability. These results indicate that topology-aware representations provide a strong inductive bias for data-efficient, cross-dataset PCG segmentation, supporting practical use when labeled data are limited.

Method: Introduces a novel text-only adaptation method that emulates the audio projection task by treating it as a text denoising task. The LLM is trained to recover clean transcripts from noisy inputs, adapting to target domains while preserving cross-modal alignment. The solution is lightweight with no architectural changes or additional parameters.

Result: Extensive evaluation on two datasets demonstrates up to 22.1% relative improvement, outperforming recent state-of-the-art text-only adaptation methods.

Conclusion: The proposed text denoising approach effectively adapts LLM-based ASR systems to new domains using only text data while maintaining cross-modal alignment, offering a lightweight and effective solution.

Abstract: Adapting automatic speech recognition (ASR) systems based on large language models (LLMs) to new domains using text-only data is a significant yet underexplored challenge. Standard fine-tuning of the LLM on target-domain text often disrupts the critical alignment between speech and text modalities learned by the projector, degrading performance. We introduce a novel text-only adaptation method that emulates the audio projection task by treating it as a text denoising task. Our approach thus trains the LLM to recover clean transcripts from noisy inputs. This process effectively adapts the model to a target domain while preserving cross-modal alignment. Our solution is lightweight, requiring no architectural changes or additional parameters. Extensive evaluation on two datasets demonstrates up to 22.1% relative improvement, outperforming recent state-of-the-art text-only adaptation methods.

Method: 1) Revisits existing framework to explicitly model both speech and acoustic noise as latent variables, jointly sampling them in the E-step. 2) Introduces new unsupervised SE framework replacing NMF noise prior with diffusion-based noise model, learned jointly with speech prior in a single conditional score model. 3) Derives two variants: implicit noise accounting and explicit noise as latent variable.

Result: Explicit noise modeling systematically improves SE performance for both NMF-based and diffusion-based noise priors. Diffusion-based noise model attains best overall quality and intelligibility among unsupervised methods under matched conditions. NMF-based explicit-noise framework shows better robustness and less degradation under mismatched conditions than several supervised baselines.

Conclusion: Explicit noise modeling in diffusion-based speech enhancement frameworks consistently improves performance. Diffusion-based noise priors offer superior performance under matched conditions, while NMF-based approaches provide better robustness to mismatched conditions, making them competitive with supervised methods.

Abstract: This paper addresses unsupervised diffusion-based single-channel speech enhancement (SE). Prior work in this direction combines a score-based diffusion model trained on clean speech with a Gaussian noise model whose covariance is structured by non-negative matrix factorization (NMF). This combination is used within an iterative expectation-maximization (EM) scheme, in which a diffusion-based posterior-sampling E-step estimates the clean speech. We first revisit this framework and propose to explicitly model both speech and acoustic noise as latent variables, jointly sampling them in the E-step instead of sampling speech alone as in previous approaches. We then introduce a new unsupervised SE framework that replaces the NMF noise prior with a diffusion-based noise model, learned jointly with the speech prior in a single conditional score model. Within this framework, we derive two variants: one that implicitly accounts for noise and one that explicitly treats noise as a latent variable. Experiments on WSJ0-QUT and VoiceBank-DEMAND show that explicit noise modeling systematically improves SE performance for both NMF-based and diffusion-based noise priors. Under matched conditions, the diffusion-based noise model attains the best overall quality and intelligibility among unsupervised methods, while under mismatched conditions the proposed NMF-based explicit-noise framework is more robust and suffers less degradation than several supervised baselines.

Method: Proposes two novel metrics: (1) Difference of Inter-annotator Consistency (DIC) - quantifies how well models capture annotator tendencies by comparing predicted vs. ground-truth inter-annotator similarity structures; (2) Behavior Alignment Explainability (BAE) - evaluates how well model explanations reflect annotator behavior and decision relevance by aligning explainability-derived with ground-truth labeling similarity structures using Multidimensional Scaling (MDS).

Result: Extensive experiments validate the effectiveness of the proposed evaluation framework, demonstrating its ability to properly assess ITL methods’ capability to capture individual annotator tendencies and provide meaningful behavioral explanations.

Conclusion: The paper presents the first unified evaluation framework for Individual Tendency Learning methods, addressing a critical gap in multi-annotator learning research by providing systematic metrics to assess whether models truly capture annotator-specific behaviors and provide meaningful explanations.

Abstract: Recent works have emerged in multi-annotator learning that shift focus from Consensus-oriented Learning (CoL), which aggregates multiple annotations into a single ground-truth prediction, to Individual Tendency Learning (ITL), which models annotator-specific labeling behavior patterns (i.e., tendency) to provide explanation analysis for understanding annotator decisions. However, no evaluation framework currently exists to assess whether ITL methods truly capture individual tendencies and provide meaningful behavioral explanations. To address this gap, we propose the first unified evaluation framework with two novel metrics: (1) Difference of Inter-annotator Consistency (DIC) quantifies how well models capture annotator tendencies by comparing predicted inter-annotator similarity structures with ground-truth; (2) Behavior Alignment Explainability (BAE) evaluates how well model explanations reflect annotator behavior and decision relevance by aligning explainability-derived with ground-truth labeling similarity structures via Multidimensional Scaling (MDS). Extensive experiments validate the effectiveness of our proposed evaluation framework.

Method: Systematic study using EAV dataset with three model categories: baseline transformers (M1), novel factorized attention mechanisms (M2), and improved CNN baselines (M3). Tested domain-specific modifications like delta MFCCs for audio, frequency-domain features for EEG, and vision delta features.

Result: Complex attention mechanisms (M2) underperformed by 5-13 percentage points due to overfitting. Simple modifications were effective: audio delta MFCCs improved accuracy from 61.9% to 65.56%; EEG frequency features achieved 67.62% (+7.62pp); vision transformer baseline reached 75.30% (exceeding ViViT); vision delta features achieved 72.68%.

Conclusion: For small-scale multimodal emotion recognition, domain knowledge and proper implementation outperform architectural complexity. Simple domain-specific modifications are more effective than sophisticated attention mechanisms on limited data.

Abstract: We present a systematic study of multimodal emotion recognition using the EAV dataset, investigating whether complex attention mechanisms improve performance on small datasets. We implement three model categories: baseline transformers (M1), novel factorized attention mechanisms (M2), and improved CNN baselines (M3). Our experiments show that sophisticated attention mechanisms consistently underperform on small datasets. M2 models achieved 5 to 13 percentage points below baselines due to overfitting and destruction of pretrained features. In contrast, simple domain-appropriate modifications proved effective: adding delta MFCCs to the audio CNN improved accuracy from 61.9% to \textbf{65.56%} (+3.66pp), while frequency-domain features for EEG achieved \textbf{67.62%} (+7.62pp over the paper baseline). Our vision transformer baseline (M1) reached \textbf{75.30%}, exceeding the paper’s ViViT result (74.5%) through domain-specific pretraining, and vision delta features achieved \textbf{72.68%} (+1.28pp over the paper CNN). These findings demonstrate that for small-scale emotion recognition, domain knowledge and proper implementation outperform architectural complexity.

Method: Hybrid model combining multitask learning for efficient data encoding with location weight module using quantum convolution operations for feature extraction.

Result: Validated on multiple Earth observation benchmarks, showing potential advantages and good generalizability of the quantum-enhanced approach.

Conclusion: Demonstrates promising potential of quantum machine learning for Earth observation data analysis despite current quantum device limitations.

Abstract: Quantum machine learning (QML) has gained increasing attention as a potential solution to address the challenges of computation requirements in the future. Earth observation (EO) has entered the era of Big Data, and the computational demands for effectively analyzing large EO data with complex deep learning models have become a bottleneck. Motivated by this, we aim to leverage quantum computing for EO data classification and explore its advantages despite the current limitations of quantum devices. This paper presents a hybrid model that incorporates multitask learning to assist efficient data encoding and employs a location weight module with quantum convolution operations to extract valid features for classification. The validity of our proposed model was evaluated using multiple EO benchmarks. Additionally, we experimentally explored the generalizability of our model and investigated the factors contributing to its advantage, highlighting the potential of QML in EO data analysis.

Method: Developed CELM by curating a large-scale clinical EEG dataset (9,922 reports with ~11,000 hours of EEG from 9,048 patients), integrating pretrained EEG foundation models with language models for multimodal learning, enabling end-to-end clinical report generation at multiple scales including recording description, background activity, epileptiform abnormalities, events/seizures, and impressions.

Result: With patient history supervision: 70%-95% average relative improvements in generation metrics (ROUGE-1, METEOR) from 0.2-0.3 to 0.4-0.6. Zero-shot without patient history: CELM achieves 0.43-0.52 vs baselines of 0.17-0.26.

Conclusion: CELM represents the first clinical EEG-to-language foundation model capable of scalable multimodal learning for automated clinical report generation from long-duration EEG recordings, with significant performance improvements over existing methods.

Abstract: Generating clinical reports that summarize abnormal patterns, diagnostic findings, and clinical interpretations from long-term EEG recordings remains labor-intensive. We curate a large-scale clinical EEG dataset with $9{,}922$ reports paired with approximately $11{,}000$ hours of EEG recordings from $9{,}048$ patients. We therefore develop CELM, the first clinical EEG-to-Language foundation model capable of summarizing long-duration, variable-length EEG recordings and performing end-to-end clinical report generation at multiple scales, including recording description, background activity, epileptiform abnormalities, events/seizures, and impressions. Experimental results show that, with patient history supervision, our method achieves $70%$–$95%$ average relative improvements in standard generation metrics (e.g., ROUGE-1 and METEOR) from $0.2$–$0.3$ to $0.4$–$0.6$. In the zero-shot setting without patient history, CELM attains generation scores in the range of $0.43$–$0.52$, compared to baselines of $0.17$–$0.26$. CELM integrates pretrained EEG foundation models with language models to enable scalable multimodal learning. We release our model and benchmark construction pipeline at [URL].

Method: Proposes FedAdaVR which incorporates an adaptive optimizer with variance reduction technique, using recent stored updates from clients even when absent. Also proposes FedAdaVR-Quant which stores client updates in quantized form to reduce memory requirements.

Result: FedAdaVR can eliminate partial client participation error, and FedAdaVR-Quant reduces memory requirements by 50%, 75%, and 87.5% while maintaining equivalent model performance. Extensive experiments on multiple datasets under IID and non-IID settings show consistent outperformance over state-of-the-art baselines.

Conclusion: FedAdaVR effectively addresses heterogeneity issues in federated learning, particularly partial client participation, with theoretical convergence guarantees and practical memory-efficient variants.

Abstract: Federated learning (FL) encounters substantial challenges due to heterogeneity, leading to gradient noise, client drift, and partial client participation errors, the last of which is the most pervasive but remains insufficiently addressed in current literature. In this paper, we propose FedAdaVR, a novel FL algorithm aimed at solving heterogeneity issues caused by sporadic client participation by incorporating an adaptive optimiser with a variance reduction technique. This method takes advantage of the most recent stored updates from clients, even when they are absent from the current training round, thereby emulating their presence. Furthermore, we propose FedAdaVR-Quant, which stores client updates in quantised form, significantly reducing the memory requirements (by 50%, 75%, and 87.5%) of FedAdaVR while maintaining equivalent model performance. We analyse the convergence behaviour of FedAdaVR under general nonconvex conditions and prove that our proposed algorithm can eliminate partial client participation error. Extensive experiments conducted on multiple datasets, under both independent and identically distributed (IID) and non-IID settings, demonstrate that FedAdaVR consistently outperforms state-of-the-art baseline methods.

Method: Proposes CausIL framework inspired by causal modeling, treating noisy state observations as proxy variables. Uses proximal causal inference for identification, with estimators for discrete/continuous spaces (adversarial RKHS procedure for continuous).

Result: CausIL demonstrates improved robustness to distribution shift compared to BC baselines on semi-simulated longitudinal data from PhysioNet/Computing in Cardiology Challenge 2019 cohort.

Conclusion: CausIL provides a principled causal framework for offline imitation learning under measurement error, enabling robust policy recovery from demonstrations without rewards or interactive expert queries.

Abstract: We study offline imitation learning (IL) when part of the decision-relevant state is observed only through noisy measurements and the distribution may change between training and deployment. Such settings induce spurious state-action correlations, so standard behavioral cloning (BC) – whether conditioning on raw measurements or ignoring them – can converge to systematically biased policies under distribution shift. We propose a general framework for IL under measurement error, inspired by explicitly modeling the causal relationships among the variables, yielding a target that retains a causal interpretation and is robust to distribution shift. Building on ideas from proximal causal inference, we introduce \texttt{CausIL}, which treats noisy state observations as proxy variables, and we provide identification conditions under which the target policy is recoverable from demonstrations without rewards or interactive expert queries. We develop estimators for both discrete and continuous state spaces; for continuous settings, we use an adversarial procedure over RKHS function classes to learn the required parameters. We evaluate \texttt{CausIL} on semi-simulated longitudinal data from the PhysioNet/Computing in Cardiology Challenge 2019 cohort and demonstrate improved robustness to distribution shift compared to BC baselines.

Method: Proposes LSFlow with: 1) Latent spherical flow policy that learns stochastic policies in compact continuous latent space via spherical flow matching, 2) Delegates feasibility to combinatorial optimization solvers that map latent samples to valid structured actions, 3) Trains value network directly in latent space to avoid repeated solver calls, 4) Introduces smoothed Bellman operator to address piecewise-constant and discontinuous value landscape from solver-based action selection.

Result: Empirically outperforms state-of-the-art baselines by an average of 20.6% across a range of challenging combinatorial RL tasks.

Conclusion: LSFlow brings the expressiveness of modern generative policies to combinatorial RL while guaranteeing feasibility by design, offering a general and expressive approach to combinatorial action spaces.

Abstract: Reinforcement learning (RL) with combinatorial action spaces remains challenging because feasible action sets are exponentially large and governed by complex feasibility constraints, making direct policy parameterization impractical. Existing approaches embed task-specific value functions into constrained optimization programs or learn deterministic structured policies, sacrificing generality and policy expressiveness. We propose a solver-induced \emph{latent spherical flow policy} that brings the expressiveness of modern generative policies to combinatorial RL while guaranteeing feasibility by design. Our method, LSFlow, learns a \emph{stochastic} policy in a compact continuous latent space via spherical flow matching, and delegates feasibility to a combinatorial optimization solver that maps each latent sample to a valid structured action. To improve efficiency, we train the value network directly in the latent space, avoiding repeated solver calls during policy optimization. To address the piecewise-constant and discontinuous value landscape induced by solver-based action selection, we introduce a smoothed Bellman operator that yields stable, well-defined learning targets. Empirically, our approach outperforms state-of-the-art baselines by an average of 20.6% across a range of challenging combinatorial RL tasks.

Method: Proposes DAJ, a reasoning-based LLM judge trained with verifiable rewards under a bi-level data-reweighted learning framework that learns data-importance weights (domain-level or instance-level) to optimize generalization on a held-out meta set aligned with target benchmarks.

Result: DAJ achieves state-of-the-art performance on LiveCodeBench and BigCodeBench, outperforming strong test-time scaling baselines and leading proprietary models.

Conclusion: The proposed data-reweighted learning framework effectively addresses distribution shift challenges in LLM judge training for test-time scaling, automatically emphasizing hard problems, in-distribution samples, and trajectory-aligned data without hand-crafted heuristics.

Abstract: Test-time scaling for code generation commonly relies on Best-of-N selection, in which multiple candidate solutions are sampled from a base model, and the best one is selected by an LLM judge. However, training reliable LLM judges is challenging due to severe distribution shifts, including imbalances between easy and hard problems, mismatches between training tasks and evaluation benchmarks, and trajectory mismatch arising from training data generated by cheaper models whose behavior differs from that of inference-time models. We propose DAJ, a reasoning-based LLM judge trained with verifiable rewards under a bi-level data-reweighted learning framework. The proposed framework learns data-importance weights (either domain-level or instance-level) to optimize generalization performance on a held-out meta set aligned with target benchmarks. To the best of our knowledge, this is the first application of data reweighting to LLM-as-a-Judge training for test-time scaling. Our approach automatically emphasizes hard problems, in-distribution samples, and trajectory-aligned data, without relying on hand-crafted heuristics. Empirically, DAJ achieves state-of-the-art performance on LiveCodeBench and BigCodeBench, outperforming strong test-time scaling baselines as well as leading proprietary models.

Method: FunPRM prompts LLMs to generate modular code organized into functions, treating functions as PRM reasoning steps. It introduces a meta-learning-based reward correction mechanism that uses clean final-solution rewards from unit tests to purify noisy partial-solution rewards.

Result: FunPRM consistently outperforms existing test-time scaling methods across five base LLMs on LiveCodeBench and BigCodeBench, achieving state-of-the-art performance on LiveCodeBench when combined with O4-mini. It also produces more readable and reusable code.

Conclusion: FunPRM effectively addresses code-specific challenges in PRM-based test-time scaling by using function-level decomposition and reward purification, significantly improving code generation performance and quality.

Abstract: Code generation is a core application of large language models (LLMs), yet LLMs still frequently fail on complex programming tasks. Given its success in mathematical reasoning, test-time scaling approaches such as Process Reward Model (PRM)-based Best-of-N selection offer a promising way to improve performance. However, existing PRMs remain ineffective for code generation due to the lack of meaningful step decomposition in code and the noise of Monte Carlo-estimated partial-solution correctness scores (rewards). To address these challenges, we propose FunPRM. FunPRM prompts LLMs to encourage modular code generation organized into functions, with functions treated as PRM reasoning steps. Furthermore, FunPRM introduces a novel meta-learning-based reward correction mechanism that leverages clean final-solution rewards obtained via a unit-test-based evaluation system to purify noisy partial-solution rewards. Experiments on LiveCodeBench and BigCodeBench demonstrate that FunPRM consistently outperforms existing test-time scaling methods across five base LLMs, notably achieving state-of-the-art performance on LiveCodeBench when combined with O4-mini. Furthermore, FunPRM produces code that is more readable and reusable for developers.

Method: AgentScore performs semantically guided optimization in exponentially large discrete rule spaces by using LLMs to propose candidate rules, followed by a deterministic verification-and-selection loop to enforce statistical validity and deployability constraints.

Result: Across eight clinical prediction tasks, AgentScore outperforms existing score-generation methods and achieves AUC comparable to more flexible interpretable models despite stronger structural constraints. On two externally validated tasks, it achieves higher discrimination than established guideline-based scores.

Conclusion: AgentScore bridges the gap between ML performance and clinical deployability by generating interpretable unit-weighted clinical checklists that meet workflow constraints while maintaining strong predictive performance.

Abstract: Modern clinical practice relies on evidence-based guidelines implemented as compact scoring systems composed of a small number of interpretable decision rules. While machine-learning models achieve strong performance, many fail to translate into routine clinical use due to misalignment with workflow constraints such as memorability, auditability, and bedside execution. We argue that this gap arises not from insufficient predictive power, but from optimizing over model classes that are incompatible with guideline deployment. Deployable guidelines often take the form of unit-weighted clinical checklists, formed by thresholding the sum of binary rules, but learning such scores requires searching an exponentially large discrete space of possible rule sets. We introduce AgentScore, which performs semantically guided optimization in this space by using LLMs to propose candidate rules and a deterministic, data-grounded verification-and-selection loop to enforce statistical validity and deployability constraints. Across eight clinical prediction tasks, AgentScore outperforms existing score-generation methods and achieves AUC comparable to more flexible interpretable models despite operating under stronger structural constraints. On two additional externally validated tasks, AgentScore achieves higher discrimination than established guideline-based scores.

Method: Insert batchwise-sampled, unlearned query and value biases to break rotational symmetry in attention, creating a preferred direction in the rotational space.

Result: Substantially improves performance of memory-efficient optimizers (narrows/closes gap to complex adaptive methods) and enables interpretable amplification of semantically meaningful token classes.

Conclusion: Minimal, principled architectural changes to attention can simultaneously improve optimization performance and model interpretability.

Abstract: The attention mechanism in its standard implementation contains extraneous rotational degrees of freedom that are carried through computation but do not affect model activations or outputs. We introduce a simple symmetry-breaking protocol that inserts a preferred direction into this rotational space through batchwise-sampled, unlearned query and value biases. This modification has two theoretically motivated and empirically validated consequences. First, it can substantially improve the performance of simple, memory-efficient optimizers, narrowing – and in some cases closing – the gap to successful but more complex memory-intensive adaptive methods. We demonstrate this by pretraining 124M parameter transformer models with four optimization algorithms (AdamW, SOAP, SGDM, and Energy Conserving Descent(ECD)) and evaluating both validation loss and downstream logical reasoning. Second, it enables an interpretable use of otherwise redundant rotational degrees of freedom, selectively amplifying semantically meaningful token classes within individual attention heads. Overall, our results show that minimal, principled architectural changes can simultaneously improve performance and interpretability.

Method: Reformulates both static and dynamic survival analysis as series of binary classification problems by discretizing event times. Censored observations are handled as examples with missing labels at certain time points. Enables tabular foundation models to perform survival analysis through in-context learning without explicit training.

Result: Proved that under standard censoring assumptions, minimizing binary classification loss recovers true survival probabilities as training set increases. Evaluation across 53 real-world datasets shows off-the-shelf tabular foundation models with this formulation outperform classical and deep learning baselines on multiple survival metrics.

Conclusion: Classification formulation enables existing tabular foundation models to perform survival analysis through in-context learning without explicit training, providing strong performance across diverse datasets.

Abstract: While tabular foundation models have achieved remarkable success in classification and regression, adapting them to model time-to-event outcomes for survival analysis is non-trivial due to right-censoring, where data observations may end before the event occurs. We develop a classification-based framework that reformulates both static and dynamic survival analysis as a series of binary classification problems by discretizing event times. Censored observations are naturally handled as examples with missing labels at certain time points. This classification formulation enables existing tabular foundation models to perform survival analysis through in-context learning without explicit training. We prove that under standard censoring assumptions, minimizing our binary classification loss recovers the true survival probabilities as the training set size increases. We demonstrate through evaluation across $53$ real-world datasets that off-the-shelf tabular foundation models with this classification formulation outperform classical and deep learning baselines on average over multiple survival metrics.

Method: Proposed a human activity recognition framework using wearable inertial sensors (accelerometer and gyroscope). Compared four classical classifiers (Logistic Regression, Random Forest, SVM, k-NN) with a novel Support Tensor Machine (STM) that preserves spatio-temporal motion dynamics through tensor representations.

Result: SVM achieved 93.33% accuracy, while other classical methods ranged 91.11-93.33%. STM significantly outperformed all with 96.67% test accuracy and 98.50% cross-validation accuracy, demonstrating superior classification across diverse activities.

Conclusion: The STM-based framework offers a scalable, low-cost solution for remote healthcare applications including elderly assistance, child monitoring, yoga feedback, and smart home wellness, particularly suitable for low-resource settings.

Abstract: Limited access to medical infrastructure forces elderly and vulnerable patients to rely on home-based care, often leading to neglect and poor adherence to therapeutic exercises such as yoga or physiotherapy. To address this gap, we propose a low-cost and automated human activity recognition (HAR) framework based on wearable inertial sensors and machine learning. Activity data, including walking, walking upstairs, walking downstairs, sitting, standing, and lying, were collected using accelerometer and gyroscope measurements. Four classical classifiers, Logistic Regression, Random Forest, Support Vector Machine (SVM), and k-Nearest Neighbors (k-NN), were evaluated and compared with the proposed Support Tensor Machine (STM). Experimental results show that SVM achieved an accuracy of 93.33 percent, while Logistic Regression, Random Forest, and k-NN achieved 91.11 percent. In contrast, STM significantly outperformed these models, achieving a test accuracy of 96.67 percent and the highest cross-validation accuracy of 98.50 percent. Unlike conventional methods, STM leverages tensor representations to preserve spatio-temporal motion dynamics, resulting in robust classification across diverse activities. The proposed framework demonstrates strong potential for remote healthcare, elderly assistance, child activity monitoring, yoga feedback, and smart home wellness, offering a scalable solution for low-resource and rural healthcare settings.

Method: DyCAST uses soft character-level alignment and explicit duration modeling to learn associations between tokens and linguistic units, supports alignment-free inference with direct duration control, and includes retrieval-augmented decoding for improved quality at low frame rates.

Result: DyCAST achieves competitive speech resynthesis quality and downstream performance while using significantly fewer tokens than fixed-frame-rate codecs.

Conclusion: DyCAST provides an efficient variable-frame-rate speech tokenization approach that reduces sequence length while maintaining quality, making it more suitable for LLM processing.

Abstract: Neural audio codecs are at the core of modern conversational speech technologies, converting continuous speech into sequences of discrete tokens that can be processed by LLMs. However, existing codecs typically operate at fixed frame rates, allocating tokens uniformly in time and producing unnecessarily long sequences. In this work, we introduce DyCAST, a Dynamic Character-Aligned Speech Tokenizer that enables variable-frame-rate tokenization through soft character-level alignment and explicit duration modeling. DyCAST learns to associate tokens with character-level linguistic units during training and supports alignment-free inference with direct control over token durations at decoding time. To improve speech resynthesis quality at low frame rates, we further introduce a retrieval-augmented decoding mechanism that enhances reconstruction fidelity without increasing bitrate. Experiments show that DyCAST achieves competitive speech resynthesis quality and downstream performance while using significantly fewer tokens than fixed-frame-rate codecs.

Method: SPECIAL adds a single server-side “anchor” to vanilla FedAvg: in each round, the server nudges uniformly sampled participating clients’ updates toward the previous global model with a lightweight proximal term, curbing cumulative drift without replay buffers or synthetic data.

Result: Theoretical analysis shows SPECIAL preserves earlier tasks with backward knowledge transfer bounds and achieves communication-efficient non-convex convergence rate O((E/NT)^(1/2)) matching single-task FedAvg while separating optimization variance from inter-task drift.

Conclusion: SPECIAL provides a simple, memory-free solution for federated domain-incremental learning with theoretical guarantees for backward knowledge transfer and efficient convergence across tasks with partial participation.

Abstract: Real-world federated systems seldom operate on static data: input distributions drift while privacy rules forbid raw-data sharing. We study this setting as Federated Domain-Incremental Learning (FDIL), where (i) clients are heterogeneous, (ii) tasks arrive sequentially with shifting domains, yet (iii) the label space remains fixed. Two theoretical pillars remain missing for FDIL under realistic deployment: a guarantee of backward knowledge transfer (BKT) and a convergence rate that holds across the sequence of all tasks with partial participation. We introduce SPECIAL (Server-Proximal Efficient Continual Aggregation for Learning), a simple, memory-free FDIL algorithm that adds a single server-side ``anchor’’ to vanilla FedAvg: in each round, the server nudges the uniformly sampled participated clients update toward the previous global model with a lightweight proximal term. This anchor curbs cumulative drift without replay buffers, synthetic data, or task-specific heads, keeping communication and model size unchanged. Our theory shows that SPECIAL (i) preserves earlier tasks: a BKT bound caps any increase in prior-task loss by a drift-controlled term that shrinks with more rounds, local epochs, and participating clients; and (ii) learns efficiently across all tasks: the first communication-efficient non-convex convergence rate for FDIL with partial participation, O((E/NT)^(1/2)), with E local epochs, T communication rounds, and N participated clients per round, matching single-task FedAvg while explicitly separating optimization variance from inter-task drift. Experimental results further demonstrate the effectiveness of SPECIAL.

Method: Formulates IR-drop estimation as dense pixel-wise regression using U-Net encoder-decoder architecture with skip connections. Trained on synthetic dataset incorporating power grid structure, cell density distribution, and switching activity. Uses MSE and PSNR for evaluation.

Result: Model accurately predicts IR-drop distributions with millisecond-level inference time, enabling fast pre-signoff screening and iterative design optimization.

Conclusion: Proposed deep learning framework serves as complementary early-stage analysis tool for rapid IR-drop insight before expensive signoff analysis, with publicly available implementation and interactive application.

Abstract: IR-drop is a critical power integrity challenge in modern VLSI designs that can cause timing degradation, reliability issues, and functional failures if not detected early in the design flow. Conventional IR-drop analysis relies on physics-based signoff tools, which provide high accuracy but incur significant computational cost and require near-final layout information, making them unsuitable for rapid early-stage design exploration. In this work, we propose a deep learning-based surrogate modeling approach for early-stage IR-drop estimation using a CNN. The task is formulated as a dense pixel-wise regression problem, where spatial physical layout features are mapped directly to IR-drop heatmaps. A U-Net-based encoder-decoder architecture with skip connections is employed to effectively capture both local and global spatial dependencies within the layout. The model is trained on a physics-inspired synthetic dataset generated by us, which incorporates key physical factors including power grid structure, cell density distribution, and switching activity. Model performance is evaluated using standard regression metrics such as Mean Squared Error (MSE) and Peak Signal-to-Noise Ratio (PSNR). Experimental results demonstrate that the proposed approach can accurately predict IR-drop distributions with millisecond-level inference time, enabling fast pre-signoff screening and iterative design optimization. The proposed framework is intended as a complementary early-stage analysis tool, providing designers with rapid IR-drop insight prior to expensive signoff analysis. The implementation, dataset generation scripts, and the interactive inference application are publicly available at: https://github.com/riteshbhadana/IR-Drop-Predictor. The live application can be accessed at: https://ir-drop-predictor.streamlit.app/.

Method: Proposes SurrogateSHAP that approximates expensive retraining through inference from pretrained models, uses gradient-boosted trees to approximate utility functions, and derives Shapley values analytically from tree-based models.

Result: Outperforms prior methods across three attribution tasks (image quality, aesthetics, product diversity) while reducing computational overhead, effectively identifies influential contributors, and localizes data sources for spurious correlations in clinical images.

Conclusion: SurrogateSHAP provides a scalable framework for fair data attribution in T2I diffusion models, enabling practical implementation of Shapley values for data marketplace compensation and safety-critical model auditing.

Abstract: As Text-to-Image (T2I) diffusion models are increasingly used in real-world creative workflows, a principled framework for valuing contributors who provide a collection of data is essential for fair compensation and sustainable data marketplaces. While the Shapley value offers a theoretically grounded approach to attribution, it faces a dual computational bottleneck: (i) the prohibitive cost of exhaustive model retraining for each sampled subset of players (i.e., data contributors) and (ii) the combinatorial number of subsets needed to estimate marginal contributions due to contributor interactions. To this end, we propose SurrogateSHAP, a retraining-free framework that approximates the expensive retraining game through inference from a pretrained model. To further improve efficiency, we employ a gradient-boosted tree to approximate the utility function and derive Shapley values analytically from the tree-based model. We evaluate SurrogateSHAP across three diverse attribution tasks: (i) image quality for DDPM-CFG on CIFAR-20, (ii) aesthetics for Stable Diffusion on Post-Impressionist artworks, and (iii) product diversity for FLUX.1 on Fashion-Product data. Across settings, SurrogateSHAP outperforms prior methods while substantially reducing computational overhead, consistently identifying influential contributors across multiple utility metrics. Finally, we demonstrate that SurrogateSHAP effectively localizes data sources responsible for spurious correlations in clinical images, providing a scalable path toward auditing safety-critical generative models.

Method: RLOs reinterpret optimization as an extended state discrete-time controlled dynamical system on Riemannian parameter manifolds. They identify Normally Attracting Invariant Manifolds (NAIM) that organize training dynamics into two stages, construct strict Lyapunov functions for convergence certification, and develop an “optimizer generator” that recovers classic algorithms while enabling new designs.

Result: The framework successfully recovers classic optimization algorithms and enables principled design of new RLOs. Geometric diagnostics validate the theory, and the resulting optimizers achieve state-of-the-art performance on large-scale benchmarks.

Conclusion: RLOs provide a unified geometric framework bridging control theory and machine learning optimization, offering systematic tools for designing stable, effective optimizers with theoretical guarantees.

Abstract: We introduce Riemannian Lyapunov Optimizers (RLOs), a family of optimization algorithms that unifies classic optimizers within one geometric framework. Unlike heuristic improvements to existing optimizers, RLOs are systematically derived from a novel control-theoretic framework that reinterprets optimization as an extended state discrete-time controlled dynamical system on a Riemannian parameter manifold. Central to this framework is the identification of a Normally Attracting Invariant Manifold (NAIM), which organizes training dynamics into two distinct stages: rapid alignment of the speed state to a target graph, followed by controlled evolution within it. We formalize this by constructing a strict Lyapunov function that certifies convergence to a target manifold. This perspective yields a constructive ``optimizer generator” that not only recovers classic algorithms but enables the principled design of RLOs. We validate our theory via geometric diagnostics and demonstrate that grounding optimizer design in control theory yields state-of-the-art performance in large-scale benchmarks. Overall, RLOs bridge control theory and modern machine learning optimization, providing a unified language and a systematic toolkit for designing stable, effective optimizers.

Method: Uses an architecture-agnostic framework with linear optimization over interpretable pairwise metrics (e.g., gradient L2 distance) to analyze mergeability across four merging methods, identifying method-specific “fingerprints” and foundational prerequisites.

Result: Found substantial variation in success drivers (46.7% metric overlap; 55.3% sign agreement) revealing method-specific patterns, but subspace overlap and gradient alignment metrics consistently emerge as method-agnostic prerequisites for compatibility.

Conclusion: Mergeability depends on both merging method and partner tasks, with subspace overlap and gradient alignment as foundational prerequisites; provides diagnostic foundation for understanding mergeability and motivates fine-tuning strategies that encourage these properties.

Abstract: Model merging combines knowledge from separately fine-tuned models, yet success factors remain poorly understood. While recent work treats mergeability as an intrinsic property, we show with an architecture-agnostic framework that it fundamentally depends on both the merging method and the partner tasks. Using linear optimization over a set of interpretable pairwise metrics (e.g., gradient L2 distance), we uncover properties correlating with post-merge performance across four merging methods. We find substantial variation in success drivers (46.7% metric overlap; 55.3% sign agreement), revealing method-specific “fingerprints”. Crucially, however, subspace overlap and gradient alignment metrics consistently emerge as foundational, method-agnostic prerequisites for compatibility. These findings provide a diagnostic foundation for understanding mergeability and motivate future fine-tuning strategies that explicitly encourage these properties.

Method: Proposes AT-GRPO with two components: (1) agent- and turn-wise grouped RL algorithm tailored for MAS, and (2) training system supporting both single- and multi-policy regimes for MAS workflow rollouts and on-policy updates.

Result: Substantial gains across game, planning, coding, and math tasks. On long-horizon planning: accuracy improved from 14.0-47.0% baseline to 96.0-99.5%. Coding tasks: average gains of 3.87-7.62%. Math tasks: average gains of 9.0-17.93%.

Conclusion: AT-GRPO successfully addresses challenges in applying on-policy RL to multi-agent LLM systems, demonstrating significant performance improvements across diverse reasoning and planning tasks through specialized grouping algorithms and flexible training infrastructure.

Abstract: Multi-agent systems (MAS) and reinforcement learning (RL) are widely used to enhance the agentic capabilities of large language models (LLMs). MAS improves task performance through role-based orchestration, while RL uses environmental rewards to learn stronger policies, such as GRPO-style optimization. However, applying on-policy RL to MAS remains underexplored and presents unique challenges. Algorithmically, standard GRPO grouping assumptions break down because prompts vary by role and by turn. System-wise, the training stack must support MAS-workflow rollouts and on-policy updates for both single-policy and multi-policy models. We propose AT-GRPO, which includes (i) an agent- and turn-wise grouped RL algorithm tailored to MAS and (ii) a training system that supports both single- and multi-policy regimes. Across game, planning, coding, and math tasks, AT-GRPO delivers substantial gains. On long-horizon planning, it increases accuracy from a 14.0 to 47.0 percent single-agent RL baseline to 96.0 to 99.5 percent. It also improves reasoning performance, with average gains of 3.87 to 7.62 percent on coding tasks and 9.0 to 17.93 percent on math. Code and environments are available at: https://github.com/pettingllms-ai/PettingLLMs.

Method: ParalESN uses structured operators and state space modeling to create high-dimensional reservoirs based on diagonal linear recurrence in complex space, enabling parallel processing of temporal data while preserving theoretical properties.

Result: ParalESN matches traditional RC accuracy on time series benchmarks with substantial computational savings. On 1-D pixel-level classification, it achieves competitive accuracy with trainable neural networks while reducing computational costs and energy consumption by orders of magnitude.

Conclusion: ParalESN provides a scalable, principled pathway for integrating Reservoir Computing within deep learning, addressing key limitations of traditional RC while maintaining theoretical guarantees and practical efficiency.

Abstract: Reservoir Computing (RC) has established itself as an efficient paradigm for temporal processing. However, its scalability remains severely constrained by (i) the necessity of processing temporal data sequentially and (ii) the prohibitive memory footprint of high-dimensional reservoirs. In this work, we revisit RC through the lens of structured operators and state space modeling to address these limitations, introducing Parallel Echo State Network (ParalESN). ParalESN enables the construction of high-dimensional and efficient reservoirs based on diagonal linear recurrence in the complex space, enabling parallel processing of temporal data. We provide a theoretical analysis demonstrating that ParalESN preserves the Echo State Property and the universality guarantees of traditional Echo State Networks while admitting an equivalent representation of arbitrary linear reservoirs in the complex diagonal form. Empirically, ParalESN matches the predictive accuracy of traditional RC on time series benchmarks, while delivering substantial computational savings. On 1-D pixel-level classification tasks, ParalESN achieves competitive accuracy with fully trainable neural networks while reducing computational costs and energy consumption by orders of magnitude. Overall, ParalESN offers a promising, scalable, and principled pathway for integrating RC within the deep learning landscape.

Method: Proposes CP4Gen, a conformal prediction approach that uses clustering-based density estimation on model-generated samples to construct prediction sets that are less sensitive to outliers and more interpretable.

Result: Extensive experiments on synthetic datasets and real-world applications (including climate emulation) show CP4Gen achieves superior performance in prediction set volume and structural simplicity compared to existing methods.

Conclusion: CP4Gen provides practitioners with a powerful tool for uncertainty estimation in conditional generative models, particularly valuable for scenarios requiring rigorous and interpretable prediction sets.

Abstract: Conditional generative models map input variables to complex, high-dimensional distributions, enabling realistic sample generation in a diverse set of domains. A critical challenge with these models is the absence of calibrated uncertainty, which undermines trust in individual outputs for high-stakes applications. To address this issue, we propose a systematic conformal prediction approach tailored to conditional generative models, leveraging density estimation on model-generated samples. We introduce a novel method called CP4Gen, which utilizes clustering-based density estimation to construct prediction sets that are less sensitive to outliers, more interpretable, and of lower structural complexity than existing methods. Extensive experiments on synthetic datasets and real-world applications, including climate emulation tasks, demonstrate that CP4Gen consistently achieves superior performance in terms of prediction set volume and structural simplicity. Our approach offers practitioners a powerful tool for uncertainty estimation associated with conditional generative models, particularly in scenarios demanding rigorous and interpretable prediction sets.

Method: Proposes ZK-HybridFL combining directed acyclic graph (DAG) ledger with dedicated sidechains and zero-knowledge proofs (ZKPs). Uses event-driven smart contracts and oracle-assisted sidechain to verify local model updates without exposing sensitive data. Includes built-in challenge mechanism for adversarial behavior detection.

Result: Achieves faster convergence, higher accuracy, lower perplexity, and reduced latency compared to Blade-FL and ChainFL on image classification and language modeling tasks. Robust against adversarial and idle nodes, supports sub-second on-chain verification with efficient gas usage, prevents invalid updates and orphanage-style attacks.

Conclusion: ZK-HybridFL provides a scalable and secure solution for decentralized federated learning across diverse environments with privacy-preserving validation capabilities.

Abstract: Federated learning (FL) enables collaborative model training while preserving data privacy, yet both centralized and decentralized approaches face challenges in scalability, security, and update validation. We propose ZK-HybridFL, a secure decentralized FL framework that integrates a directed acyclic graph (DAG) ledger with dedicated sidechains and zero-knowledge proofs (ZKPs) for privacy-preserving model validation. The framework uses event-driven smart contracts and an oracle-assisted sidechain to verify local model updates without exposing sensitive data. A built-in challenge mechanism efficiently detects adversarial behavior. In experiments on image classification and language modeling tasks, ZK-HybridFL achieves faster convergence, higher accuracy, lower perplexity, and reduced latency compared to Blade-FL and ChainFL. It remains robust against substantial fractions of adversarial and idle nodes, supports sub-second on-chain verification with efficient gas usage, and prevents invalid updates and orphanage-style attacks. This makes ZK-HybridFL a scalable and secure solution for decentralized FL across diverse environments.

Method: BWG casts workflow generation as Bayesian inference over posterior distribution on workflows. It uses parallel look-ahead rollouts for importance weighting and sequential in-loop refiner for pool-wide improvements. Instantiated as BayesFlow algorithm.

Result: Across six benchmark datasets, BayesFlow improves accuracy by up to 9 percentage points over SOTA workflow generation baselines and up to 65 percentage points over zero-shot prompting.

Conclusion: BWG provides a principled upgrade to search-based workflow design with theoretical grounding and strong empirical performance.

Abstract: Automatic workflow generation is the process of automatically synthesizing sequences of LLM calls, tool invocations, and post-processing steps for complex end-to-end tasks. Most prior methods cast this task as an optimization problem with limited theoretical grounding. We propose to cast workflow generation as Bayesian inference over a posterior distribution on workflows, and introduce \textbf{Bayesian Workflow Generation (BWG)}, a sampling framework that builds workflows step-by-step using parallel look-ahead rollouts for importance weighting and a sequential in-loop refiner for pool-wide improvements. We prove that, without the refiner, the weighted empirical distribution converges to the target posterior. We instantiate BWG as \textbf{BayesFlow}, a training-free algorithm for workflow construction. Across six benchmark datasets, BayesFlow improves accuracy by up to 9 percentage points over SOTA workflow generation baselines and by up to 65 percentage points over zero-shot prompting, establishing BWG as a principled upgrade to search-based workflow design. Code will be available on https://github.com/BoYuanVisionary/BayesFlow.

Method: Derives exact moment matching for probit, GeLU, ReLU, Heaviside, and sine activation functions for both feedforward and generalized residual layers using mathematical analysis of Gaussian propagation.

Result: Achieves orders-of-magnitude improvements (up to millionfold) in KL divergence error on random networks, competitive statistical calibration on real data, and hundredfold improvements over state-of-the-art deterministic inference methods.

Conclusion: The method provides exact solutions for Gaussian propagation through neural networks with various activations, offering substantial improvements in uncertainty quantification and inference accuracy.

Abstract: We study the problem of propagating the mean and covariance of a general multivariate Gaussian distribution through a deep (residual) neural network using layer-by-layer moment matching. We close a longstanding gap by deriving exact moment matching for the probit, GeLU, ReLU (as a limit of GeLU), Heaviside (as a limit of probit), and sine activation functions; for both feedforward and generalized residual layers. On random networks, we find orders-of-magnitude improvements in the KL divergence error metric, up to a millionfold, over popular alternatives. On real data, we find competitive statistical calibration for inference under epistemic uncertainty in the input. On a variational Bayes network, we show that our method attains hundredfold improvements in KL divergence from Monte Carlo ground truth over a state-of-the-art deterministic inference method. We also give an a priori error bound and a preliminary analysis of stochastic feedforward neurons, which have recently attracted general interest.

Method: Proposes a novel optimal stealthy attack formulation considering different detectability degrees, develops a normalization-based methodology for evaluating effectiveness-detectability trade-offs, and creates BayesClean defense against stealthy attacks.

Result: The proposed stealthy attack bypasses state-of-the-art defenses, and BayesClean improves defense performance when attacks are stealthy and poisoning points are significant.

Conclusion: The paper addresses practical poisoning threats to regression models with novel attack formulations and defense mechanisms, advancing the field of model robustness against adversarial data poisoning.

Abstract: Regression models are widely used in industrial processes, engineering and in natural and physical sciences, yet their robustness to poisoning has received less attention. When it has, studies often assume unrealistic threat models and are thus less useful in practice. In this paper, we propose a novel optimal stealthy attack formulation that considers different degrees of detectability and show that it bypasses state-of-the-art defenses. We further propose a new methodology based on normalization of objectives to evaluate different trade-offs between effectiveness and detectability. Finally, we develop a novel defense (BayesClean) against stealthy attacks. BayesClean improves on previous defenses when attacks are stealthy and the number of poisoning points is significant.

Method: Created SCALAR benchmark with ≈100,000 DFT-validated structures spanning atomic to nanoparticle scales (up to 18,000 atoms). Includes three tasks: CIF to property prediction, Chain-of-Thought physics reasoning, and inverse retrieval. Uses structured metrics for numeric error, hallucination, consistency, monotonic reasoning, validity, and retrieval regret.

Result: Experiments show large, model-dependent performance shifts under explicit reasoning, often reducing hallucination and error but frequently destabilizing consistency or validity. Geometric scale generalization cannot be inferred from accuracy alone.

Conclusion: SCALAR reveals critical gaps in materials foundation models’ geometric reasoning capabilities and provides a comprehensive benchmark for evaluating scale generalization and structural consistency in materials science applications.

Abstract: Large language models are increasingly applied to materials science reasoning, yet their behavior under physically structured distribution shifts remains poorly understood. We introduce SCALAR (Structural Consistency And Logic Across Regimes), a benchmark for evaluating geometric scale generalization and its connection to structural hallucination, consistency, and reasoning in materials foundation models. Given canonical crystal representations, models must reason about derived nanoparticle structures obtained through supercell expansion and geometric truncation across length scales spanning a few atoms to over 18,000 atoms, totaling $\approx$100,000 structures from DFT-validated unit cells. SCALAR defines three tasks. (i) CIF to property prediction. (ii) A Chain-of-Thought variant with explicit physics-grounded reasoning. (iii) Inverse retrieval identifying crystals from candidates given target properties. Outputs are evaluated via structured metrics capturing numeric error, hallucination, cross-prompt consistency, monotonic reasoning, output validity, and retrieval regret. Experiments across diverse foundation models reveal large, model-dependent shifts under explicit reasoning, often reducing hallucination and error, but frequently destabilizing consistency or validity. These results demonstrate that geometric scale generalization cannot be inferred from accuracy alone. Supplementary materials are available at https://github.com/KurbanIntelligenceLab/SCALAR.

Method: Theoretical analysis showing static alignment provides no guarantee of post-update alignment due to overparameterization, plus empirical validation across privacy, jailbreak safety, and behavioral honesty domains in LLMs.

Result: Models can pass all standard black-box alignment tests yet become severely misaligned after a single benign update. The capacity to hide latent adversarial behavior increases with model scale.

Conclusion: Static evaluation protocols are inadequate; there’s urgent need for post-update-robust alignment evaluation methods.

Abstract: Large Language Models (LLMs) are rarely static and are frequently updated in practice. A growing body of alignment research has shown that models initially deemed “aligned” can exhibit misaligned behavior after fine-tuning, such as forgetting jailbreak safety features or re-surfacing knowledge that was intended to be forgotten. These works typically assume that the initial model is aligned based on static black-box evaluation, i.e., the absence of undesired responses to a fixed set of queries. In contrast, we formalize model alignment in both the static and post-update settings and uncover a fundamental limitation of black-box evaluation. We theoretically show that, due to overparameterization, static alignment provides no guarantee of post-update alignment for any update dataset. Moreover, we prove that static black-box probing cannot distinguish a model that is genuinely post-update robust from one that conceals an arbitrary amount of adversarial behavior which can be activated by even a single benign gradient update. We further validate these findings empirically in LLMs across three core alignment domains: privacy, jailbreak safety, and behavioral honesty. We demonstrate the existence of LLMs that pass all standard black-box alignment tests, yet become severely misaligned after a single benign update. Finally, we show that the capacity to hide such latent adversarial behavior increases with model scale, confirming our theoretical prediction that post-update misalignment grows with the number of parameters. Together, our results highlight the inadequacy of static evaluation protocols and emphasize the urgent need for post-update-robust alignment evaluation.

Method: Prediction-Augmented Gaussian Process Upper Confidence Bound (PA-GP-UCB) uses control-variates estimator from joint Gaussian process posterior to correct prediction bias and reduce uncertainty, leveraging both oracles and offline data.

Result: PA-GP-UCB preserves standard GP-UCB regret rate with strictly smaller leading constant controlled by prediction quality and offline data coverage. Outperforms baselines on synthetic benchmarks and real-world human behavioral hypothesis evaluation using LLM predictions.

Conclusion: PA-GP-UCB provides general, sample-efficient framework for hypothesis generation under expensive feedback, effectively combining expensive ground-truth with cheap predictions and offline data.

Abstract: Many real-world optimization problems involve an expensive ground-truth oracle (e.g., human evaluation, physical experiments) and a cheap, low-fidelity prediction oracle (e.g., machine learning models, simulations). Meanwhile, abundant offline data (e.g., past experiments and predictions) are often available and can be used to pretrain powerful predictive models, as well as to provide an informative prior. We propose Prediction-Augmented Gaussian Process Upper Confidence Bound (PA-GP-UCB), a novel Bayesian optimization algorithm that leverages both oracles and offline data to achieve provable gains in sample efficiency for the ground-truth oracle queries. PA-GP-UCB employs a control-variates estimator derived from a joint Gaussian process posterior to correct prediction bias and reduce uncertainty. We prove that PA-GP-UCB preserves the standard regret rate of GP-UCB while achieving a strictly smaller leading constant that is explicitly controlled by prediction quality and offline data coverage. Empirically, PA-GP-UCB converges faster than Vanilla GP-UCB and naive prediction-augmented GP-UCB baselines on synthetic benchmarks and on a real-world hypothesis evaluation task grounded in human behavioral data, where predictions are provided by large language models. These results establish PA-GP-UCB as a general and sample-efficient framework for hypothesis generation under expensive feedback.

Method: Combines: (1) group-action on divergence-free flows, (2) graph-attention encoder to learn feature-conditioned weights over symmetry-preserving actions, and (3) lightweight Tikhonov refinement via implicit bilevel optimization. Uses GATv2 layers to encode graph features into per-edge embeddings, then attention-guided selection of physics-aware group actions.

Result: Outperforms state-of-the-art baselines in RMSE, MAE and correlation metrics across three real-world flow benchmarks (traffic, power, bike).

Conclusion: FlowSymm effectively combines physical constraints with data-driven learning for network flow recovery, demonstrating superior performance on real-world flow systems.

Abstract: Recovering missing flows on the edges of a network, while exactly respecting local conservation laws, is a fundamental inverse problem that arises in many systems such as transportation, energy, and mobility. We introduce FlowSymm, a novel architecture that combines (i) a group-action on divergence-free flows, (ii) a graph-attention encoder to learn feature-conditioned weights over these symmetry-preserving actions, and (iii) a lightweight Tikhonov refinement solved via implicit bilevel optimization. The method first anchors the given observation on a minimum-norm divergence-free completion. We then compute an orthonormal basis for all admissible group actions that leave the observed flows invariant and parameterize the valid solution subspace, which shows an Abelian group structure under vector addition. A stack of GATv2 layers then encodes the graph and its edge features into per-edge embeddings, which are pooled over the missing edges and produce per-basis attention weights. This attention-guided process selects a set of physics-aware group actions that preserve the observed flows. Finally, a scalar Tikhonov penalty refines the missing entries via a convex least-squares solver, with gradients propagated implicitly through Cholesky factorization. Across three real-world flow benchmarks (traffic, power, bike), FlowSymm outperforms state-of-the-art baselines in RMSE, MAE and correlation metrics.

Method: Introduces a federated framework for LLM routing where clients learn a shared routing policy from local offline query-model evaluation data. Supports both parametric multilayer perceptron routers and nonparametric K-means routers. Handles heterogeneous client query distributions and non-uniform model coverage through federated collaboration.

Result: Across two benchmarks, federated collaboration improves the accuracy-cost frontier over client-local routers. Improvements come from increased effective model coverage and better query generalization. Theoretical results validate that federated training reduces routing suboptimality.

Conclusion: Federated learning enables effective LLM routing without centralizing privacy-sensitive data, overcoming limitations of both centralized and purely local approaches. The framework supports diverse router architectures and handles real-world heterogeneity in client data distributions and model coverage.

Abstract: Large language models (LLMs) are increasingly accessed as remotely hosted services by edge and enterprise clients that cannot run frontier models locally. Since models vary widely in capability and price, routing queries to models that balance quality and inference cost is essential. Existing router approaches assume access to centralized query-model evaluation data. However, these data are often fragmented across clients, such as end users and organizations, and are privacy-sensitive, which makes centralizing data infeasible. Additionally, per-client router training is ineffective since local evaluation data is limited and covers only a restricted query distribution and a biased subset of model evaluations. We introduce the first federated framework for LLM routing, enabling clients to learn a shared routing policy from local offline query-model evaluation data. Our framework supports both parametric multilayer perceptron router and nonparametric K-means router under heterogeneous client query distributions and non-uniform model coverage. Across two benchmarks, federated collaboration improves the accuracy-cost frontier over client-local routers, both via increased effective model coverage and better query generalization. Our theoretical results also validate that federated training reduces routing suboptimality.

Method: Uses approximate differential privacy with Matrix Factorization mechanism, introducing a novel mean estimation specific factorization that is both efficient and accurate.

Result: Achieves asymptotically lower mean-squared error bounds in continual mean estimation under user-level differential privacy compared to pure DP approaches.

Conclusion: The proposed approximate DP approach with specialized factorization provides more practical and accurate continual mean estimation while maintaining user-level privacy guarantees.

Abstract: We study continual mean estimation, where data vectors arrive sequentially and the goal is to maintain accurate estimates of the running mean. We address this problem under user-level differential privacy, which protects each user’s entire dataset even when they contribute multiple data points. Previous work on this problem has focused on pure differential privacy. While important, this approach limits applicability, as it leads to overly noisy estimates. In contrast, we analyze the problem under approximate differential privacy, adopting recent advances in the Matrix Factorization mechanism. We introduce a novel mean estimation specific factorization, which is both efficient and accurate, achieving asymptotically lower mean-squared error bounds in continual mean estimation under user-level differential privacy.

Method: Combines a Graph Transformer (GT) model to capture spatial topology and signal strength dynamics with a novel Spatially-Adaptive Conformal Prediction (SACP) method that provides region-specific uncertainty estimates.

Result: Achieves state-of-the-art localization accuracy on large-scale real-world datasets while delivering robust and spatially adaptive reliability guarantees.

Conclusion: SAC-GT provides a comprehensive solution for indoor localization that addresses both accuracy and reliability requirements, making it suitable for real-world deployment in location-based services.

Abstract: Indoor localization is a critical enabler for a wide range of location-based services in smart environments, including navigation, asset tracking, and safety-critical applications. Recent graph-based models leverage spatial relationships between Wire-less Fidelity (Wi-Fi) Access Points (APs) and devices, offering finer localization granularity, but fall short in quantifying prediction uncertainty, a key requirement for real-world deployment. In this paper, we propose Spatially-Adaptive Conformal Graph Transformer (SAC-GT), a framework for accurate and reliable indoor localization. SAC-GT integrates a Graph Transformer (GT) model that captures network’s spatial topology and signal strength dynamics, with a novel Spatially-Adaptive Conformal Prediction (SACP) method that provides region-specific uncertainty estimates. This allows SAC-GT to produce not only precise two-dimensional (2D) location predictions but also statistically valid confidence regions tailored to varying environmental conditions. Extensive evaluations on a large-scale real-world dataset demonstrate that the proposed SAC-GT solution achieves state-of-the-art localization accuracy while delivering robust and spatially adaptive reliability guarantees.

Method: SCOPE uses reinforcement learning to make reasoning-based predictions by retrieving how models behave on similar problems, rather than relying on fixed model names. It explicitly predicts both accuracy and cost metrics for models.

Result: SCOPE achieves significant improvements: boosts accuracy by up to 25.7% when performance is prioritized, or cuts costs by up to 95.1% when efficiency is the main concern.

Conclusion: SCOPE provides a flexible, scalable routing framework that goes beyond simple model selection by enabling dynamic cost-performance trade-offs and adapting to new, unseen models through retrieval-based reasoning.

Abstract: Model routing chooses which language model to use for each query. By sending easy queries to cheaper models and hard queries to stronger ones, it can significantly reduce inference cost while maintaining high accuracy. However, most existing routers treat this as a fixed choice among a small set of models, which makes them hard to adapt to new models or changing budget constraints. In this paper, we propose SCOPE (Scalable and Controllable Outcome Performance Estimator), a routing framework that goes beyond model selection by predicting their cost and performance. Trained with reinforcement learning, SCOPE makes reasoning-based predictions by retrieving how models behave on similar problems, rather than relying on fixed model names, enabling it to work with new, unseen models. Moreover, by explicitly predicting how accurate and how expensive a model will be, it turns routing into a dynamic decision problem, allowing users to easily control the trade-off between accuracy and cost. Experiments show that SCOPE is more than just a cost-saving tool. It flexibly adapts to user needs: it can boost accuracy by up to 25.7% when performance is the priority, or cut costs by up to 95.1% when efficiency matters most.

Method: Proposes Stratified Importance Sampling (SIS) framework that combines importance sampling with stratification. Uses noisy proxies for stratification and importance weights, doesn’t require optimal proposal distributions or strata.

Result: Theoretical analysis shows SIS yields unbiased estimators with finite-sample MSE improvements over both importance sampling and stratified random sampling. Experiments across binary and multiclass tasks demonstrate consistent efficiency improvements under fixed label budgets.

Conclusion: SIS provides a principled, label-efficient, and operationally lightweight methodology for post-deployment model monitoring that addresses practical constraints in production environments.

Abstract: Monitoring the performance of classification models in production is critical yet challenging due to strict labeling budgets, one-shot batch acquisition of labels and extremely low error rates. We propose a general framework based on Stratified Importance Sampling (SIS) that directly addresses these constraints in model monitoring. While SIS has previously been applied in specialized domains, our theoretical analysis establishes its broad applicability to the monitoring of classification models. Under mild conditions, SIS yields unbiased estimators with strict finite-sample mean squared error (MSE) improvements over both importance sampling (IS) and stratified random sampling (SRS). The framework does not rely on optimally defined proposal distributions or strata: even with noisy proxies and sub-optimal stratification, SIS can improve estimator efficiency compared to IS or SRS individually, though extreme proposal mismatch may limit these gains. Experiments across binary and multiclass tasks demonstrate consistent efficiency improvements under fixed label budgets, underscoring SIS as a principled, label-efficient, and operationally lightweight methodology for post-deployment model monitoring.

Method: Proposes MolField, a hyper-network-based framework that models each molecule as a continuous function over 3D space (molecular field). Uses canonicalized coordinates for SE(3) invariance, structured weight tokenization, and trains a sequence-based hyper-network to model a shared prior over molecular fields.

Result: Evaluation on molecular dynamics and property prediction shows that treating molecules as continuous functions fundamentally changes how molecular representations generalize across tasks and yields downstream behavior that is stable to how molecules are discretized or queried.

Conclusion: Formulating molecular learning in function space by treating molecules as continuous functions over 3D space provides a more physically consistent representation that improves generalization and stability across molecular learning tasks.

Abstract: Molecular representations fundamentally shape how machine learning systems reason about molecular structure and physical properties. Most existing approaches adopt a discrete pipeline: molecules are encoded as sequences, graphs, or point clouds, mapped to fixed-dimensional embeddings, and then used for task-specific prediction. This paradigm treats molecules as discrete objects, despite their intrinsically continuous and field-like physical nature. We argue that molecular learning can instead be formulated as learning in function space. Specifically, we model each molecule as a continuous function over three-dimensional (3D) space and treat this molecular field as the primary object of representation. From this perspective, conventional molecular representations arise as particular sampling schemes of an underlying continuous object. We instantiate this formulation with MolField, a hyper-network-based framework that learns distributions over molecular fields. To ensure physical consistency, these functions are defined over canonicalized coordinates, yielding invariance to global SE(3) transformations. To enable learning directly over functions, we introduce a structured weight tokenization and train a sequence-based hyper-network to model a shared prior over molecular fields. We evaluate MolField on molecular dynamics and property prediction. Our results show that treating molecules as continuous functions fundamentally changes how molecular representations generalize across tasks and yields downstream behavior that is stable to how molecules are discretized or queried.

Method: MAAT formulates state reconstruction in reproducing kernel Hilbert space, incorporating structural priors (non-negativity, conservation laws, domain-specific models) into reconstruction objective while handling heterogeneous sampling and measurement granularity.

Result: Across twelve scientific benchmarks and multiple noise regimes, MAAT substantially reduces state-estimation MSE for trajectories and derivatives used by downstream symbolic regression compared to strong baselines.

Conclusion: MAAT provides a principled interface between fragmented sensor data and symbolic regression, enabling better recovery of governing equations from noisy, partial observations.

Abstract: Recovering governing equations from data is central to scientific discovery, yet existing methods often break down under noisy, partial observations, or rely on black-box latent dynamics that obscure mechanism. We introduce MAAT (Model Aware Approximation of Trajectories), a framework for symbolic discovery built on knowledge-informed Kernel State Reconstruction. MAAT formulates state reconstruction in a reproducing kernel Hilbert space and directly incorporates structural and semantic priors such as non-negativity, conservation laws, and domain-specific observation models into the reconstruction objective, while accommodating heterogeneous sampling and measurement granularity. This yields smooth, physically consistent state estimates with analytic time derivatives, providing a principled interface between fragmented sensor data and symbolic regression. Across twelve diverse scientific benchmarks and multiple noise regimes, MAAT substantially reduces state-estimation MSE for trajectories and derivatives used by downstream symbolic regression relative to strong baselines.

Method: BALANS constructs a sparse affinity matrix using batch-aware local scales (Gaussian kernel calibrated by k-th nearest neighbor distances) and adaptive sampling that prioritizes rows with low neighbor coverage, retaining only strongest affinities per row.

Result: BALANS scales to large collections with nearly linear time complexity, improves runtime over existing methods without sacrificing correction quality, and has proven optimal sample complexity with approximation guarantees.

Conclusion: BALANS provides an efficient, scalable batch-correction solution for large-scale Cell Painting data that maintains correction quality while dramatically improving computational efficiency.

Abstract: Cell Painting is a microscopy-based, high-content imaging assay that produces rich morphological profiles of cells and can support drug discovery by quantifying cellular responses to chemical perturbations. At scale, however, Cell Painting data is strongly affected by batch effects arising from differences in laboratories, instruments, and protocols, which can obscure biological signal. We present BALANS (Batch Alignment via Local Affinities and Subsampling), a scalable batch-correction method that aligns samples across batches by constructing a smoothed affinity matrix from pairwise distances. Given $n$ data points, BALANS builds a sparse affinity matrix $A \in \mathbb{R}^{n \times n}$ using two ideas. (i) For points $i$ and $j$, it sets a local scale using the distance from $i$ to its $k$-th nearest neighbor within the batch of $j$, then computes $A_{ij}$ via a Gaussian kernel calibrated by these batch-aware local scales. (ii) Rather than forming all $n^2$ entries, BALANS uses an adaptive sampling procedure that prioritizes rows with low cumulative neighbor coverage and retains only the strongest affinities per row, yielding a sparse but informative approximation of $A$. We prove that this sampling strategy is order-optimal in sample complexity and provides an approximation guarantee, and we show that BALANS runs in nearly linear time in $n$. Experiments on diverse real-world Cell Painting datasets and controlled large-scale synthetic benchmarks demonstrate that BALANS scales to large collections while improving runtime over native implementations of widely used batch-correction methods, without sacrificing correction quality.

Method: Proposes a noise correlation strategy that only correlates noise with the immediately preceding iteration and cancels a controlled portion using pseudorandom noise regeneration, eliminating the need to store past noise vectors

Result: Method requires no additional memory beyond standard DP-SGD, has minimal computational overhead, and empirically demonstrates improved accuracy over DP-SGD

Conclusion: Provides a practical, memory-efficient alternative to existing DP-SGD extensions with correlated noise while maintaining privacy guarantees and improving accuracy

Abstract: Differentially private stochastic gradient descent (DP-SGD) is the gold standard for training machine learning models with formal differential privacy guarantees. Several recent extensions improve its accuracy by introducing correlated noise across training iterations. Matrix factorization mechanisms are a prominent example, but they correlate noise across many iterations and require storing previously added noise vectors, leading to substantial memory overhead in some settings. In this work, we propose a new noise correlation strategy that correlates noise only with the immediately preceding iteration and cancels a controlled portion of it. Our method relies on noise regeneration using a pseudorandom noise generator, eliminating the need to store past noise. As a result, it requires no additional memory beyond standard DP-SGD. We show that the computational overhead is minimal and empirically demonstrate improved accuracy over DP-SGD.

Method: Derived an exact and analytical knowledge gradient acquisition function specifically for preferential Bayesian optimization. The method addresses the computational challenge of computing non-Gaussian posteriors in the look-ahead step that was previously considered intractable.

Result: The exact knowledge gradient performs strongly on a suite of benchmark problems, often outperforming existing acquisition functions. The paper also presents a case study illustrating limitations of the knowledge gradient in certain scenarios.

Conclusion: Successfully extended the knowledge gradient to preferential Bayesian optimization by deriving an exact analytical solution, overcoming previous computational barriers and demonstrating strong performance on benchmark problems.

Abstract: The knowledge gradient is a popular acquisition function in Bayesian optimization (BO) for optimizing black-box objectives with noisy function evaluations. Many practical settings, however, allow only pairwise comparison queries, yielding a preferential BO problem where direct function evaluations are unavailable. Extending the knowledge gradient to preferential BO is hindered by its computational challenge. At its core, the look-ahead step in the preferential setting requires computing a non-Gaussian posterior, which was previously considered intractable. In this paper, we address this challenge by deriving an exact and analytical knowledge gradient for preferential BO. We show that the exact knowledge gradient performs strongly on a suite of benchmark problems, often outperforming existing acquisition functions. In addition, we also present a case study illustrating the limitation of the knowledge gradient in certain scenarios.

Method: Integrates quantum-inspired reinforcement learning framework with controllable spin-chain analogy, real-time AIoT signals, and multi-objective reward function unifying fidelity, security, and carbon costs. Uses value-based and ensemble updates with window-normalized reward components for stabilized training.

Result: In simulation, the method exhibits smooth convergence, strong late-episode performance, and graceful degradation under representative noise channels, outperforming standard learned and model-based references.

Conclusion: Demonstrates potential for quantum-inspired AIoT frameworks to drive secure, eco-conscious supply chain operations at scale, laying groundwork for globally connected infrastructures that meet both consumer and environmental needs responsibly.

Abstract: Modern supply chains must balance high-speed logistics with environmental impact and security constraints, prompting a surge of interest in AI-enabled Internet of Things (AIoT) solutions for global commerce. However, conventional supply chain optimization models often overlook crucial sustainability goals and cyber vulnerabilities, leaving systems susceptible to both ecological harm and malicious attacks. To tackle these challenges simultaneously, this work integrates a quantum-inspired reinforcement learning framework that unifies carbon footprint reduction, inventory management, and cryptographic-like security measures. We design a quantum-inspired reinforcement learning framework that couples a controllable spin-chain analogy with real-time AIoT signals and optimizes a multi-objective reward unifying fidelity, security, and carbon costs. The approach learns robust policies with stabilized training via value-based and ensemble updates, supported by window-normalized reward components to ensure commensurate scaling. In simulation, the method exhibits smooth convergence, strong late-episode performance, and graceful degradation under representative noise channels, outperforming standard learned and model-based references, highlighting its robust handling of real-time sustainability and risk demands. These findings reinforce the potential for quantum-inspired AIoT frameworks to drive secure, eco-conscious supply chain operations at scale, laying the groundwork for globally connected infrastructures that responsibly meet both consumer and environmental needs.

Method: Introduces three parametric tasks with controllable exploration difficulty in continuous and discrete environments, tests state-of-the-art models against simple explore-exploit baselines, and studies two interventions: parallel budget splitting and periodic history summarization.

Result: Models show systematic under-exploration and suboptimal solutions, often performing significantly worse than simple baselines, with weak scaling as budget increases. Parallel execution surprisingly improves performance despite theoretical no-gain prediction, and history summarization preserves key discoveries and further enhances exploration.

Conclusion: Current language models have fundamental limitations in efficient exploration of interactive environments, but lightweight interventions like parallel execution and history summarization can mitigate these issues and improve performance.

Abstract: We evaluate language models on their ability to explore interactive environments under a limited interaction budget. We introduce three parametric tasks with controllable exploration difficulty, spanning continuous and discrete environments. Across state-of-the-art models, we find systematic under-exploration and suboptimal solutions, with performance often significantly worse than simple explore–exploit heuristic baselines and scaling weakly as the budget increases. Finally, we study two lightweight interventions: splitting a fixed budget into parallel executions, which surprisingly improves performance despite a no-gain theoretical result for our tasks, and periodically summarizing the interaction history, which preserves key discoveries and further improves exploration.

Method: 1) Presents first systematic, non-asymptotic analysis of outlier suppression for block Hadamard rotations, revealing that outlier suppression is limited by input vector geometry; 2) Introduces MixQuant framework that redistributes activation mass via permutations before rotation; 3) Develops greedy mass diffusion algorithm to calibrate permutations by equalizing expected blockwise ℓ₁ norms; 4) Identifies permutation-equivariant regions in transformer architectures to merge permutations into model weights without inference overhead.

Result: MixQuant consistently improves accuracy across all block sizes, recovering up to 90% of the full-vector rotation perplexity when quantizing Llama3 1B to INT4 with block size 16, compared to 46% without permutations.

Conclusion: The paper provides fundamental insights into block rotation-based outlier suppression and introduces an effective permutation-based framework that significantly improves quantization accuracy for large language models without adding inference overhead.

Abstract: Recent post-training quantization (PTQ) methods have adopted block rotations to diffuse outliers prior to rounding. While this reduces the overhead of full-vector rotations, the effect of block structure on outlier suppression remains poorly understood. To fill this gap, we present the first systematic, non-asymptotic analysis of outlier suppression for block Hadamard rotations. Our analysis reveals that outlier suppression is fundamentally limited by the geometry of the input vector. In particular, post-rotation outliers are deterministically minimized when the pre-rotation $\ell_1$ norm mass is evenly distributed across blocks. Guided by these insights, we introduce MixQuant, a block rotation-aware PTQ framework that redistributes activation mass via permutations prior to rotation. We propose a greedy mass diffusion algorithm to calibrate permutations by equalizing the expected blockwise $\ell_1$ norms. To avoid adding inference overhead, we identify permutation-equivariant regions in transformer architectures to merge the resulting permutations into model weights before deployment. Experiments show that MixQuant consistently improves accuracy across all block sizes, recovering up to 90% of the full-vector rotation perplexity when quantizing Llama3 1B to INT4 with block size 16, compared to 46% without permutations.

Method: Model policy representations as expectations of state-action feature maps with respect to occupancy measures. Use set-based architecture to encode state-action samples into latent embeddings, decode both policies and value functions. Employ variational generative approach for smooth latent space and contrastive learning to align latent distances with value function differences.

Result: The method creates a smooth latent space where gradient-based optimization can be performed directly, enabling novel behavior synthesis where policies can be steered to satisfy previously unseen value function constraints without additional training.

Conclusion: The proposed framework successfully learns policy representations that support gradient-based behavior steering in latent space, allowing for flexible adaptation to new value function constraints at test time without retraining.

Abstract: Given a Markov decision process (MDP), we seek to learn representations for a range of policies to facilitate behavior steering at test time. As policies of an MDP are uniquely determined by their occupancy measures, we propose modeling policy representations as expectations of state-action feature maps with respect to occupancy measures. We show that these representations can be approximated uniformly for a range of policies using a set-based architecture. Our model encodes a set of state-action samples into a latent embedding, from which we decode both the policy and its value functions corresponding to multiple rewards. We use variational generative approach to induce a smooth latent space, and further shape it with contrastive learning so that latent distances align with differences in value functions. This geometry permits gradient-based optimization directly in the latent space. Leveraging this capability, we solve a novel behavior synthesis task, where policies are steered to satisfy previously unseen value function constraints without additional training.

Method: Formalizes recoverability through Expected Recovery Regret (ERR) which quantifies deviation from optimal recovery policy under stochastic execution noise. Derives first-order relationship between ERR and empirical Efficiency Score (ES), creating a falsifiable quantitative law of recovery dynamics. Empirically validates across five tool-use benchmarks with controlled perturbations, diagnostic reasoning, and real-world APIs.

Result: Predicted regret under the ERR-ES law closely matched observed post-failure regret measured from Monte Carlo rollouts (delta ≤ 0.05) across model scales, perturbation regimes, and recovery horizons. Recoverability is shown to be a governed property of interaction dynamics rather than artifact of model scale or architecture.

Conclusion: Recoverability in language agents follows a measurable law, providing theoretical foundation for execution-level robustness. The ERR-ES framework offers predictive understanding of self-recovery capabilities in tool-using agents.

Abstract: Language model agents often appear capable of self-recovery after failing tool call executions, yet this behavior lacks a formal explanation. We present a predictive theory that resolves this gap by showing that recoverability follows a measurable law. To elaborate, we formalize recoverability through Expected Recovery Regret (ERR), which quantifies the deviation of a recovery policy from the optimal one under stochastic execution noise, and derive a first-order relationship between ERR and an empirical observable quantity, the Efficiency Score (ES). This yields a falsifiable first-order quantitative law of recovery dynamics in tool-using agents. We empirically validate the law across five tool-use benchmarks spanning controlled perturbations, diagnostic reasoning, and real-world APIs. Across model scales, perturbation regimes, and recovery horizons, predicted regret under the ERR-ES law closely matched observed post-failure regret measured from Monte Carlo rollouts, within delta less than or equal to 0.05. Our results reveal that recoverability is not an artifact of model scale or architecture, but a governed property of interaction dynamics, providing a theoretical foundation for execution-level robustness in language agents.

Method: Exploits cone geometry of relative translation invariant quadratic Wasserstein space, introduces relative Wasserstein angle and orthogonal projection distance, proves flatness of filling cones, derives closed-form expressions in 1D, and develops stochastic manifold optimization algorithm for high dimensions.

Result: Shows that moment-matching Gaussian is not the W₂-nearest Gaussian, demonstrates relative Wasserstein angle is more robust than Wasserstein distance, and that proposed nearest Gaussian provides better approximation than moment matching in FID score evaluation.

Conclusion: The geometric framework provides meaningful measures of non-Gaussianity and reveals limitations of traditional moment-matching approaches, offering improved Gaussian approximations for distribution analysis.

Abstract: We study the problem of quantifying how far an empirical distribution deviates from Gaussianity under the framework of optimal transport. By exploiting the cone geometry of the relative translation invariant quadratic Wasserstein space, we introduce two novel geometric quantities, the relative Wasserstein angle and the orthogonal projection distance, which provide meaningful measures of non-Gaussianity. We prove that the filling cone generated by any two rays in this space is flat, ensuring that angles, projections, and inner products are rigorously well-defined. This geometric viewpoint recasts Gaussian approximation as a projection problem onto the Gaussian cone and reveals that the commonly used moment-matching Gaussian can \emph{not} be the (W_2)-nearest Gaussian for a given empirical distribution. In one dimension, we derive closed-form expressions for the proposed quantities and extend them to several classical distribution families, including uniform, Laplace, and logistic distributions; while in high dimensions, we develop an efficient stochastic manifold optimization algorithm based on a semi-discrete dual formulation. Experiments on synthetic data and real-world feature distributions demonstrate that the relative Wasserstein angle is more robust than the Wasserstein distance and that the proposed nearest Gaussian provides a better approximation than moment matching in the evaluation of Fréchet Inception Distance (FID) scores.

Method: Formalizes safety constraints as a partially ordered set (poset) and treats safety composition as a structural property of the policy class. Proposes PoSafeNet, a differentiable neural safety layer that enforces safety via sequential closed-form projection under poset-consistent constraint orderings, enabling adaptive selection or mixing of valid safety executions.

Result: Experiments on multi-obstacle navigation, constrained robot manipulation, and vision-based autonomous driving demonstrate improved feasibility, robustness, and scalability over unstructured and differentiable quadratic program-based safety layers.

Conclusion: The poset-structured safety formulation and PoSafeNet provide a principled approach to handling heterogeneous safety constraints with partial priority relations, enabling more flexible and robust safe learning for robotic systems.

Abstract: Safe learning is essential for deploying learningbased controllers in safety-critical robotic systems, yet existing approaches often enforce multiple safety constraints uniformly or via fixed priority orders, leading to infeasibility and brittle behavior. In practice, safety requirements are heterogeneous and admit only partial priority relations, where some constraints are comparable while others are inherently incomparable. We formalize this setting as poset-structured safety, modeling safety constraints as a partially ordered set and treating safety composition as a structural property of the policy class. Building on this formulation, we propose PoSafeNet, a differentiable neural safety layer that enforces safety via sequential closed-form projection under poset-consistent constraint orderings, enabling adaptive selection or mixing of valid safety executions while preserving priority semantics by construction. Experiments on multi-obstacle navigation, constrained robot manipulation, and vision-based autonomous driving demonstrate improved feasibility, robustness, and scalability over unstructured and differentiable quadratic program-based safety layers.

Method: The researchers use real-world conversation traces and fine-grained energy measurements to analyze how input length, output length, and model size affect energy consumption in LLM interactions, with polite messages serving as a reproducible test case.

Result: The study provides quantified energy costs of polite messages and reveals how different factors (input/output length, model size) impact energy consumption, offering actionable insights for building more sustainable LLM applications.

Conclusion: Understanding and mitigating the energy cost of LLM interactions is crucial for sustainable AI deployment, especially as user adoption grows and billions of prompts are processed daily in real-world contexts like chat applications.

Abstract: Being polite is free - or is it? In this paper, we quantify the energy cost of seemingly innocuous messages such as ``thank you’’ when interacting with large language models, often used by users to convey politeness. Using real-world conversation traces and fine-grained energy measurements, we quantify how input length, output length and model size affect energy use. While politeness is our motivating example, it also serves as a controlled and reproducible proxy for measuring the energy footprint of a typical LLM interaction. Our findings provide actionable insights for building more sustainable and efficient LLM applications, especially in increasingly widespread real-world contexts like chat. As user adoption grows and billions of prompts are processed daily, understanding and mitigating this cost becomes crucial - not just for efficiency, but for sustainable AI deployment.

Method: The authors formalize residual knowledge vulnerability, prove its inevitability in high-dimensional settings, and propose RURK (Residual Unlearning with Robust Knowledge) - a fine-tuning strategy that penalizes the model’s ability to re-recognize perturbed forget samples.

Result: Experiments on vision benchmarks with deep neural networks show that residual knowledge is prevalent across existing unlearning methods, and the proposed RURK approach effectively prevents this vulnerability.

Conclusion: Machine unlearning methods need stronger privacy guarantees against adversarial perturbations, and the proposed RURK framework provides an effective solution to mitigate residual knowledge risks in unlearned models.

Abstract: Machine unlearning offers a practical alternative to avoid full model re-training by approximately removing the influence of specific user data. While existing methods certify unlearning via statistical indistinguishability from re-trained models, these guarantees do not naturally extend to model outputs when inputs are adversarially perturbed. In particular, slight perturbations of forget samples may still be correctly recognized by the unlearned model - even when a re-trained model fails to do so - revealing a novel privacy risk: information about the forget samples may persist in their local neighborhood. In this work, we formalize this vulnerability as residual knowledge and show that it is inevitable in high-dimensional settings. To mitigate this risk, we propose a fine-tuning strategy, named RURK, that penalizes the model’s ability to re-recognize perturbed forget samples. Experiments on vision benchmarks with deep neural networks demonstrate that residual knowledge is prevalent across existing unlearning methods and that our approach effectively prevents residual knowledge.

Method: Detailed empirical study of LLM inference energy and latency on NVIDIA H100 GPUs, analyzing impact of quantization, batch size, and serving configuration using Hugging Face’s Text Generation Inference server.

Result: Lower-precision formats only yield energy gains in compute-bound regimes; batching improves energy efficiency especially in memory-bound phases like decoding; structured request timing (arrival shaping) can reduce per-request energy by up to 100 times.

Conclusion: Sustainable LLM deployment depends not only on model internals but also on orchestration of the serving stack, motivating phase-aware energy profiling and system-level optimizations for greener AI services.

Abstract: Large Language Models (LLMs) are increasingly deployed in production, contributing towards shifting the burden in terms of computational resources and energy demands from training to inference. While prior work has examined the energy cost of inference per prompt or per token, we highlight how \emph{system-level design choices} - such as numerical precision, batching strategy, and request scheduling - can lead to orders-of-magnitude differences in energy consumption for the same model. We perform a detailed empirical study of LLM inference energy and latency on NVIDIA H100 GPUs, analyzing the impact of quantization, batch size, and serving configuration (e.g., with Hugging Face’s Text Generation Inference server). Our results reveal that lower-precision formats only yield energy gains in compute-bound regimes; that batching improves energy efficiency, especially in memory-bound phases like decoding; and that structured request timing (arrival shaping) can reduce per-request energy by up to 100 times. We argue that sustainable LLM deployment depends not only on model internals, but also on the orchestration of the serving stack. Our findings motivate phase-aware energy profiling and system-level optimizations for greener AI services.

Method: FIRE couples tabular foundation models to perform zero-shot in-context Bayesian inference via a high-fidelity correction model conditioned on the low-fidelity model’s posterior predictive distributions, enabling cross-fidelity information transfer via distributional summaries.

Result: Across 31 benchmark problems including synthetic and real-world tasks (DrivAerNet, LCBench), FIRE delivers stronger performance-time trade-off than seven state-of-the-art GP-based or deep learning MF regression methods, ranking highest in accuracy and uncertainty quantification with runtime advantages.

Conclusion: FIRE provides an effective training-free approach for multi-fidelity regression with better efficiency and generalization than traditional methods, though limitations include context window constraints and dependence on pre-trained TFM quality.

Abstract: Multi-fidelity (MF) regression often operates in regimes of extreme data imbalance, where the commonly-used Gaussian-process (GP) surrogates struggle with cubic scaling costs and overfit to sparse high-fidelity observations, limiting efficiency and generalization in real-world applications. We introduce FIRE, a training-free MF framework that couples tabular foundation models (TFMs) to perform zero-shot in-context Bayesian inference via a high-fidelity correction model conditioned on the low-fidelity model’s posterior predictive distributions. This cross-fidelity information transfer via distributional summaries captures heteroscedastic errors, enabling robust residual learning without model retraining. Across 31 benchmark problems spanning synthetic and real-world tasks (e.g., DrivAerNet, LCBench), FIRE delivers a stronger performance-time trade-off than seven state-of-the-art GP-based or deep learning MF regression methods, ranking highest in accuracy and uncertainty quantification with runtime advantages. Limitations include context window constraints and dependence on the quality of the pre-trained TFM’s.

Method: PABLO uses scientific LLMs pretrained on chemistry/biology literature in a hierarchical agentic system to generate and iteratively refine biological candidates through language-based reasoning.

Result: Achieves state-of-the-art performance on GuacaMol molecular design and antimicrobial peptide optimization, with improved sample efficiency and final objective values. PABLO-optimized peptides showed strong activity against drug-resistant pathogens in vitro.

Conclusion: Agentic formulation offers advantages for realistic biological design including semantic task descriptions, retrieval-augmented knowledge, and complex constraints, showing practical potential for therapeutic discovery.

Abstract: Many key challenges in biological design-such as small-molecule drug discovery, antimicrobial peptide development, and protein engineering-can be framed as black-box optimization over vast, complex structured spaces. Existing methods rely mainly on raw structural data and struggle to exploit the rich scientific literature. While large language models (LLMs) have been added to these pipelines, they have been confined to narrow roles within structure-centered optimizers. We instead cast biological black-box optimization as a fully agentic, language-based reasoning process. We introduce Purely Agentic BLack-box Optimization (PABLO), a hierarchical agentic system that uses scientific LLMs pretrained on chemistry and biology literature to generate and iteratively refine biological candidates. On both the standard GuacaMol molecular design and antimicrobial peptide optimization tasks, PABLO achieves state-of-the-art performance, substantially improving sample efficiency and final objective values over established baselines. Compared to prior optimization methods that incorporate LLMs, PABLO achieves competitive token usage per run despite relying on LLMs throughout the optimization loop. Beyond raw performance, the agentic formulation offers key advantages for realistic design: it naturally incorporates semantic task descriptions, retrieval-augmented domain knowledge, and complex constraints. In follow-up in vitro validation, PABLO-optimized peptides showed strong activity against drug-resistant pathogens, underscoring the practical potential of PABLO for therapeutic discovery.

Method: G-Substrate framework with two key mechanisms: (1) a unified structural schema ensuring compatibility of graph representations across heterogeneous modalities and tasks, and (2) an interleaved role-based training strategy that exposes the same graph structure to multiple functional roles during learning.

Result: Experiments across multiple domains, modalities, and tasks show that G-Substrate outperforms both task-isolated learning methods and naive multi-task learning approaches.

Conclusion: Treating graph structure as a persistent substrate that accumulates knowledge across learning contexts enables more effective representation learning that transfers structural regularities across heterogeneous modalities and tasks.

Abstract: Graphs provide a natural representation of relational structure that arises across diverse domains. Despite this ubiquity, graph structure is typically learned in a modality- and task-isolated manner, where graph representations are constructed within individual task contexts and discarded thereafter. As a result, structural regularities across modalities and tasks are repeatedly reconstructed rather than accumulated at the level of intermediate graph representations. This motivates a representation-learning question: how should graph structure be organized so that it can persist and accumulate across heterogeneous modalities and tasks? We adopt a representation-centric perspective in which graph structure is treated as a structural substrate that persists across learning contexts. To instantiate this perspective, we propose G-Substrate, a graph substrate framework that organizes learning around shared graph structures. G-Substrate comprises two complementary mechanisms: a unified structural schema that ensures compatibility among graph representations across heterogeneous modalities and tasks, and an interleaved role-based training strategy that exposes the same graph structure to multiple functional roles during learning. Experiments across multiple domains, modalities, and tasks show that G-Substrate outperforms task-isolated and naive multi-task learning methods.

Method: SAIR uses an LLM as an in-context reinforcement learning controller that improves its policy online from reward-labeled interaction histories without gradient updates. It combines Pareto-dominance reward shaping with provable separation margin, surprisal-guided experience retrieval for context efficiency, and fine-grained GPU rate control via user-space CUDA interception.

Result: On four ML serving pipelines under three workload patterns, SAIR achieves the best or tied-best P99 latency and effective resource cost among deployed baselines, improving P99 latency by up to 50% and reducing effective cost by up to 97% (under GPU rate-control assumptions), with 86% bottleneck detection accuracy and no offline training.

Conclusion: SAIR demonstrates that LLMs can effectively serve as in-context RL controllers for complex autoscaling problems in ML inference pipelines, achieving significant performance improvements without requiring offline training or gradient updates.

Abstract: Multi-stage ML inference pipelines are difficult to autoscale due to heterogeneous resources, cross-stage coupling, and dynamic bottleneck migration. We present SAIR, an autoscaling framework that uses an LLM as an in-context reinforcement learning controller, improving its policy online from reward-labeled interaction histories without gradient updates. SAIR combines Pareto-dominance reward shaping with a provable separation margin, surprisal-guided experience retrieval for context efficiency, and fine-grained GPU rate control via user-space CUDA interception. We provide regret analysis decomposing error into retrieval coverage and LLM selection components. On four ML serving pipelines under three workload patterns, SAIR achieves the best or tied-best P99 latency and effective resource cost among deployed baselines, improving P99 by up to 50% and reducing effective cost by up to 97% (under GPU rate-control assumptions), with 86% bottleneck detection accuracy and no offline training.

Method: SIREN attributes root causes by estimating score functions of data likelihood and computing attribution via integrated gradients that accumulate score contributions along paths from outliers toward normal data distributions.

Result: Extensive experiments on synthetic random graphs and real-world cloud service and supply chain datasets show SIREN outperforms state-of-the-art baselines in both attribution accuracy and computational efficiency.

Conclusion: SIREN provides a tractable and uncertainty-aware approach for root cause attribution in nonlinear, high-dimensional, and heteroscedastic causal models, satisfying key mathematical axioms.

Abstract: Identifying the root causes of outliers is a fundamental problem in causal inference and anomaly detection. Traditional approaches based on heuristics or counterfactual reasoning often struggle under uncertainty and high-dimensional dependencies. We introduce SIREN, a novel and scalable method that attributes the root causes of outliers by estimating the score functions of the data likelihood. Attribution is computed via integrated gradients that accumulate score contributions along paths from the outlier toward the normal data distribution. Our method satisfies three of the four classic Shapley value axioms - dummy, efficiency, and linearity - as well as an asymmetry axiom derived from the underlying causal structure. Unlike prior work, SIREN operates directly on the score function, enabling tractable and uncertainty-aware root cause attribution in nonlinear, high-dimensional, and heteroscedastic causal models. Extensive experiments on synthetic random graphs and real-world cloud service and supply chain datasets show that SIREN outperforms state-of-the-art baselines in both attribution accuracy and computational efficiency.

Method: Analyzes gradient descent for training two-layer KANs, deriving general bounds for training dynamics, generalization, and utility under differential privacy. Specializes analysis to logistic loss under NTK-separable assumption, establishing theoretical results for optimization rates, generalization rates, and privacy-utility tradeoffs.

Result: Shows polylogarithmic network width suffices for GD to achieve optimization rate of O(1/T) and generalization rate of O(1/n). In private setting, obtains utility bound of O(√d/(nε)) matching classical lower bounds. Reveals polylogarithmic width is both sufficient and necessary under differential privacy, creating qualitative gap between private and non-private regimes.

Conclusion: Provides first comprehensive theoretical analysis of KANs’ training dynamics, generalization, and privacy properties. Theoretical insights can guide practical choices like network width selection and early stopping. Reveals fundamental differences between private and non-private training regimes for KANs.

Abstract: Kolmogorov–Arnold Networks (KANs) have recently emerged as a structured alternative to standard MLPs, yet a principled theory for their training dynamics, generalization, and privacy properties remains limited. In this paper, we analyze gradient descent (GD) for training two-layer KANs and derive general bounds that characterize their training dynamics, generalization, and utility under differential privacy (DP). As a concrete instantiation, we specialize our analysis to logistic loss under an NTK-separable assumption, where we show that polylogarithmic network width suffices for GD to achieve an optimization rate of order $1/T$ and a generalization rate of order $1/n$, with $T$ denoting the number of GD iterations and $n$ the sample size. In the private setting, we characterize the noise required for $(ε,δ)$-DP and obtain a utility bound of order $\sqrt{d}/(nε)$ (with $d$ the input dimension), matching the classical lower bound for general convex Lipschitz problems. Our results imply that polylogarithmic width is not only sufficient but also necessary under differential privacy, revealing a qualitative gap between non-private (sufficiency only) and private (necessity also emerges) training regimes. Experiments further illustrate how these theoretical insights can guide practical choices, including network width selection and early stopping.

Method: Developed MM-OpenFGL benchmark with 19 multimodal datasets across 7 domains, 8 simulation strategies for modality/topology variations, 6 downstream tasks, and 57 state-of-the-art methods via modular API.

Result: Extensive experiments investigate multimodal federated graph learning from necessity, effectiveness, robustness, and efficiency perspectives, providing valuable insights for future research.

Conclusion: MM-OpenFGL bridges the gap in multimodal federated graph learning research by providing the first comprehensive benchmark for systematic evaluation and future development.

Abstract: Multimodal-attributed graphs (MMAGs) provide a unified framework for modeling complex relational data by integrating heterogeneous modalities with graph structures. While centralized learning has shown promising performance, MMAGs in real-world applications are often distributed across isolated platforms and cannot be shared due to privacy concerns or commercial constraints. Federated graph learning (FGL) offers a natural solution for collaborative training under such settings; however, existing studies largely focus on single-modality graphs and do not adequately address the challenges unique to multimodal federated graph learning (MMFGL). To bridge this gap, we present MM-OpenFGL, the first comprehensive benchmark that systematically formalizes the MMFGL paradigm and enables rigorous evaluation. MM-OpenFGL comprises 19 multimodal datasets spanning 7 application domains, 8 simulation strategies capturing modality and topology variations, 6 downstream tasks, and 57 state-of-the-art methods implemented through a modular API. Extensive experiments investigate MMFGL from the perspectives of necessity, effectiveness, robustness, and efficiency, offering valuable insights for future research on MMFGL.

Method: Created MetaLead, a fully human-annotated ML leaderboard dataset that captures ALL experimental results (not just best ones), includes extra metadata like experimental types (baseline, proposed method, variations), and explicitly separates train and test datasets for cross-domain assessment.

Result: MetaLead provides a powerful resource for more transparent and nuanced evaluations across ML research by offering complete result transparency, experiment-type guided comparisons, and cross-domain assessment capabilities.

Conclusion: MetaLead addresses limitations of existing leaderboard datasets by providing comprehensive experimental results with rich metadata, enabling more transparent, nuanced, and cross-domain ML evaluations through its enriched structure.

Abstract: Leaderboards are crucial in the machine learning (ML) domain for benchmarking and tracking progress. However, creating leaderboards traditionally demands significant manual effort. In recent years, efforts have been made to automate leaderboard generation, but existing datasets for this purpose are limited by capturing only the best results from each paper and limited metadata. We present MetaLead, a fully human-annotated ML Leaderboard dataset that captures all experimental results for result transparency and contains extra metadata, such as the result experimental type: baseline, proposed method, or variation of proposed method for experiment-type guided comparisons, and explicitly separates train and test dataset for cross-domain assessment. This enriched structure makes MetaLead a powerful resource for more transparent and nuanced evaluations across ML research.

Method: Combines counterfactual data augmentation with contrastive learning. Uses dynamic treatments design with structural neighborhood exploration to generate high-quality counterfactual data that quantifies temporal interaction pattern changes.

Result: Extensive experiments on multiple real-world datasets show CoDCL significantly improves state-of-the-art baseline models in dynamic network prediction tasks.

Conclusion: Integrating counterfactual data augmentation into dynamic representation learning plays a critical role in improving model performance and robustness to structural changes in dynamic networks.

Abstract: The rapid growth and continuous structural evolution of dynamic networks make effective predictions increasingly challenging. To enable prediction models to adapt to complex temporal environments, they need to be robust to emerging structural changes. We propose a dynamic network learning framework CoDCL, which combines counterfactual data augmentation with contrastive learning to address this deficiency.Furthermore, we devise a comprehensive strategy to generate high-quality counterfactual data, combining a dynamic treatments design with efficient structural neighborhood exploration to quantify the temporal changes in interaction patterns.Crucially, the entire CoDCL is designed as a plug-and-play universal module that can be seamlessly integrated into various existing temporal graph models without requiring architectural modifications.Extensive experiments on multiple real-world datasets demonstrate that CoDCL significantly gains state-of-the-art baseline models in the field of dynamic networks, confirming the critical role of integrating counterfactual data augmentation into dynamic representation learning.

Method: Instead of estimating advantages, the method bifurcates K outcomes into positive and negative sets, then maximizes the likelihood of positive outcomes. This is framed as an online instantiation of (multi-label) noise contrastive estimation for LLM reasoning.

Result: Demonstrates competitive performance on challenging math benchmarks against strong baselines like DAPO and online DPO.

Conclusion: The explicit contrastive learning approach provides a simpler alternative to GRPO’s advantage estimation while maintaining competitive performance on reasoning tasks.

Abstract: GRPO is a standard approach to endowing pretrained LLMs with reasoning capabilities. It estimates the advantage of an outcome from a group of $K$ outcomes, and promotes those with positive advantages inside a trust region. Since GRPO discriminates between good and bad outcomes softly, it benefits from additional refinements such as asymmetric clipping and zero-variance data filtering. While effective, these refinements require significant empirical insight and can be challenging to identify. We instead propose an explicit contrastive learning approach. Instead of estimating advantages, we bifurcate $K$ outcomes into positive and negative sets, then maximize the likelihood of positive outcomes. Our approach can be viewed as an online instantiation of (multi-label) noise contrastive estimation for LLM reasoning. We validate our method by demonstrating competitive performance on a suite of challenging math benchmarks against strong baselines such as DAPO and online DPO.

Method: Introduces asynchronous updates across both parallelism axes: 1) For pipeline parallelism, uses weight look-ahead approach; 2) For data parallelism, introduces asynchronous sparse averaging with exponential moving average based correction mechanism.

Result: Experiments on large-scale language models (up to 1B parameters) show the approach matches performance of fully synchronous baseline while significantly reducing communication overhead.

Conclusion: The proposed asynchronous methods effectively address the communication bottleneck in distributed training while maintaining model performance, enabling more scalable training without requiring co-located clusters.

Abstract: Data and pipeline parallelism are key strategies for scaling neural network training across distributed devices, but their high communication cost necessitates co-located computing clusters with fast interconnects, limiting their scalability. We address this communication bottleneck by introducing asynchronous updates across both parallelism axes, relaxing the co-location requirement at the expense of introducing staleness between pipeline stages and data parallel replicas. To mitigate staleness, for pipeline parallelism, we adopt a weight look-ahead approach, and for data parallelism, we introduce an asynchronous sparse averaging method equipped with an exponential moving average based correction mechanism. We provide convergence guarantees for both sparse averaging and asynchronous updates. Experiments on large-scale language models (up to \em 1B parameters) demonstrate that our approach matches the performance of the fully synchronous baseline, while significantly reducing communication overhead.

Method: Through extensive experiments studying when weak priors succeed or fail, combined with theoretical analysis based on Bayesian consistency to determine conditions under which high-dimensional measurements make the posterior concentrate near the true signal.

Result: Weak priors succeed when measurements are highly informative (e.g., many observed pixels) and fail in certain regimes. The theory provides principled justification for when weak diffusion priors can be reliably used.

Conclusion: Diffusion models trained on mismatched data can serve as effective priors for inverse problems under specific conditions, particularly when measurements provide sufficient information about the target signal.

Abstract: Can a diffusion model trained on bedrooms recover human faces? Diffusion models are widely used as priors for inverse problems, but standard approaches usually assume a high-fidelity model trained on data that closely match the unknown signal. In practice, one often must use a mismatched or low-fidelity diffusion prior. Surprisingly, these weak priors often perform nearly as well as full-strength, in-domain baselines. We study when and why inverse solvers are robust to weak diffusion priors. Through extensive experiments, we find that weak priors succeed when measurements are highly informative (e.g., many observed pixels), and we identify regimes where they fail. Our theory, based on Bayesian consistency, gives conditions under which high-dimensional measurements make the posterior concentrate near the true signal. These results provide a principled justification on when weak diffusion priors can be used reliably.

Method: Developed a system using LLM-powered web research agents to automatically generate and resolve forecasting questions at scale. Generated 1499 diverse real-world questions and resolved them months later. Evaluated question quality and resolution accuracy, and tested forecasting performance of different LLMs.

Result: System produces verifiable, unambiguous questions ~96% of the time (exceeding Metaculus), resolves questions at ~95% accuracy. More intelligent LLMs perform better (Brier scores: Gemini 3 Pro 0.134, GPT-5 0.149, Gemini 2.5 Flash 0.179). Question decomposition strategy improved Brier scores (0.132 vs 0.141).

Conclusion: LLM-powered agents can automate high-quality forecasting question generation and resolution at scale, producing diverse real-world questions with high verifiability and resolution accuracy, enabling better evaluation and improvement of forecasting systems.

Abstract: Forecasting future events is highly valuable in decision-making and is a robust measure of general intelligence. As forecasting is probabilistic, developing and evaluating AI forecasters requires generating large numbers of diverse and difficult questions, and accurately resolving them. Previous efforts to automate this laborious work relied on recurring data sources (e.g., weather, stocks), limiting diversity and utility. In this work, we present a system for generating and resolving high-quality forecasting questions automatically and at scale using LLM-powered web research agents. We use this system to generate 1499 diverse, real-world forecasting questions, and to resolve them several months later. We estimate that our system produces verifiable, unambiguous questions approximately 96% of the time, exceeding the rate of Metaculus, a leading human-curated forecasting platform. We also find that our system resolves questions at approximately 95% accuracy. We verify that forecasting agents powered by more intelligent LLMs perform better on these questions (Brier score of 0.134 for Gemini 3 Pro, 0.149 for GPT-5, and 0.179 for Gemini 2.5 Flash). Finally, we demonstrate how our system can be leveraged to directly improve forecasting, by evaluating a question decomposition strategy on a generated question set, yielding a significant improvement in Brier scores (0.132 vs. 0.141).

Method: Novel weight-based interpretation framework that measures functional effects through direct weight interactions without requiring activation data

Result: 1/4 of features directly predict output tokens; features actively participate in attention mechanisms with depth-dependent structure; semantic and non-semantic features show distinct distribution profiles in attention circuits

Conclusion: Provides the missing out-of-context half of SAE feature interpretability by analyzing functional roles through weight interactions

Abstract: Sparse autoencoders (SAEs) have emerged as a powerful technique for decomposing language model representations into interpretable features. Current interpretation methods infer feature semantics from activation patterns, but overlook that features are trained to reconstruct activations that serve computational roles in the forward pass. We introduce a novel weight-based interpretation framework that measures functional effects through direct weight interactions, requiring no activation data. Through three experiments on Gemma-2 and Llama-3.1 models, we demonstrate that (1) 1/4 of features directly predict output tokens, (2) features actively participate in attention mechanisms with depth-dependent structure, and (3) semantic and non-semantic feature populations exhibit distinct distribution profiles in attention circuits. Our analysis provides the missing out-of-context half of SAE feature interpretability.

Method: HeaPA maintains a bounded, evolving prompt pool using heap-based boundary sampling to track the capability frontier, expands the pool via on-policy augmentation with asynchronous validation, and stabilizes queries through topology-aware re-estimation and controlled reinsertion.

Result: Across two training corpora, two training recipes, and seven benchmarks, HeaPA consistently improves accuracy and reaches target performance with fewer computations while maintaining comparable wall-clock time.

Conclusion: HeaPA’s frontier-focused sampling and on-policy pool growth significantly improve RLVR training efficiency, with benefits scaling with model size, offering a practical solution for efficient reasoning task training.

Abstract: RLVR is now a standard way to train LLMs on reasoning tasks with verifiable outcomes, but when rollout generation dominates the cost, efficiency depends heavily on which prompts you sample and when. In practice, prompt pools are often static or only loosely tied to the model’s learning progress, so uniform sampling can’t keep up with the shifting capability frontier and ends up wasting rollouts on prompts that are already solved or still out of reach. Existing approaches improve efficiency through filtering, curricula, adaptive rollout allocation, or teacher guidance, but they typically assume a fixed pool-which makes it hard to support stable on-policy pool growth-or they add extra teacher cost and latency. We introduce HeaPA (Heap Sampling and On-Policy Query Augmentation), which maintains a bounded, evolving pool, tracks the frontier using heap-based boundary sampling, expands the pool via on-policy augmentation with lightweight asynchronous validation, and stabilizes correlated queries through topology-aware re-estimation of pool statistics and controlled reinsertion. Across two training corpora, two training recipes, and seven benchmarks, HeaPA consistently improves accuracy and reaches target performance with fewer computations while keeping wall-clock time comparable. Our analyses suggest these gains come from frontier-focused sampling and on-policy pool growth, with the benefits becoming larger as model scale increases. Our code is available at https://github.com/horizon-rl/HeaPA.

Method: Theoretical decomposition of Masked Diffusion objective into Signal (feature learning) and Noise (implicit regularization) regimes. Training nanoGPT with MD objective on k-parity problems, and optimizing mask probability distribution based on theoretical insights.

Result: MD objective fundamentally alters learning landscape, enabling rapid simultaneous generalization without grokking. Optimized mask distribution improves perplexity for 50M-parameter models and achieves 8.8% and 5.8% performance gains on 8B-parameter models for pre-training and fine-tuning respectively.

Conclusion: Masked Diffusion Language Models offer superior generalization properties compared to auto-regressive models, with theoretical insights enabling practical optimizations that scale effectively to large models.

Abstract: Masked Diffusion Language Models have recently emerged as a powerful generative paradigm, yet their generalization properties remain understudied compared to their auto-regressive counterparts. In this work, we investigate these properties within the setting of the $k$-parity problem (computing the XOR sum of $k$ relevant bits), where neural networks typically exhibit grokking – a prolonged plateau of chance-level performance followed by sudden generalization. We theoretically decompose the Masked Diffusion (MD) objective into a Signal regime which drives feature learning, and a Noise regime which serves as an implicit regularizer. By training nanoGPT using MD objective on the $k$-parity problem, we demonstrate that MD objective fundamentally alters the learning landscape, enabling rapid and simultaneous generalization without experiencing grokking. Furthermore, we leverage our theoretical insights to optimize the distribution of the mask probability in the MD objective. Our method significantly improves perplexity for 50M-parameter models and achieves superior results across both pre-training from scratch and supervised fine-tuning. Specifically, we observe performance gains peaking at $8.8%$ and $5.8%$, respectively, on 8B-parameter models, confirming the scalability and effectiveness of our framework in large-scale masked diffusion language model regimes.

Method: TGPM conceptualizes interactions as interaction patches via temporally-biased random walks to capture multi-scale structural semantics and long-range dependencies. Uses Transformer-based backbone for global temporal regularities, with self-supervised pre-training tasks (masked token modeling and next-time prediction) to encode network evolution laws.

Result: TGPM consistently achieves state-of-the-art performance in both transductive and inductive link prediction, demonstrating exceptional cross-domain transferability.

Conclusion: TGPM provides a foundation framework that shifts focus from task-centric approaches to learning generalized evolving patterns, enabling better discovery of transferable temporal evolution mechanisms in dynamic systems.

Abstract: Temporal graph learning is pivotal for deciphering dynamic systems, where the core challenge lies in explicitly modeling the underlying evolving patterns that govern network transformation. However, prevailing methods are predominantly task-centric and rely on restrictive assumptions – such as short-term dependency modeling, static neighborhood semantics, and retrospective time usage. These constraints hinder the discovery of transferable temporal evolution mechanisms. To address this, we propose the Temporal Graph Pattern Machine (TGPM), a foundation framework that shifts the focus toward directly learning generalized evolving patterns. TGPM conceptualizes each interaction as an interaction patch synthesized via temporally-biased random walks, thereby capturing multi-scale structural semantics and long-range dependencies that extend beyond immediate neighborhoods. These patches are processed by a Transformer-based backbone designed to capture global temporal regularities while adapting to context-specific interaction dynamics. To further empower the model, we introduce a suite of self-supervised pre-training tasks – specifically masked token modeling and next-time prediction – to explicitly encode the fundamental laws of network evolution. Extensive experiments show that TGPM consistently achieves state-of-the-art performance in both transductive and inductive link prediction, demonstrating exceptional cross-domain transferability.

Method: LOFT operates in low-dimensional feature subspaces using principal projections that maximize information of remaining data while diminishing forgetting data. It optimizes a small projection matrix plugged into the pretrained model, requiring only one-shot feature fetching instead of repetitive raw data access.

Result: Extensive experiments show LOFT achieves significantly lower computational overhead and superior unlearning performance across diverse models, datasets, tasks, and applications.

Conclusion: LOFT provides an efficient, privacy-preserving solution for machine unlearning by operating in low-dimensional feature subspaces, addressing key limitations of existing methods.

Abstract: Machine Unlearning (MU) aims at removing the influence of specific data from a pretrained model while preserving performance on the remaining data. In this work, a novel perspective for MU is presented upon low-dimensional feature subspaces, which gives rise to the potentials of separating the remaining and forgetting data herein. This separability motivates our LOFT, a method that proceeds unlearning in a LOw-dimensional FeaTure subspace from the pretrained model skithrough principal projections, which are optimized to maximally capture the information of the remaining data and meanwhile diminish that of the forgetting data. In training, LOFT simply optimizes a small-size projection matrix flexibly plugged into the pretrained model, and only requires one-shot feature fetching from the pretrained backbone instead of repetitively accessing the raw data. Hence, LOFT mitigates two critical issues in mainstream MU methods, i.e., the privacy leakage risk from massive data reload and the inefficiency of updates to the entire pretrained model. Extensive experiments validate the significantly lower computational overhead and superior unlearning performance of LOFT across diverse models, datasets, tasks, and applications. Code is anonymously available at https://anonymous.4open.science/r/4352/.

Method: Models molecules as composite exponential-family distributions and defines generative flows along exponential geodesics under the Fisher-Rao metric. Introduces Evolving Exponential Geodesic Flow (EvoEGF-Mol) with dynamically concentrating distributions instead of static Dirac targets, using progressive-parameter-refinement architecture for stable training.

Result: Achieves reference-level PoseBusters passing rate (93.4%) on CrossDock, demonstrating remarkable geometric precision and interaction fidelity. Outperforms baselines on real-world MolGenBench tasks by recovering bioactive scaffolds and generating candidates meeting established MedChem filters.

Conclusion: The information-geometric approach with evolving exponential geodesics provides a principled framework for SBDD that addresses manifold mismatches and enables stable training with high-quality molecular generation.

Abstract: Structure-Based Drug Design (SBDD) aims to discover bioactive ligands. Conventional approaches construct probability paths separately in Euclidean and probabilistic spaces for continuous atomic coordinates and discrete chemical categories, leading to a mismatch with the underlying statistical manifolds. We address this issue from an information-geometric perspective by modeling molecules as composite exponential-family distributions and defining generative flows along exponential geodesics under the Fisher-Rao metric. To avoid the instantaneous trajectory collapse induced by geodesics directly targeting Dirac distributions, we propose Evolving Exponential Geodesic Flow for SBDD (EvoEGF-Mol), which replaces static Dirac targets with dynamically concentrating distributions, ensuring stable training via a progressive-parameter-refinement architecture. Our model approaches a reference-level PoseBusters passing rate (93.4%) on CrossDock, demonstrating remarkable geometric precision and interaction fidelity, while outperforming baselines on real-world MolGenBench tasks by recovering bioactive scaffolds and generating candidates that meet established MedChem filters.

Method: Two-stage training approach: first unrewarded exploration phase where LLMs organize task-relevant knowledge without reward constraints, followed by reward-based learning phase. Extensive experiments across multiple model families and diverse task domains.

Result: LLMs show modest performance improvements during unrewarded exploration, with further enhancement when rewards are introduced. Models trained with this two-stage approach achieve higher competence than those trained with purely reward-based reinforcement learning.

Conclusion: Latent learning dynamics exist in LLMs, suggesting that unrewarded exploration can improve learning efficiency and performance, offering insights beyond traditional reward-centric approaches.

Abstract: Latent learning, classically theorized by Tolman, shows that biological agents (e.g., rats) can acquire internal representations of their environment without rewards, enabling rapid adaptation once rewards are introduced. In contrast, from a cognitive science perspective, reward learning remains overly dependent on external feedback, limiting flexibility and generalization. Although recent advances in the reasoning capabilities of large language models (LLMs), such as OpenAI-o1 and DeepSeek-R1, mark a significant breakthrough, these models still rely primarily on reward-centric reinforcement learning paradigms. Whether and how the well-established phenomenon of latent learning in psychology can inform or emerge within LLMs’ training remains largely unexplored. In this work, we present novel findings from our experiments that LLMs also exhibit the latent learning dynamics. During an initial phase of unrewarded exploration, LLMs display modest performance improvements, as this phase allows LLMs to organize task-relevant knowledge without being constrained by reward-driven biases, and performance is further enhanced once rewards are introduced. LLMs post-trained under this two-stage exploration regime ultimately achieve higher competence than those post-trained with reward-based reinforcement learning throughout. Beyond these empirical observations, we also provide theoretical analyses for our experiments explaining why unrewarded exploration yields performance gains, offering a mechanistic account of these dynamics. Specifically, we conducted extensive experiments across multiple model families and diverse task domains to establish the existence of the latent learning dynamics in LLMs.

Method: Proposes a teacher-student framework: 1) Train single-task teacher models using distributed RL, 2) Continually distill these teachers into a central generalist model using policy distillation, 3) Employ mixture-of-experts architecture and replay-based approach to enhance plasticity and stability.

Result: Extensive experiments on Meta-World benchmark show the framework recovers over 85% of teacher performance while constraining task-wise forgetting to within 10%.

Conclusion: Decoupling continual RL into single-task RL training and multi-task policy distillation enables efficient continual learning with minimal forgetting, leveraging the strengths of both approaches.

Abstract: Continual Reinforcement Learning (CRL) aims to develop lifelong learning agents to continuously acquire knowledge across diverse tasks while mitigating catastrophic forgetting. This requires efficiently managing the stability-plasticity dilemma and leveraging prior experience to rapidly generalize to novel tasks. While various enhancement strategies for both aspects have been proposed, achieving scalable performance by directly applying RL to sequential task streams remains challenging. In this paper, we propose a novel teacher-student framework that decouples CRL into two independent processes: training single-task teacher models through distributed RL and continually distilling them into a central generalist model. This design is motivated by the observation that RL excels at solving single tasks, while policy distillation – a relatively stable supervised learning process – is well aligned with large foundation models and multi-task learning. Moreover, a mixture-of-experts (MoE) architecture and a replay-based approach are employed to enhance the plasticity and stability of the continual policy distillation process. Extensive experiments on the Meta-World benchmark demonstrate that our framework enables efficient continual RL, recovering over 85% of teacher performance while constraining task-wise forgetting to within 10%.

Method: TTCS initializes two policies from the same pretrained model: a question synthesizer and reasoning solver. They co-evolve through iterative optimization where the synthesizer generates progressively challenging question variants tailored to the solver’s current capability, while the solver updates using self-consistency rewards from multiple responses on both original and synthetic questions.

Result: Experiments show TTCS consistently strengthens reasoning ability on challenging mathematical benchmarks and transfers to general-domain tasks across different LLM backbones.

Conclusion: TTCS demonstrates a scalable path toward dynamically constructing test-time curricula for self-evolving LLMs, addressing limitations of existing test-time training methods.

Abstract: Test-Time Training offers a promising way to improve the reasoning ability of large language models (LLMs) by adapting the model using only the test questions. However, existing methods struggle with difficult reasoning problems for two reasons: raw test questions are often too difficult to yield high-quality pseudo-labels, and the limited size of test sets makes continuous online updates prone to instability. To address these limitations, we propose TTCS, a co-evolving test-time training framework. Specifically, TTCS initializes two policies from the same pretrained model: a question synthesizer and a reasoning solver. These policies evolve through iterative optimization: the synthesizer generates progressively challenging question variants conditioned on the test questions, creating a structured curriculum tailored to the solver’s current capability, while the solver updates itself using self-consistency rewards computed from multiple sampled responses on both original test and synthetic questions. Crucially, the solver’s feedback guides the synthesizer to generate questions aligned with the model’s current capability, and the generated question variants in turn stabilize the solver’s test-time training. Experiments show that TTCS consistently strengthens the reasoning ability on challenging mathematical benchmarks and transfers to general-domain tasks across different LLM backbones, highlighting a scalable path towards dynamically constructing test-time curricula for self-evolving. Our code and implementation details are available at https://github.com/XMUDeepLIT/TTCS.

Method: TA-GRPO generates semantically equivalent transformed variants of each question via paraphrasing, variable renaming, and format changes, then computes advantages by pooling rewards across the entire group of variants.

Result: Experiments show consistent Pass@k improvements: gains up to 9.84 points on competition math (AMC12, AIME24) and 5.05 points on out-of-distribution scientific reasoning (GPQA-Diamond).

Conclusion: TA-GRPO reduces zero-gradient probability and improves generalization via reduced train-test distribution shift by training on diverse phrasings that promote multiple solution strategies.

Abstract: Large language models trained via next-token prediction are fundamentally pattern-matchers: sensitive to superficial phrasing variations even when the underlying problem is identical. Group Relative Policy Optimization (GRPO) was designed to improve reasoning, but in fact it worsens this situation through two failure modes: diversity collapse, where training amplifies a single solution strategy while ignoring alternatives of gradient signal, and gradient diminishing, where a large portion of questions yield zero gradients because all rollouts receive identical rewards. We propose TA-GRPO (Transform-Augmented GRPO), which generates semantically equivalent transformed variants of each question (via paraphrasing, variable renaming, and format changes) and computes advantages by pooling rewards across the entire group. This pooled computation ensures mixed rewards even when the original question is too easy or too hard, while training on diverse phrasings promotes multiple solution strategies. We provide theoretical justification showing that TA-GRPO reduces zero-gradient probability and improves generalization via reduced train-test distribution shift. Experiments on mathematical reasoning benchmarks show consistent Pass@k improvements, with gains up to 9.84 points on competition math (AMC12, AIME24) and 5.05 points on out-of-distribution scientific reasoning (GPQA-Diamond).

Method: 1) Categorize LoRA variants along four axes: rank, optimization dynamics, initialization, and integration with Mixture-of-Experts; 2) Develop unified theoretical framework for low-rank update dynamics; 3) Create LoRAFactory modular codebase with unified interface; 4) Conduct large-scale evaluation across natural language generation, understanding, and image classification tasks.

Result: Key findings: LoRA and variants show pronounced sensitivity to learning rate choices compared to other hyperparameters; with proper hyperparameter configurations, LoRA consistently matches or surpasses most variants’ performance.

Conclusion: This unified study provides systematic understanding of LoRA variants, revealing that many proposed improvements may not be necessary when proper hyperparameter tuning is applied, and offers tools for standardized evaluation.

Abstract: Low-Rank Adaptation (LoRA) is a fundamental parameter-efficient fine-tuning method that balances efficiency and performance in large-scale neural networks. However, the proliferation of LoRA variants has led to fragmentation in methodology, theory, code, and evaluation. To this end, this work presents the first unified study of LoRA variants, offering a systematic taxonomy, unified theoretical review, structured codebase, and standardized empirical assessment. First, we categorize LoRA variants along four principal axes: rank, optimization dynamics, initialization, and integration with Mixture-of-Experts. Then, we review their relationships and evolution within a common theoretical framework focused on low-rank update dynamics. Further, we introduce LoRAFactory, a modular codebase that implements variants through a unified interface, supporting plug-and-play experimentation and fine-grained analysis. Last, using this codebase, we conduct a large-scale evaluation across natural language generation, natural language understanding, and image classification tasks, systematically exploring key hyperparameters. Our results uncover several findings, notably: LoRA and its variants exhibit pronounced sensitivity to the choices of learning rate compared to other hyperparameters; moreover, with proper hyperparameter configurations, LoRA consistently matches or surpasses the performance of most of its variants.

Method: STARS monitors latent dynamics by detecting L2 distance spikes in hidden states to identify Cognitive Pivots (reasoning transitions). It uses geometric trajectory analysis to diagnose transition structure and injects state-aware language cues to steer the model in real-time.

Result: Experiments across diverse benchmarks show STARS efficiently reduces redundant loops while improving accuracy through adaptive correction of erroneous reasoning trajectories.

Conclusion: STARS provides a robust, unsupervised mechanism to optimize reasoning processes in Large Reasoning Models without requiring additional fine-tuning.

Abstract: While Large Reasoning Models (LRMs) have achieved remarkable performance by scaling test-time compute, they frequently suffer from Cognitive Inertia, a failure pattern manifesting as either overthinking (inertia of motion) or reasoning rigidity (inertia of direction). Existing detection methods, typically relying on superficial textual heuristics like self-correction tokens, often fail to capture the model’s unvoiced internal conflicts. To address this, we propose STARS (Spike-Triggered Adaptive Reasoning Steering), a training-free framework designed to rectify cognitive inertia by monitoring latent dynamics. STARS identifies Cognitive Pivots-critical moments of reasoning transition-by detecting distinct L2 distance spikes in the hidden states. Upon detection, the framework employs geometric trajectory analysis to diagnose the structural nature of the transition and injects state-aware language cues to steer the model in real-time. Our experiments across diverse benchmarks confirm that STARS efficiently curtails redundant loops while improving accuracy through the adaptive correction of erroneous trajectories. STARS offers a robust, unsupervised mechanism to optimize the reasoning process of LRMs without requiring additional fine-tuning.

Method: Proposes Elastic Spectral State Space Models (ES-SSM) with Hankel spectral filtering over SSM, coupled with lightweight input-adaptive gates trained under randomized spectral budgets. Uses shared masked normalization over ordered spectral channels to concentrate predictive capability in low-index components.

Result: Single ES-SSM model trained once can be truncated to provide competitive performance compared with Transformer and SSM baselines at similar parameter scales. Shows smooth budget-performance curves across various runtime budgets and truncation levels.

Conclusion: ES-SSM offers efficient single-model training with flexible runtime adaptation to different computational constraints, demonstrating strong performance across text, logic, retrieval, vision, and audio tasks.

Abstract: Foundation models are typically trained at a fixed computational capacity, while real-world applications require deployment across platforms with different resource constraints. Current approaches usually rely on training families of model variants or model distillation, which requires additional training and supports only a pre-selected set of sizes rather than fine-grained adaptation at runtime. In this paper, we propose Elastic Spectral State Space Models (ES-SSM), which require only one-time training at full capacity, but can be directly truncated into arbitrary scales for budgeted, runtime inference without retraining. Our ES-SSM builds on Hankel spectral filtering over a state space model (SSM), coupled with a lightweight input-adaptive gate trained under randomized spectral budgets. Using a shared masked normalization rule over the ordered spectral channels, we encourage predictive capability to concentrate in low-index components, while higher-index components act primarily as refinement. We test our algorithm across long-sequence benchmarks spanning text, logic, retrieval, vision, and audio. We demonstrate that a single ES-SSM model trained once can be truncated to provide competitive performance compared with modern Transformer and SSM baselines at similar parameter scales. Furthermore, by testing under various runtime budgets, we observe smooth and stable budget-performance curves over a wide range of truncation levels.

Method: Introduces boundary enrichment metric B to measure chunk start concentration on high-surprisal positions, then proposes Sombrero with confidence-alignment boundary loss and confidence-weighted smoothing at input level to steer boundaries toward predictive difficulty.

Result: On 1B scale across UTF-8 corpora (English/German text, code, math), Sombrero improves accuracy-efficiency trade-off and yields boundaries that consistently align compute with hard-to-predict positions.

Conclusion: The proposed boundary enrichment metric and Sombrero method enable better steering of computational resources in hierarchical sequence models by aligning boundaries with positions of high predictive difficulty.

Abstract: Hierarchical sequence models replace fixed tokenization with learned segmentations that compress long byte sequences for efficient autoregressive modeling. While recent end-to-end methods can learn meaningful boundaries from the language-modeling objective alone, it remains difficult to quantitatively assess and systematically steer where compute is spent. We introduce a router-agnostic metric of boundary quality, boundary enrichment B, which measures how strongly chunk starts concentrate on positions with high next-byte surprisal. Guided by this metric, we propose Sombrero, which steers boundary placement toward predictive difficulty via a confidence-alignment boundary loss and stabilizes boundary learning by applying confidence-weighted smoothing at the input level rather than on realized chunks. On 1B scale, across UTF-8 corpora covering English and German text as well as code and mathematical content, Sombrero improves the accuracy-efficiency trade-off and yields boundaries that more consistently align compute with hard-to-predict positions.

Method: Gradual Fine-Tuning (GFT) defines a temperature-controlled sequence of intermediate objectives that smoothly interpolate between pretrained and target drifts for stochastic flows. It approaches the true target as temperature approaches zero, enabling use of suitable couplings (e.g., optimal transport) while preserving correctness.

Result: GFT improves convergence stability and shortens probability paths, resulting in faster inference while maintaining generation quality comparable to standard fine-tuning. Theoretical convergence results are proven for both marginal and conditional GFT objectives.

Conclusion: GFT provides a theoretically grounded and practically effective alternative for scalable adaptation of flow matching models under distribution shift, positioning it as a valuable framework for fine-tuning flow-based generative models.

Abstract: Fine-tuning flow matching models is a central challenge in settings with limited data, evolving distributions, or strict efficiency demands, where unconstrained fine-tuning can erode the accuracy and efficiency gains learned during pretraining. Prior work has produced theoretical guarantees and empirical advances for reward-based fine-tuning formulations, but these methods often impose restrictions on permissible drift structure or training techniques. In this work, we propose Gradual Fine-Tuning (GFT), a principled framework for fine-tuning flow-based generative models when samples from the target distribution are available. For stochastic flows, GFT defines a temperature-controlled sequence of intermediate objectives that smoothly interpolate between the pretrained and target drifts, approaching the true target as the temperature approaches zero. We prove convergence results for both marginal and conditional GFT objectives, enabling the use of suitable (e.g., optimal transport) couplings during GFT while preserving correctness. Empirically, GFT improves convergence stability and shortens probability paths, resulting in faster inference, while maintaining generation quality comparable to standard fine-tuning. Our results position GFT as a theoretically grounded and practically effective alternative for scalable adaptation of flow matching models under distribution shift.

Method: Introduces an information-theoretic framework defining “action sufficiency” - a condition on goal representations necessary for optimal action selection. Shows that standard log-loss training of low-level policies naturally induces action-sufficient representations. Proves value sufficiency does not imply action sufficiency.

Result: Empirical verification shows action sufficiency is more strongly associated with control success than value sufficiency in discrete environments. Actor-derived representations consistently outperform representations learned via value estimation on popular benchmarks.

Conclusion: Action-sufficient goal representations are crucial for hierarchical offline RL, and actor-derived representations provide better performance than value-based approaches for long-horizon goal-conditioned tasks.

Abstract: Hierarchical policies in offline goal-conditioned reinforcement learning (GCRL) addresses long-horizon tasks by decomposing control into high-level subgoal planning and low-level action execution. A critical design choice in such architectures is the goal representation-the compressed encoding of goals that serves as the interface between these levels. Existing approaches commonly derive goal representations while learning value functions, implicitly assuming that preserving information sufficient for value estimation is adequate for optimal control. We show that this assumption can fail, even when the value estimation is exact, as such representations may collapse goal states that need to be differentiated for action learning. To address this, we introduce an information-theoretic framework that defines action sufficiency, a condition on goal representations necessary for optimal action selection. We prove that value sufficiency does not imply action sufficiency and empirically verify that the latter is more strongly associated with control success in a discrete environment. We further demonstrate that standard log-loss training of low-level policies naturally induces action-sufficient representations. Our experimental results a popular benchmark demonstrate that our actor-derived representations consistently outperform representations learned via value estimation.

Method: MoVE introduces a global bank of learnable value embeddings shared across all attention layers. For each sequence step, a differentiable soft gating mechanism dynamically mixes retrieved concepts from this bank into the standard value projection, allowing parametric memory to scale independently of network depth by increasing embedding slots.

Result: MoVE consistently outperforms standard and layer-wise memory baselines in both text and image generation tasks, enabling “memory-dense” models that achieve lower perplexity and higher fidelity than dense counterparts at comparable compute budgets.

Conclusion: MoVE successfully decouples memory from compute in autoregressive modeling, establishing a new axis for scaling capacity that enables more efficient model architectures across different modalities.

Abstract: Autoregressive sequence modeling stands as the cornerstone of modern Generative AI, powering results across diverse modalities ranging from text generation to image generation. However, a fundamental limitation of this paradigm is the rigid structural coupling of model capacity to computational cost: expanding a model’s parametric memory – its repository of factual knowledge or visual patterns – traditionally requires deepening or widening the network, which incurs a proportional rise in active FLOPs. In this work, we introduce $\textbf{MoVE (Mixture of Value Embeddings)}$, a mechanism that breaks this coupling and establishes a new axis for scaling capacity. MoVE decouples memory from compute by introducing a global bank of learnable value embeddings shared across all attention layers. For every step in the sequence, the model employs a differentiable soft gating mechanism to dynamically mix retrieved concepts from this bank into the standard value projection. This architecture allows parametric memory to be scaled independently of network depth by simply increasing the number of embedding slots. We validate MoVE through strictly controlled experiments on two representative applications of autoregressive modeling: Text Generation and Image Generation. In both domains, MoVE yields consistent performance improvements over standard and layer-wise memory baselines, enabling the construction of “memory-dense” models that achieve lower perplexity and higher fidelity than their dense counterparts at comparable compute budgets.

Method: Dual Replay with Experience Enhancement (DREE) framework to improve learning efficiency and mitigate catastrophic forgetting under continual drift

Result: DREE effectively learns new tasks, preserves prior knowledge, improves generalization to unseen tasks, and can be applied to diverse existing neural solvers

Conclusion: DREE addresses the practical challenge of continual task drift in neural VRP solvers with limited training resources per task

Abstract: Existing neural solvers for vehicle routing problems (VRPs) are typically trained either in a one-off manner on a fixed set of pre-defined tasks or in a lifelong manner on several tasks arriving sequentially, assuming sufficient training on each task. Both settings overlook a common real-world property: problem patterns may drift continually over time, yielding massive tasks sequentially arising while offering only limited training resources per task. In this paper, we study a novel lifelong learning paradigm for neural VRP solvers under continually drifting tasks over learning time steps, where sufficient training for any given task at any time is not available. We propose Dual Replay with Experience Enhancement (DREE), a general framework to improve learning efficiency and mitigate catastrophic forgetting under such drift. Extensive experiments show that, under such continual drift, DREE effectively learns new tasks, preserves prior knowledge, improves generalization to unseen tasks, and can be applied to diverse existing neural solvers.

Method: Train transformers on synthetic arithmetic tasks, conduct extensive ablations and fine-grained diagnostic metrics to analyze learning patterns and skill acquisition order.

Result: Transformers don’t learn skills sequentially like humans; they acquire skills in reverse order or parallel, causing “shattered compositionality” that leads to mixing errors under distribution shifts.

Conclusion: There’s a fundamental mismatch between transformer learning behavior and desired skill compositions, with implications for reasoning reliability, OOD robustness, and alignment that isn’t solved by scaling or reasoning techniques.

Abstract: Large language models (LLMs) often exhibit unexpected errors or unintended behavior, even at scale. While recent work reveals the discrepancy between LLMs and humans in skill compositions, the learning dynamics of skill compositions and the underlying cause of non-human behavior remain elusive. In this study, we investigate the mechanism of learning dynamics by training transformers on synthetic arithmetic tasks. Through extensive ablations and fine-grained diagnostic metrics, we discover that transformers do not reliably build skill compositions according to human-like sequential rules. Instead, they often acquire skills in reverse order or in parallel, which leads to unexpected mixing errors especially under distribution shifts–a phenomenon we refer to as shattered compositionality. To explain these behaviors, we provide evidence that correlational matching to the training data, rather than causal or procedural composition, shapes learning dynamics. We further show that shattered compositionality persists in modern LLMs and is not mitigated by pure model scaling or scratchpad-based reasoning. Our results reveal a fundamental mismatch between a model’s learning behavior and desired skill compositions, with implications for reasoning reliability, out-of-distribution robustness, and alignment.

Method: The authors leverage recent results on Transformer continuity to prove theoretical limitations of perplexity. They show that if a compact decoder-only Transformer predicts any sequence accurately and confidently, there must exist another sequence with very low perplexity but incorrect predictions. They also analytically study iso-perplexity plots to examine perplexity’s selection properties.

Result: The paper proves that perplexity can be an unsuitable metric for model selection because: 1) Models that generalize well must have sequences with low perplexity but incorrect predictions, and 2) Perplexity doesn’t always select more accurate models - increased confidence must be accompanied by proportional accuracy gains for selection.

Conclusion: Perplexity has fundamental limitations as a model selection metric due to mathematical properties of Transformer models. The analysis provides rigorous justification for empirical observations about perplexity’s shortcomings and suggests caution in using perplexity for model evaluation and selection.

Abstract: Perplexity – a function measuring a model’s overall level of “surprise” when encountering a particular output – has gained significant traction in recent years, both as a loss function and as a simple-to-compute metric of model quality. Prior studies have pointed out several limitations of perplexity, often from an empirical manner. Here we leverage recent results on Transformer continuity to show in a rigorous manner how perplexity may be an unsuitable metric for model selection. Specifically, we prove that, if there is any sequence that a compact decoder-only Transformer model predicts accurately and confidently – a necessary pre-requisite for strong generalisation – it must imply existence of another sequence with very low perplexity, but not predicted correctly by that same model. Further, by analytically studying iso-perplexity plots, we find that perplexity will not always select for the more accurate model – rather, any increase in model confidence must be accompanied by a commensurate rise in accuracy for the new model to be selected.

Method: The study proposes SCOPE-PD framework that collects multimodal clinical assessment data from PPMI study, applies multiple ML techniques, selects the best model, and uses SHAP-based analysis for interpretability of feature contributions.

Result: Random Forest achieved 98.66% accuracy using combined subjective and objective features. Top contributing features identified were tremor, bradykinesia, and facial expression from MDS-UPDRS test.

Conclusion: SCOPE-PD demonstrates that integrating multimodal clinical data with explainable AI can provide accurate and interpretable Parkinson’s disease predictions, enabling personalized health decisions.

Abstract: Parkinson’s disease (PD) is a chronic and complex neurodegenerative disorder influenced by genetic, clinical, and lifestyle factors. Predicting this disease early is challenging because it depends on traditional diagnostic methods that face issues of subjectivity, which commonly delay diagnosis. Several objective analyses are currently in practice to help overcome the challenges of subjectivity; however, a proper explanation of these analyses is still lacking. While machine learning (ML) has demonstrated potential in supporting PD diagnosis, existing approaches often rely on subjective reports only and lack interpretability for individualized risk estimation. This study proposes SCOPE-PD, an explainable AI-based prediction framework, by integrating subjective and objective assessments to provide personalized health decisions. Subjective and objective clinical assessment data are collected from the Parkinson’s Progression Markers Initiative (PPMI) study to construct a multimodal prediction framework. Several ML techniques are applied to these data, and the best ML model is selected to interpret the results. Model interpretability is examined using SHAP-based analysis. The Random Forest algorithm achieves the highest accuracy of 98.66 percent using combined features from both subjective and objective test data. Tremor, bradykinesia, and facial expression are identified as the top three contributing features from the MDS-UPDRS test in the prediction of PD.

Method: Formulated as mixed-integer non-convex optimization problem. Derived closed-form optimal flight altitude under VLC channel gain threshold, then optimized horizontal trajectory using twin delayed deep deterministic policy gradient algorithm with novel pheromone-driven reward mechanism.

Result: Optimal altitude reduces flight distance by up to 35% compared to baselines. Pheromone-driven reward mechanism shortens convergence steps by approximately 50%, showing significant efficiency gains.

Conclusion: Proposed framework effectively optimizes UAV trajectory for VLC data collection, achieving substantial improvements in flight efficiency and convergence speed through combined analytical and reinforcement learning approaches.

Abstract: Recently, the integration of unmanned aerial vehicle (UAV) and visible light communication (VLC) technologies has emerged as a promising solution to offer flexible communication and efficient lighting. This letter investigates the three-dimensional trajectory planning in a UAV-assisted VLC system, where a UAV is dispatched to collect data from ground users (GUs). The core objective is to develop a trajectory planning framework that minimizes UAV flight distance, which is equivalent to maximizing the data collection efficiency. This issue is formulated as a challenging mixed-integer non-convex optimization problem. To tackle it, we first derive a closed-form optimal flight altitude under specific VLC channel gain threshold. Subsequently, we optimize the UAV horizontal trajectory by integrating a novel pheromone-driven reward mechanism with the twin delayed deep deterministic policy gradient algorithm, which enables adaptive UAV motion strategy in complex environments. Simulation results validate that the derived optimal altitude effectively reduces the flight distance by up to 35% compared to baseline methods. Additionally, the proposed reward mechanism significantly shortens the convergence steps by approximately 50%, demonstrating notable efficiency gains in the context of UAV-assisted VLC data collection.

Method: REKD uses knowledge distillation where student RE models learn from both teacher rationales and predictions, in addition to their own RE optimization. Method is neural-model agnostic and works with any backbone model.

Result: Experiments with BERT and ViT variants across IMDB, CIFAR-10, and CIFAR-100 show REKD significantly improves predictive performance of student RE models.

Conclusion: REKD effectively enhances rationale extraction models by leveraging teacher rationales through knowledge distillation, improving performance while maintaining interpretability.

Abstract: Because of the pervasive use of deep neural networks (DNNs), especially in high-stakes domains, the interpretability of DNNs has received increased attention. The general idea of rationale extraction (RE) is to provide an interpretable-by-design framework for DNNs via a select-predict architecture where two neural networks learn jointly to perform feature selection and prediction, respectively. Given only the remote supervision from the final task prediction, the process of learning to select subsets of features (or \emph{rationales}) requires searching in the space of all possible feature combinations, which is computationally challenging and even harder when the base neural networks are not sufficiently capable. To improve the predictive performance of RE models that are based on less capable or smaller neural networks (i.e., the students), we propose \textbf{REKD} (\textbf{R}ationale \textbf{E}xtraction with \textbf{K}nowledge \textbf{D}istillation) where a student RE model learns from the rationales and predictions of a teacher (i.e., a \emph{rationalist}) in addition to the student’s own RE optimization. This structural adjustment to RE aligns well with how humans could learn effectively from interpretable and verifiable knowledge. Because of the neural-model agnostic nature of the method, any black-box neural network could be integrated as a backbone model. To demonstrate the viability of REKD, we conduct experiments with multiple variants of BERT and vision transformer (ViT) models. Our experiments across language and vision classification datasets (i.e., IMDB movie reviews, CIFAR 10 and CIFAR 100) show that REKD significantly improves the predictive performance of the student RE models.

Method: Proposes Variational Bayesian Flow Network (VBFN) with variational lifting to a tractable joint Gaussian variational belief family governed by structured precisions. Each Bayesian update solves a symmetric positive definite linear system, enabling coupled node and edge updates in a single fusion step. Uses sample-agnostic sparse precisions from representation-induced dependency graphs to avoid label leakage while enforcing node-edge consistency.

Result: On synthetic and molecular graph datasets, VBFN improves fidelity and diversity and surpasses baseline methods.

Conclusion: VBFN provides a principled approach for graph generation that naturally handles node-edge coupling through structured variational beliefs, outperforming factorized methods.

Abstract: Graph generation aims to sample discrete node and edge attributes while satisfying coupled structural constraints. Diffusion models for graphs often adopt largely factorized forward-noising, and many flow-matching methods start from factorized reference noise and coordinate-wise interpolation, so node-edge coupling is not encoded by the generative geometry and must be recovered implicitly by the core network, which can be brittle after discrete decoding. Bayesian Flow Networks (BFNs) evolve distribution parameters and naturally support discrete generation. But classical BFNs typically rely on factorized beliefs and independent channels, which limit geometric evidence fusion. We propose Variational Bayesian Flow Network (VBFN), which performs a variational lifting to a tractable joint Gaussian variational belief family governed by structured precisions. Each Bayesian update reduces to solving a symmetric positive definite linear system, enabling coupled node and edge updates within a single fusion step. We construct sample-agnostic sparse precisions from a representation-induced dependency graph, thereby avoiding label leakage while enforcing node-edge consistency. On synthetic and molecular graph datasets, VBFN improves fidelity and diversity, and surpasses baseline methods.

Method: Introduces learnable permutation cost matrix to quantify swapping costs, differentiable bipartite matching solver to obtain optimal binary permutation matrix, and sparsity optimization loss to directly optimize permutation operator end-to-end.

Result: Extensively validated on vision and language Transformers, achieving state-of-the-art permutation results for structured sparsity.

Conclusion: The proposed end-to-end learnable permutation framework effectively addresses the limitations of heuristic approaches and improves structured sparsity pruning performance.

Abstract: Structured sparsity has emerged as a popular model pruning technique, widely adopted in various architectures, including CNNs, Transformer models, and especially large language models (LLMs) in recent years. A promising direction to further improve post-pruning performance is weight permutation, which reorders model weights into patterns more amenable to pruning. However, the exponential growth of the permutation search space with the scale of Transformer architectures forces most methods to rely on greedy or heuristic algorithms, limiting the effectiveness of reordering. In this work, we propose a novel end-to-end learnable permutation framework. Our method introduces a learnable permutation cost matrix to quantify the cost of swapping any two input channels of a given weight matrix, a differentiable bipartite matching solver to obtain the optimal binary permutation matrix given a cost matrix, and a sparsity optimization loss function to directly optimize the permutation operator. We extensively validate our approach on vision and language Transformers, demonstrating that our method achieves state-of-the-art permutation results for structured sparsity.

Method: Constructed a minimalist baseline for disentangling factors: one rollout per query per round, outcome reward as training signal without advantage trick, and batch size of 32. This connects to batched contextual bandit learning. Designed an experiment pipeline to examine marginal gains of factors like advantage and number of rollouts across three base models and two datasets.

Result: The systematic analysis revealed new understanding of how various design choices affect learning and generalization dynamics, and identified which ones are critical and deserve more research effort.

Conclusion: The paper provides principled answers to fundamental questions about reinforcement fine-tuning design choices, moving beyond anecdotal evidence to systematic understanding of what matters most for learning and generalization.

Abstract: The reinforcement fine-tuning area is undergoing an explosion papers largely on optimizing design choices. Though performance gains are often claimed, inconsistent conclusions also arise from time to time, making the progress illusive. Reflecting on this illusion, we still lack principled answers to two fundamental questions: 1) what is the role of each design choice? 2) which ones are critical? This paper aims to shed light on them. The underlying challenge is that design choices are entangled together, making their contribution to learning and generalization difficult to attribute. To address this challenge, we first construct a minimalist baseline for disentangling factors: one rollout per query in each round, the outcome reward serving as the training signal without any advantage trick, and a batch size of thirty-two. This baseline connects to batched contextual bandit learning, which facilitates experimental analysis. Centering around this baseline, we design an experiment pipeline, examining the marginal gains of factors like advantage, number of rollouts, etc. Experiments on three base models and two datasets, not only reveal new understanding on the role of various design choices on learning and generalization dynamics, but also identify critical ones that deserve more effort.

Method: Proposes L2D-SLDS, a factorized switching linear-Gaussian state-space model with context-dependent regime transitions, a shared global factor for cross-expert information transfer, and per-expert idiosyncratic states. Uses one-step-ahead predictive beliefs with an IDS-inspired routing rule that balances predicted cost against information gained about latent regime and shared factor.

Result: Experiments show improvements over contextual-bandit baselines and a no-shared-factor ablation, demonstrating the effectiveness of the proposed approach for expert routing in non-stationary environments.

Conclusion: The proposed L2D-SLDS model with shared global factors and information-directed routing provides an effective framework for learning to defer in non-stationary time series with partial feedback and dynamic expert availability.

Abstract: We study Learning to Defer for non-stationary time series with partial feedback and time-varying expert availability. At each time step, the router selects an available expert, observes the target, and sees only the queried expert’s prediction. We model signed expert residuals using L2D-SLDS, a factorized switching linear-Gaussian state-space model with context-dependent regime transitions, a shared global factor enabling cross-expert information transfer, and per-expert idiosyncratic states. The model supports expert entry and pruning via a dynamic registry. Using one-step-ahead predictive beliefs, we propose an IDS-inspired routing rule that trades off predicted cost against information gained about the latent regime and shared factor. Experiments show improvements over contextual-bandit baselines and a no-shared-factor ablation.

Method: Proposes Mem-T agent with hierarchical memory database for dynamic updates and multi-turn retrieval, and MoT-GRPO framework using tree-guided reinforcement learning with memory operation tree backpropagation and hindsight credit assignment.

Result: Mem-T outperforms frameworks like A-Mem and Mem0 by up to 14.92%, operates on favorable accuracy-efficiency Pareto frontier, and reduces inference tokens per query by ~24.45% relative to GAM without performance loss.

Conclusion: The proposed approach enables effective end-to-end optimization of memory management policies through dense step-wise supervision, achieving both high performance and efficiency.

Abstract: Memory agents, which depart from predefined memory-processing pipelines by endogenously managing the processing, storage, and retrieval of memories, have garnered increasing attention for their autonomy and adaptability. However, existing training paradigms remain constrained: agents often traverse long-horizon sequences of memory operations before receiving sparse and delayed rewards, which hinders truly end-to-end optimization of memory management policies. To address this limitation, we introduce Mem-T, an autonomous memory agent that interfaces with a lightweight hierarchical memory database to perform dynamic updates and multi-turn retrieval over streaming inputs. To effectively train long-horizon memory management capabilities, we further propose MoT-GRPO, a tree-guided reinforcement learning framework that transforms sparse terminal feedback into dense, step-wise supervision via memory operation tree backpropagation and hindsight credit assignment, thereby enabling the joint optimization of memory construction and retrieval. Extensive experiments demonstrate that Mem-T is (1) high-performing, surpassing frameworks such as A-Mem and Mem0 by up to $14.92%$, and (2) economical, operating on a favorable accuracy-efficiency Pareto frontier and reducing inference tokens per query by $\sim24.45%$ relative to GAM without sacrificing performance.

Method: The proposed algorithm comprises three components: (1) a model-based module that uses the target distribution for guided but slow sampling, (2) a model-free module that learns patterns from previous samples to enable fast, reflexive sampling without evaluating the expensive target distribution, and (3) an episodic-control module that supports rapid sampling by recalling specific past samples.

Result: The approach advances Bayesian methods and facilitates their application to large-scale statistical machine learning problems, particularly in Bayesian deep learning with proper uncertainty quantification.

Conclusion: The neuroscience-inspired multi-system approach provides an efficient sampling algorithm for Bayesian inference that balances computational cost and flexibility, enabling practical application to large-scale machine learning problems with principled uncertainty quantification.

Abstract: Humans learn efficiently from their environment by engaging multiple interacting neural systems that support distinct yet complementary forms of control, including model-based (goal-directed) planning, model-free (habitual) responding, and episodic memory-based learning. Model-based mechanisms compute prospective action values using an internal model of the environment, supporting flexible but computationally costly planning; model-free mechanisms cache value estimates and build heuristics that enable fast, efficient habitual responding; and memory-based mechanisms allow rapid adaptation from individual experience. In this work, we aim to elucidate the computational principles underlying this biological efficiency and translate them into a sampling algorithm for scalable Bayesian inference through effective exploration of the posterior distribution. More specifically, our proposed algorithm comprises three components: a model-based module that uses the target distribution for guided but computationally slow sampling; a model-free module that uses previous samples to learn patterns in the parameter space, enabling fast, reflexive sampling without directly evaluating the expensive target distribution; and an episodic-control module that supports rapid sampling by recalling specific past events (i.e., samples). We show that this approach advances Bayesian methods and facilitates their application to large-scale statistical machine learning problems. In particular, we apply our proposed framework to Bayesian deep learning, with an emphasis on proper and principled uncertainty quantification.

Method: Proposes new objectives for studying language identification and generation without realizability assumptions. Uses theoretical analysis to characterize these problems in the agnostic setup and derives statistical rates.

Result: Obtains novel characterizations and nearly tight statistical rates for both language identification and generation problems in the agnostic setup.

Conclusion: The work provides a more general theoretical framework for language identification and generation by removing realizability assumptions, offering new insights and tight statistical bounds for these fundamental problems.

Abstract: Recent works on language identification and generation have established tight statistical rates at which these tasks can be achieved. These works typically operate under a strong realizability assumption: that the input data is drawn from an unknown distribution necessarily supported on some language in a given collection. In this work, we relax this assumption of realizability entirely, and impose no restrictions on the distribution of the input data. We propose objectives to study both language identification and generation in this more general “agnostic” setup. Across both problems, we obtain novel interesting characterizations and nearly tight rates.

Method: Proposes EUGens (Efficient, Unified and General dense layers) that use random features to approximate standard FFLs with input norm dependence, unifying existing efficient FFL extensions and enabling linear-time inference.

Result: Integration into Transformers and MLPs yields up to 27% faster inference and 30% memory efficiency improvements across image classification, language model pre-training, and 3D scene reconstruction tasks.

Conclusion: EUGens enable scalable deployment of large-scale neural networks by reducing computational overhead while maintaining expressive power, with potential for real-world applications.

Abstract: Efficient neural networks are essential for scaling machine learning models to real-time applications and resource-constrained environments. Fully-connected feedforward layers (FFLs) introduce computation and parameter count bottlenecks within neural network architectures. To address this challenge, in this work, we propose a new class of dense layers that generalize standard fully-connected feedforward layers, \textbf{E}fficient, \textbf{U}nified and \textbf{Gen}eral dense layers (EUGens). EUGens leverage random features to approximate standard FFLs and go beyond them by incorporating a direct dependence on the input norms in their computations. The proposed layers unify existing efficient FFL extensions and improve efficiency by reducing inference complexity from quadratic to linear time. They also lead to \textbf{the first} unbiased algorithms approximating FFLs with arbitrary polynomial activation functions. Furthermore, EuGens reduce the parameter count and computational overhead while preserving the expressive power and adaptability of FFLs. We also present a layer-wise knowledge transfer technique that bypasses backpropagation, enabling efficient adaptation of EUGens to pre-trained models. Empirically, we observe that integrating EUGens into Transformers and MLPs yields substantial improvements in inference speed (up to \textbf{27}%) and memory efficiency (up to \textbf{30}%) across a range of tasks, including image classification, language model pre-training, and 3D scene reconstruction. Overall, our results highlight the potential of EUGens for the scalable deployment of large-scale neural networks in real-world scenarios.

Method: Proposes conserved quantity correction, a model-agnostic technique that incorporates physical conservation criteria into deep learning models to enforce conservation laws during auto-regressive predictions.

Result: Demonstrates consistent improvement in long-term stability of auto-regressive neural operator models across different architectures, with spectral analysis revealing limitations in handling high-frequency components important for turbulent flows.

Conclusion: Physical conservation constraints significantly improve neural operator stability, but current architectures need better handling of high-frequency components for modeling turbulent flows, suggesting future work on frequency-aware architectures.

Abstract: Deep learning has been proposed as an efficient alternative for the numerical approximation of PDE solutions, offering fast, iterative simulation of PDEs through the approximation of solution operators. However, deep learning solutions have struggle to perform well over long prediction durations due to the accumulation of auto-regressive error, which is compounded by the inability of models to conserve physical quantities. In this work, we present conserved quantity correction, a model-agnostic technique for incorporation physical conservation criteria within deep learning models. Our results demonstrate consistent improvement in the long-term stability of auto-regressive neural operator models, regardless of the model architecture. Furthermore, we analyze the performance of neural operators from the spectral domain, highlighting significant limitations of present architectures. These results highlight the need for future work to consider architectures that place specific emphasis on high frequency components, which are integral to the understanding and modeling of turbulent flows.

Method: FOCUS dynamically focuses computation on decodable tokens and evicts non-decodable ones on-the-fly by leveraging the correlation between attention-derived token importance and token-wise decoding probability.

Result: FOCUS achieves up to 3.52× throughput improvement over LMDeploy while preserving or improving generation quality across multiple benchmarks.

Conclusion: FOCUS enables scalable throughput for Diffusion LLMs by addressing computational inefficiencies in decoding, making DLLMs more practical for deployment.

Abstract: Diffusion Large Language Models (DLLMs) offer a compelling alternative to Auto-Regressive models, but their deployment is constrained by high decoding cost. In this work, we identify a key inefficiency in DLLM decoding: while computation is parallelized over token blocks, only a small subset of tokens is decodable at each diffusion step, causing most compute to be wasted on non-decodable tokens. We further observe a strong correlation between attention-derived token importance and token-wise decoding probability. Based on this insight, we propose FOCUS – an inference system designed for DLLMs. By dynamically focusing computation on decodable tokens and evicting non-decodable ones on-the-fly, FOCUS increases the effective batch size, alleviating compute limitations and enabling scalable throughput. Empirical evaluations demonstrate that FOCUS achieves up to 3.52$\times$ throughput improvement over the production-grade engine LMDeploy, while preserving or improving generation quality across multiple benchmarks. The FOCUS system is publicly available on GitHub: https://github.com/sands-lab/FOCUS.

Method: FedDis uses a dual-branch architecture: a Personalized Bank learns client-specific factors, and a Global Pattern Bank distills common knowledge. A mutual information minimization objective enforces informational orthogonality between branches for effective disentanglement.

Result: Comprehensive experiments on four real-world benchmark datasets show FedDis consistently achieves state-of-the-art performance, promising efficiency, and superior expandability.

Conclusion: FedDis successfully addresses non-IID challenges in federated traffic prediction through causal disentanglement, enabling robust cross-client knowledge transfer while preserving local adaptability.

Abstract: Federated learning offers a promising paradigm for privacy-preserving traffic prediction, yet its performance is often challenged by the non-identically and independently distributed (non-IID) nature of decentralized traffic data. Existing federated methods frequently struggle with this data heterogeneity, typically entangling globally shared patterns with client-specific local dynamics within a single representation. In this work, we postulate that this heterogeneity stems from the entanglement of two distinct generative sources: client-specific localized dynamics and cross-client global spatial-temporal patterns. Motivated by this perspective, we introduce FedDis, a novel framework that, to the best of our knowledge, is the first to leverage causal disentanglement for federated spatial-temporal prediction. Architecturally, FedDis comprises a dual-branch design wherein a Personalized Bank learns to capture client-specific factors, while a Global Pattern Bank distills common knowledge. This separation enables robust cross-client knowledge transfer while preserving high adaptability to unique local environments. Crucially, a mutual information minimization objective is employed to enforce informational orthogonality between the two branches, thereby ensuring effective disentanglement. Comprehensive experiments conducted on four real-world benchmark datasets demonstrate that FedDis consistently achieves state-of-the-art performance, promising efficiency, and superior expandability.

Method: Proposes Median-Centered Group Relative Policy Optimization (MC-GRPO) that replaces mean baseline with median baseline, which is less sensitive to outlier rewards. Generates G+1 rollouts, uses median as baseline, excludes the median pivot rollout from backpropagation to maintain same gradient cost as standard G-rollout training.

Result: MC-GRPO consistently improves stability and final accuracy in low-rollout regime across various GRPO-family methods and model scales, reducing the gap between G=2 and G=8 to within 1%.

Conclusion: Median-centered training effectively addresses advantage sign flip issues in small-rollout scenarios, providing a simple yet effective solution for resource-constrained language model training.

Abstract: Group-relative policy optimization methods train language models by generating multiple rollouts per prompt and normalizing rewards with a shared mean reward baseline. In resource-constrained settings where the rollout budget is small, accuracy often degrades. We find that noise in the shared baseline induces advantage sign flips, where some rollouts receive an incorrect advantage sign, and the update direction is reversed. To address this, we propose Median-Centered Group Relative Policy Optimization (MC-GRPO), a simple and effective solution for small-rollout training. Our main idea is to replace the mean baseline with a median baseline: the median is far less sensitive to outlier rewards than the mean, mitigating the sign flips under small rollout size (G). We generate one additional rollout for median reference (G+1), and compute advantages by using the group median. With an odd-sized group, exactly one completion is the median and receives zero advantage, we exclude this pivot rollout from backpropagation so the number of gradient-contributing samples per prompt remains G, preserving the core update cost of standard G-rollout training. Across various GRPO-family methods and a wide range of models and scales, this median-centered training consistently improves stability and final accuracy in the low-rollout regime, reducing the gap between G=2 and G=8 to within 1%. Code is available at https://github.com/lotusroot-kim/MC-GRPO

Method: Proposes a non-intrusive graph-based bot detection framework that models user session behavior through graph representations and applies inductive graph neural networks for classification, capturing both relational structure and behavioral semantics

Result: Experiments on real-world e-commerce traffic show the inductive graph model outperforms session-level multilayer perceptron baselines in AUC and F1 scores, with additional tests demonstrating robustness under adversarial perturbations and effective generalization to unseen sessions and URLs

Conclusion: The framework is deployment-friendly, integrates with existing systems without client-side instrumentation, supports real-time inference and incremental updates, making it suitable for practical e-commerce security deployments

Abstract: Malicious bots pose a growing threat to e-commerce platforms by scraping data, hoarding inventory, and perpetrating fraud. Traditional bot mitigation techniques, including IP blacklists and CAPTCHA-based challenges, are increasingly ineffective or intrusive, as modern bots leverage proxies, botnets, and AI-assisted evasion strategies. This work proposes a non-intrusive graph-based bot detection framework for e-commerce that models user session behavior through a graph representation and applies an inductive graph neural network for classification. The approach captures both relational structure and behavioral semantics, enabling accurate identification of subtle automated activity that evades feature-based methods. Experiments on real-world e-commerce traffic demonstrate that the proposed inductive graph model outperforms a strong session-level multilayer perceptron baseline in terms of AUC and F1 score. Additional adversarial perturbation and cold-start simulations show that the model remains robust under moderate graph modifications and generalizes effectively to previously unseen sessions and URLs. The proposed framework is deployment-friendly, integrates with existing systems without client-side instrumentation, and supports real-time inference and incremental updates, making it suitable for practical e-commerce security deployments.

Method: FedCARE uses gradient ascent for efficient forgetting when target data are locally available, and employs data-free model inversion to construct class-level proxies of shared knowledge. It integrates a pseudo-sample generator, conflict-aware projected gradient ascent for utility-preserving unlearning, and a recovery strategy that suppresses rollback toward the pre-unlearning model.

Result: Extensive experiments on multiple datasets and model architectures under both IID and non-IID settings show that FedCARE achieves effective forgetting, improved utility retention, and reduced relearning risk compared to state-of-the-art federated unlearning baselines.

Conclusion: FedCARE provides a unified and low-overhead federated unlearning framework that supports client, instance, and class-level unlearning while addressing key challenges of existing methods.

Abstract: Federated learning (FL) enables collaborative model training without centralizing raw data, but privacy regulations such as the right to be forgotten require FL systems to remove the influence of previously used training data upon request. Retraining a federated model from scratch is prohibitively expensive, motivating federated unlearning (FU). However, existing FU methods suffer from high unlearning overhead, utility degradation caused by entangled knowledge, and unintended relearning during post-unlearning recovery. In this paper, we propose FedCARE, a unified and low overhead FU framework that enables conflict-aware unlearning and relearning-resistant recovery. FedCARE leverages gradient ascent for efficient forgetting when target data are locally available and employs data free model inversion to construct class level proxies of shared knowledge. Based on these insights, FedCARE integrates a pseudo-sample generator, conflict-aware projected gradient ascent for utility preserving unlearning, and a recovery strategy that suppresses rollback toward the pre-unlearning model. FedCARE supports client, instance, and class level unlearning with modest overhead. Extensive experiments on multiple datasets and model architectures under both IID and non-IID settings show that FedCARE achieves effective forgetting, improved utility retention, and reduced relearning risk compared to state of the art FU baselines.

Method: MGMT uses Graph Transformer encoders to map each graph into a shared latent space, selects task-relevant supernodes via attention, builds a meta-graph connecting functionally aligned supernodes across graphs, and applies additional Graph Transformer layers for joint reasoning.

Result: MGMT consistently outperforms existing state-of-the-art models in graph-level prediction tasks on both synthetic datasets and real-world neuroscience applications, while providing interpretable representations.

Conclusion: MGMT establishes a unified framework for structured multi-graph learning with built-in interpretability, advancing representation techniques for graph-based data.

Abstract: Multi-graph learning is crucial for extracting meaningful signals from collections of heterogeneous graphs. However, effectively integrating information across graphs with differing topologies, scales, and semantics, often in the absence of shared node identities, remains a significant challenge. We present the Multi-Graph Meta-Transformer (MGMT), a unified, scalable, and interpretable framework for cross-graph learning. MGMT first applies Graph Transformer encoders to each graph, mapping structure and attributes into a shared latent space. It then selects task-relevant supernodes via attention and builds a meta-graph that connects functionally aligned supernodes across graphs using similarity in the latent space. Additional Graph Transformer layers on this meta-graph enable joint reasoning over intra- and inter-graph structure. The meta-graph provides built-in interpretability: supernodes and superedges highlight influential substructures and cross-graph alignments. Evaluating MGMT on both synthetic datasets and real-world neuroscience applications, we show that MGMT consistently outperforms existing state-of-the-art models in graph-level prediction tasks while offering interpretable representations that facilitate scientific discoveries. Our work establishes MGMT as a unified framework for structured multi-graph learning, advancing representation techniques in domains where graph-based data plays a central role.

Method: Proposes local-global multimodal contrastive learning with: 1) AttentiveFP encoder for local functional groups, 2) Graph Transformer for global molecular topology, 3) Self-supervised contrastive alignment between these representations, 4) Contrastive learning between chemically enriched textual descriptions and original SMILES, and 5) Dual Cross-attention multimodal fusion during fine-tuning with molecular fingerprints.

Result: Extensive experiments on MoleculeNet benchmarks show LGM-CL achieves consistent and competitive performance across both classification and regression tasks, validating the effectiveness of the unified local-global and multimodal representation learning approach.

Conclusion: The framework successfully integrates molecular structure and chemical semantics through multimodal contrastive learning, demonstrating improved molecular property prediction by capturing complementary information from different modalities.

Abstract: Accurate molecular property prediction requires integrating complementary information from molecular structure and chemical semantics. In this work, we propose LGM-CL, a local-global multimodal contrastive learning framework that jointly models molecular graphs and textual representations derived from SMILES and chemistry-aware augmented texts. Local functional group information and global molecular topology are captured using AttentiveFP and Graph Transformer encoders, respectively, and aligned through self-supervised contrastive learning. In addition, chemically enriched textual descriptions are contrasted with original SMILES to incorporate physicochemical semantics in a task-agnostic manner. During fine-tuning, molecular fingerprints are further integrated via Dual Cross-attention multimodal fusion. Extensive experiments on MoleculeNet benchmarks demonstrate that LGM-CL achieves consistent and competitive performance across both classification and regression tasks, validating the effectiveness of unified local-global and multimodal representation learning.

Method: Lethe uses Reshape-Rectify-Restore pipeline: 1) Train temporary adapter with gradient ascent on unlearning data for magnified updates, 2) Use as corrective signals for layer-wise rectification on remaining updates, 3) Remove adapter and perform short recovery on retained data.

Result: Lethe supports unlearning at all levels (client, class, sample) and maintains superior persistence with Resurfacing Rate <1% in most cases even after numerous rounds of follow-up training.

Conclusion: Lethe effectively addresses knowledge resurfacing in federated unlearning, ensuring persistent erasure during continued training through decorrelation of unlearned and retained knowledge.

Abstract: Federated unlearning (FU) aims to erase designated client-level, class-level, or sample-level knowledge from a global model. Existing studies commonly assume that the collaboration ends up with the unlearning operation, overlooking the follow-up situation where the federated training continues over the remaining data.We identify a critical failure mode, termed Knowledge resurfacing, by revealing that continued training can re-activate unlearned knowledge and cause the removed influence to resurface in the global model. To address this, we propose Lethe, a novel federated unlearning method that de-correlates knowledge to be unlearned from knowledge to be retained, ensuring persistent erasure during continued training.Lethe follows a Reshape–Rectify–Restore pipeline: a temporary adapter is first trained with gradient ascent on the unlearning data to obtain magnified updates, which is then used as corrective signals to diverge layer-wise rectification on the remaining updates in two streams. Finally, the adapter is removed and a short recovery stage is performed on the retained data. Our experiments show that Lethe supports unlearning in the federated system at all levels in a unified manner and maintains superior persistence (Resurfacing Rate <1% in most cases) even after numerous rounds of follow-up training.

Method: Uses cross-domain pre-trained time-series representation models with independent feature tuning network, meta-variable-based low rank multivariate fusion mechanism, and zero-initialized regressor for stable few-shot fine-tuning

Result: Achieves effective RUL prediction with less than 1% of target equipment samples, outperforms conventional supervised and few-shot approaches while reducing data requirements

Conclusion: Demonstrates substantial benefits from cross-domain pre-training for RUL prediction, challenging the view that knowledge transfer only works within similar devices

Abstract: The application of data-driven remaining useful life (RUL) prediction has long been constrained by the availability of large amount of degradation data. Mainstream solutions such as domain adaptation and meta-learning still rely on large amounts of historical degradation data from equipment that is identical or similar to the target, which imposes significant limitations in practical applications. This study investigates PEFT-MuTS, a Parameter-Efficient Fine-Tuning framework for few-shot RUL prediction, built on cross-domain pre-trained time-series representation models. Contrary to the widely held view that knowledge transfer in RUL prediction can only occur within similar devices, we demonstrate that substantial benefits can be achieved through pre-training process with large-scale cross-domain time series datasets. A independent feature tuning network and a meta-variable-based low rank multivariate fusion mechanism are developed to enable the pre-trained univariate time-series representation backbone model to fully exploit the multivariate relationships in degradation data for downstream RUL prediction task. Additionally, we introduce a zero-initialized regressor that stabilizes the fine-tuning process under few-shot conditions. Experiments on aero-engine and industrial bearing datasets demonstrate that our method can achieve effective RUL prediction even when less than 1% of samples of target equipment are used. Meanwhile, it substantially outperforms conventional supervised and few-shot approaches while markedly reducing the data required to achieve high predictive accuracy. Our code is available at https://github.com/fuen1590/PEFT-MuTS.

Method: Proposes consensus mechanism as a drop-in replacement for attention, formulates it as a graphical model, and introduces a hybrid consensus-attention framework. Provides extensive empirical analysis across text, DNA, and protein modalities, plus theoretical analysis characterizing consensus properties.

Result: Consensus stabilizes transformer training across a wider effective range of learning rates. The hybrid framework preserves performance while improving stability, demonstrated through learning rate sweeps on multiple modalities.

Conclusion: Consensus mechanism offers architectural solution to transformer training instability, with hybrid approach balancing stability and performance. This addresses fundamental limitation of attention-based transformers under learning rate overspecification.

Abstract: Standard attention-based transformers are known to exhibit instability under learning rate overspecification during training, particularly at high learning rates. While various methods have been proposed to improve resilience to such overspecification by modifying the optimization procedure, fundamental architectural innovations to this end remain underexplored. In this work, we illustrate that the consensus mechanism, a drop-in replacement for attention, stabilizes transformer training across a wider effective range of learning rates. We formulate consensus as a graphical model and provide extensive empirical analysis demonstrating improved stability across learning rate sweeps on text, DNA, and protein modalities. We further propose a hybrid consensus-attention framework that preserves performance while improving stability. We provide theoretical analysis characterizing the properties of consensus.

Method: A framework that dynamically interleaves formal symbolic verification with natural language generation, providing real-time feedback to detect and rectify errors. Uses a two-stage training pipeline combining formal logic verification-guided supervised fine-tuning and policy optimization.

Result: 7B and 14B models outperform state-of-the-art baselines by average margins of 10.4% and 14.2% respectively across six benchmarks spanning mathematical, logical, and general reasoning.

Conclusion: Formal verification can serve as a scalable mechanism to significantly improve LLM reasoning performance by actively penalizing intermediate fallacies during the reasoning chain.

Abstract: Large Language Models (LLMs) show remarkable capabilities, yet their stochastic next-token prediction creates logical inconsistencies and reward hacking that formal symbolic systems avoid. To bridge this gap, we introduce a formal logic verification-guided framework that dynamically interleaves formal symbolic verification with the natural language generation process, providing real-time feedback to detect and rectify errors as they occur. Distinguished from previous neuro-symbolic methods limited by passive post-hoc validation, our approach actively penalizes intermediate fallacies during the reasoning chain. We operationalize this framework via a novel two-stage training pipeline that synergizes formal logic verification-guided supervised fine-tuning and policy optimization. Extensive evaluation on six benchmarks spanning mathematical, logical, and general reasoning demonstrates that our 7B and 14B models outperform state-of-the-art baselines by average margins of 10.4% and 14.2%, respectively. These results validate that formal verification can serve as a scalable mechanism to significantly push the performance boundaries of advanced LLM reasoning.

Method: GUDA approximates counterfactual models by applying machine unlearning to a shared full-data model instead of training from scratch. It quantifies group influence using differences in ELBO (evidence lower bound) scores between the full model and each unlearned counterfactual model.

Result: Experiments on CIFAR-10 and artistic style attribution with Stable Diffusion show GUDA identifies primary contributing groups more reliably than semantic similarity, gradient-based attribution, and instance-level unlearning approaches, while achieving 100x speedup on CIFAR-10 over LOGO retraining.

Conclusion: GUDA provides an efficient and effective solution for group-level training data attribution in diffusion models, bridging the gap between instance-level attribution and practical group-level analysis needs.

Abstract: Training-data attribution for vision generative models aims to identify which training data influenced a given output. While most methods score individual examples, practitioners often need group-level answers (e.g., artistic styles or object classes). Group-wise attribution is counterfactual: how would a model’s behavior on a generated sample change if a group were absent from training? A natural realization of this counterfactual is Leave-One-Group-Out (LOGO) retraining, which retrains the model with each group removed; however, it becomes computationally prohibitive as the number of groups grows. We propose GUDA (Group Unlearning-based Data Attribution) for diffusion models, which approximates each counterfactual model by applying machine unlearning to a shared full-data model instead of training from scratch. GUDA quantifies group influence using differences in a likelihood-based scoring rule (ELBO) between the full model and each unlearned counterfactual. Experiments on CIFAR-10 and artistic style attribution with Stable Diffusion show that GUDA identifies primary contributing groups more reliably than semantic similarity, gradient-based attribution, and instance-level unlearning approaches, while achieving x100 speedup on CIFAR-10 over LOGO retraining.

Method: StoMPP (Stochastic Masked Partial Progressive Binarization) uses layerwise stochastic masking to progressively replace differentiable clipped weights/activations with hard binary step functions. It only backpropagates through the unfrozen (clipped) subset, avoiding straight-through estimators entirely.

Result: StoMPP outperforms BinaryConnect-style STE baselines, with gains increasing with depth: +18.0 on CIFAR-10, +13.5 on CIFAR-100, and +3.8 on ImageNet for ResNet-50 BNN. For binary-weight networks, achieves 91.2% accuracy on CIFAR-10 and 69.5% on CIFAR-100 with ResNet-50.

Conclusion: Progressive freezing with stochastic masking (StoMPP) provides an effective alternative to straight-through estimators for training binary neural networks, improving accuracy and depth scaling under binarization constraints.

Abstract: We investigate progressive freezing as an alternative to straight-through estimators (STE) for training binary networks from scratch. Under controlled training conditions, we find that while global progressive freezing works for binary-weight networks, it fails for full binary neural networks due to activation-induced gradient blockades. We introduce StoMPP (Stochastic Masked Partial Progressive Binarization), which uses layerwise stochastic masking to progressively replace differentiable clipped weights/activations with hard binary step functions, while only backpropagating through the unfrozen (clipped) subset (i.e., no straight-through estimator). Under a matched minimal training recipe, StoMPP improves accuracy over a BinaryConnect-style STE baseline, with gains that increase with depth (e.g., for ResNet-50 BNN: +18.0 on CIFAR-10, +13.5 on CIFAR-100, and +3.8 on ImageNet; for ResNet-18: +3.1, +4.7, and +1.3). For binary-weight networks, StoMPP achieves 91.2% accuracy on CIFAR-10 and 69.5% on CIFAR-100 with ResNet-50. We analyze training dynamics under progressive freezing, revealing non-monotonic convergence and improved depth scaling under binarization constraints.

Method: Proposes monitoring task vector’s growth rate using solely server-side parameters to determine optimal stopping point without validation data. Framework works with various state-of-the-art FL methods.

Result: Achieves comparable performance to validation-based early stopping on skin lesion and blood cell classification tasks. Spends average of 47/20 rounds to achieve over 12.5%/10.3% higher performance than validation-based early stopping.

Conclusion: First work to propose early stopping framework for FL without using validation data, offering practical solution for FL deployment with reduced computational costs and privacy risks.

Abstract: Federated Learning (FL) facilitates decentralized collaborative learning without transmitting raw data. However, reliance on fixed global rounds or validation data for hyperparameter tuning hinders practical deployment by incurring high computational costs and privacy risks. To address this, we propose a data-free early stopping framework that determines the optimal stopping point by monitoring the task vector’s growth rate using solely server-side parameters. The numerical results on skin lesion/blood cell classification demonstrate that our approach is comparable to validation-based early stopping across various state-of-the-art FL methods. In particular, the proposed framework spends an average of 47/20 (skin lesion/blood cell) rounds to achieve over 12.5%/10.3% higher performance than early stopping based on validation data. To the best of our knowledge, this is the first work to propose an early stopping framework for FL methods without using any validation data.

Method: Uses empirical and theoretical analyses with novel generalization analysis using Wasserstein distance to study graph structure impact, specifically fan-out size. Examines non-isotropic effects of batch size and fan-out size on GNN convergence and generalization.

Result: Reveals that full-graph training doesn’t always yield better model performance or computational efficiency than well-tuned smaller mini-batch settings. Provides practical guidance for tuning batch size and fan-out size under resource constraints.

Conclusion: The choice between full-graph and mini-batch GNN training depends on careful tuning of batch size and fan-out size, with smaller mini-batch settings potentially outperforming full-graph training when properly configured.

Abstract: Full-graph and mini-batch Graph Neural Network (GNN) training approaches have distinct system design demands, making it crucial to choose the appropriate approach to develop. A core challenge in comparing these two GNN training approaches lies in characterizing their model performance (i.e., convergence and generalization) and computational efficiency. While a batch size has been an effective lens in analyzing such behaviors in deep neural networks (DNNs), GNNs extend this lens by introducing a fan-out size, as full-graph training can be viewed as mini-batch training with the largest possible batch size and fan-out size. However, the impact of the batch and fan-out size for GNNs remains insufficiently explored. To this end, this paper systematically compares full-graph vs. mini-batch training of GNNs through empirical and theoretical analyses from the view points of the batch size and fan-out size. Our key contributions include: 1) We provide a novel generalization analysis using the Wasserstein distance to study the impact of the graph structure, especially the fan-out size. 2) We uncover the non-isotropic effects of the batch size and the fan-out size in GNN convergence and generalization, providing practical guidance for tuning these hyperparameters under resource constraints. Finally, full-graph training does not always yield better model performance or computational efficiency than well-tuned smaller mini-batch settings. The implementation can be found in the github link: https://github.com/LIUMENGFAN-gif/GNN_fullgraph_minibatch_training.

Method: Theoretical analysis of consistency models from a flow map-based perspective to understand training stability and convergence behavior. Revisits self-distillation as practical remedy for suboptimal convergence, reformulating it to avoid excessive gradient norms for stable optimization.

Result: The analysis clarifies how training instability leads to degenerate solutions. The improved self-distillation strategy enables stable training without reliance on pretrained diffusion models for initialization, extending applicability to diffusion-based policy learning.

Conclusion: Flow map-based analysis provides unified theoretical understanding of consistency model instability. The proposed self-distillation reformulation enables stable training from scratch, broadening applicability beyond image generation to other domains like policy learning.

Abstract: Consistency models have been proposed for fast generative modeling, achieving results competitive with diffusion and flow models. However, these methods exhibit inherent instability and limited reproducibility when training from scratch, motivating subsequent work to explain and stabilize these issues. While these efforts have provided valuable insights, the explanations remain fragmented, and the theoretical relationships remain unclear. In this work, we provide a theoretical examination of consistency models by analyzing them from a flow map-based perspective. This joint analysis clarifies how training stability and convergence behavior can give rise to degenerate solutions. Building on these insights, we revisit self-distillation as a practical remedy for certain forms of suboptimal convergence and reformulate it to avoid excessive gradient norms for stable optimization. We further demonstrate that our strategy extends beyond image generation to diffusion-based policy learning, without reliance on a pretrained diffusion model for initialization, thereby illustrating its broader applicability.

Method: Introduces unified interpretation of periodicity from abstract algebra and reasoning perspective (single and composite periodicity). Constructs Coper benchmark for composite periodicity with two OOD settings: Hollow and Extrapolation. Conducts experiments to test Transformers’ periodicity generalization capabilities.

Result: Experiments reveal that periodicity generalization in Transformers is limited - models can memorize periodic data during training but cannot generalize to unseen composite periodicity in OOD scenarios.

Conclusion: Transformers have fundamental limitations in periodicity generalization, highlighting a gap in OOD capabilities compared to humans. The Coper benchmark and analysis provide insights into why Transformers struggle with this basic reasoning task.

Abstract: Large language models (LLMs) based on the Transformer have demonstrated strong performance across diverse tasks. However, current models still exhibit substantial limitations in out-of-distribution (OOD) generalization compared with humans. We investigate this gap through periodicity, one of the basic OOD scenarios. Periodicity captures invariance amid variation. Periodicity generalization represents a model’s ability to extract periodic patterns from training data and generalize to OOD scenarios. We introduce a unified interpretation of periodicity from the perspective of abstract algebra and reasoning, including both single and composite periodicity, to explain why Transformers struggle to generalize periodicity. Then we construct Coper about composite periodicity, a controllable generative benchmark with two OOD settings, Hollow and Extrapolation. Experiments reveal that periodicity generalization in Transformers is limited, where models can memorize periodic data during training, but cannot generalize to unseen composite periodicity. We release the source code to support future research.

Method: Operationalizes the METRIC-framework with a collection of data quality metrics (metric library), each documented with metric cards containing definitions, applicability, examples, pitfalls, and recommendations. Provides decision trees for selecting appropriate metrics for specific use cases.

Result: Demonstrates the approach on the PTB-XL ECG dataset, showing practical application of data quality evaluation for medical ML tasks.

Conclusion: Provides a practical toolkit for fit-for-purpose evaluation of training and test data, representing a first step toward establishing trustworthy AI in medicine through systematic data quality assessment.

Abstract: Machine learning (ML) in medicine has transitioned from research to concrete applications aimed at supporting several medical purposes like therapy selection, monitoring and treatment. Acceptance and effective adoption by clinicians and patients, as well as regulatory approval, require evidence of trustworthiness. A major factor for the development of trustworthy AI is the quantification of data quality for AI model training and testing. We have recently proposed the METRIC-framework for systematically evaluating the suitability (fit-for-purpose) of data for medical ML for a given task. Here, we operationalize this theoretical framework by introducing a collection of data quality metrics - the metric library - for practically measuring data quality dimensions. For each metric, we provide a metric card with the most important information, including definition, applicability, examples, pitfalls and recommendations, to support the understanding and implementation of these metrics. Furthermore, we discuss strategies and provide decision trees for choosing an appropriate set of data quality metrics from the metric library given specific use cases. We demonstrate the impact of our approach exemplarily on the PTB-XL ECG-dataset. This is a first step to enable fit-for-purpose evaluation of training and test data in practice as the base for establishing trustworthy AI in medicine.

Method: SQUAD integrates early-exit mechanisms with distributed ensemble learning using quorum-based stopping criteria. It collects intermediate predictions incrementally until consensus is reached. QUEST (Quorum Search Technique) is a Neural Architecture Search method that selects early-exit learners with optimized hierarchical diversity to ensure complementary predictions at each layer.

Result: Improves test accuracy up to 5.95% compared to state-of-the-art dynamic solutions with comparable computational cost, and reduces inference latency up to 70.60% compared to static ensembles while maintaining good accuracy.

Conclusion: The consensus-driven approach with optimized hierarchical diversity yields statistically robust early exits, providing better accuracy-latency trade-offs than existing methods through improved uncertainty estimation.

Abstract: Early-exit neural networks have become popular for reducing inference latency by allowing intermediate predictions when sufficient confidence is achieved. However, standard approaches typically rely on single-model confidence thresholds, which are frequently unreliable due to inherent calibration issues. To address this, we introduce SQUAD (Scalable Quorum Adaptive Decisions), the first inference scheme that integrates early-exit mechanisms with distributed ensemble learning, improving uncertainty estimation while reducing the inference time. Unlike traditional methods that depend on individual confidence scores, SQUAD employs a quorum-based stopping criterion on early-exit learners by collecting intermediate predictions incrementally in order of computational complexity until a consensus is reached and halting the computation at that exit if the consensus is statistically significant. To maximize the efficacy of this voting mechanism, we also introduce QUEST (Quorum Search Technique), a Neural Architecture Search method to select early-exit learners with optimized hierarchical diversity, ensuring learners are complementary at every intermediate layer. This consensus-driven approach yields statistically robust early exits, improving the test accuracy up to 5.95% compared to state-of-the-art dynamic solutions with a comparable computational cost and reducing the inference latency up to 70.60% compared to static ensembles while maintaining a good accuracy.

Method: Proposes using Vision-Language Models (VLMs) to generate promptable representations that separate controllable changes from noise in an unsupervised way. These representations serve as targets during LAM training. Benchmarks various VLMs to evaluate their representation quality and robustness to different prompts/hyperparameters.

Result: Found substantial variation in VLM performance for promptable representations, with newer VLMs sometimes performing worse than older ones. Simply asking VLMs to ignore distractors significantly improves latent action quality, yielding up to 6x increase in downstream success rates on Distracting MetaWorld benchmark.

Conclusion: VLMs can effectively provide promptable representations that filter out action-correlated distractors, substantially improving Latent Action Models. The approach demonstrates the importance of proper prompting and reveals surprising performance variations among different VLMs.

Abstract: Latent Action Models (LAMs) have rapidly gained traction as an important component in the pre-training pipelines of leading Vision-Language-Action models. However, they fail when observations contain action-correlated distractors, often encoding noise instead of meaningful latent actions. Humans, on the other hand, can effortlessly distinguish task-relevant motions from irrelevant details in any video given only a brief task description. In this work, we propose to utilize the common-sense reasoning abilities of Vision-Language Models (VLMs) to provide promptable representations, effectively separating controllable changes from the noise in unsupervised way. We use these representations as targets during LAM training and benchmark a wide variety of popular VLMs, revealing substantial variation in the quality of promptable representations as well as their robustness to different prompts and hyperparameters. Interestingly, we find that more recent VLMs may perform worse than older ones. Finally, we show that simply asking VLMs to ignore distractors can substantially improve latent action quality, yielding up to a six-fold increase in downstream success rates on Distracting MetaWorld.

Method: Proposes Low-Rank Decomposed Scaling (LoRDS), modeling scaling manifold as continuous low-rank matrices (S = BA). This breaks spatial constraints of blocks, enabling high-fidelity PTQ initialization, iterative optimization, joint QAT of weights and scaling factors, and high-rank multiplicative PEFT adaptation.

Result: Outperforms state-of-the-art baselines across various model families. On Llama3-8B: achieves 27.0% accuracy improvement at 3 bits over NormalFloat quantization, 1.5x inference speedup on RTX 4090, and 9.6% PEFT performance improvement on downstream tasks over 4bit QLoRA.

Conclusion: LoRDS offers a robust, integrated solution for unified compression and adaptation of LLMs, providing element-wise quantization efficiency comparable to block-wise methods with superior expressive power and no additional inference overhead.

Abstract: Current quantization methods for LLMs predominantly rely on block-wise structures to maintain efficiency, often at the cost of representational flexibility. In this work, we demonstrate that element-wise quantization can be made as efficient as block-wise scaling while providing strictly superior expressive power by modeling the scaling manifold as continuous low-rank matrices ($S = BA$). We propose Low-Rank Decomposed Scaling (LoRDS), a unified framework that rethinks quantization granularity through this low-rank decomposition. By “breaking the blocks” of spatial constraints, LoRDS establishes a seamless efficiency lifecycle: it provides high-fidelity PTQ initialization refined via iterative optimization, enables joint QAT of weights and scaling factors, and facilitates high-rank multiplicative PEFT adaptation. Unlike additive PEFT approaches such as QLoRA, LoRDS enables high-rank weight updates within a low-rank budget while incurring no additional inference overhead. Supported by highly optimized Triton kernels, LoRDS consistently outperforms state-of-the-art baselines across various model families in both quantization and downstream fine-tuning tasks. Notably, on Llama3-8B, our method achieves up to a 27.0% accuracy improvement at 3 bits over NormalFloat quantization and delivers a 1.5x inference speedup on NVIDIA RTX 4090 while enhancing PEFT performance by 9.6% on downstream tasks over 4bit QLoRA, offering a robust and integrated solution for unified compression and adaptation of LLMs.

Method: Analyzed neural networks’ generalization performance, their representational alignment with each other, and their alignment with human neural activity. Investigated geometric properties of learned representations, particularly local vs. global intrinsic dimension measures.

Result: Found that generalization performance, model-model alignment, and model-brain alignment are all significantly correlated. Lower local intrinsic dimension consistently associated with stronger alignment and better generalization. Increasing model capacity and training data reduces local intrinsic dimension.

Conclusion: Local intrinsic dimension serves as a unifying descriptor of representational convergence in both artificial and biological neural systems, explaining the benefits of scaling in neural networks.

Abstract: Recent work has found that neural networks with stronger generalization tend to exhibit higher representational alignment with one another across architectures and training paradigms. In this work, we show that models with stronger generalization also align more strongly with human neural activity. Moreover, generalization performance, model–model alignment, and model–brain alignment are all significantly correlated with each other. We further show that these relationships can be explained by a single geometric property of learned representations: the local intrinsic dimension of embeddings. Lower local dimension is consistently associated with stronger model–model alignment, stronger model–brain alignment, and better generalization, whereas global dimension measures fail to capture these effects. Finally, we find that increasing model capacity and training data scale systematically reduces local intrinsic dimension, providing a geometric account of the benefits of scaling. Together, our results identify local intrinsic dimension as a unifying descriptor of representational convergence in artificial and biological systems.

Method: Restructures single LoRA module as decomposable Rank-1 Expert Pool; learns dynamic composition of sparse task-specific updates guided by [CLS] token semantics; uses Activation-Guided Orthogonal (AGO) loss to orthogonalize critical LoRA weights across tasks; enables domain-aware learning with minimal inter-task interference.

Result: State-of-the-art results across multiple settings, surpassing zero-shot upper bounds in generalization; reduces trainable parameters by 96.7% compared to baseline; eliminates need for external datasets or task-ID discriminators; merged LoRAs retain fewer weights with no inference latency.

Conclusion: The proposed framework provides computationally lightweight continual learning for VLMs through sparse composition and orthogonalization, achieving strong performance while minimizing forgetting and parameter overhead.

Abstract: Continual learning (CL) in vision-language models (VLMs) faces significant challenges in improving task adaptation and avoiding catastrophic forgetting. Existing methods usually have heavy inference burden or rely on external knowledge, while Low-Rank Adaptation (LoRA) has shown potential in reducing these issues by enabling parameter-efficient tuning. However, considering directly using LoRA to alleviate the catastrophic forgetting problem is non-trivial, we introduce a novel framework that restructures a single LoRA module as a decomposable Rank-1 Expert Pool. Our method learns to dynamically compose a sparse, task-specific update by selecting from this expert pool, guided by the semantics of the [CLS] token. In addition, we propose an Activation-Guided Orthogonal (AGO) loss that orthogonalizes critical parts of LoRA weights across tasks. This sparse composition and orthogonalization enable fewer parameter updates, resulting in domain-aware learning while minimizing inter-task interference and maintaining downstream task performance. Extensive experiments across multiple settings demonstrate state-of-the-art results in all metrics, surpassing zero-shot upper bounds in generalization. Notably, it reduces trainable parameters by 96.7% compared to the baseline method, eliminating reliance on external datasets or task-ID discriminators. The merged LoRAs retain less weights and incur no inference latency, making our method computationally lightweight.

Method: Proposes considering a confidence set around the empirical observational distribution and obtaining the intersection of causal effect bounds for all distributions in this set. Uses neural causal models to solve min-max and max-min problems over all possible distributions and structural causal models.

Result: Extensive experiments on synthetic and real-world datasets show the algorithm can determine when collecting more samples won’t help identify the best action, guiding practitioners to collect more variables or consider randomized studies.

Conclusion: The framework provides a principled way to separate sample uncertainty from non-identifiability uncertainty in causal inference, helping practitioners make informed decisions about data collection strategies.

Abstract: Causal inference from observational data provides strong evidence for the best action in decision-making without performing expensive randomized trials. The effect of an action is usually not identifiable under unobserved confounding, even with an infinite amount of data. Recent work uses neural networks to obtain practical bounds to such causal effects, which is often an intractable problem. However, these approaches may overfit to the dataset and be overconfident in their causal effect estimates. Moreover, there is currently no systematic approach to disentangle how much of the width of causal effect bounds is due to fundamental non-identifiability versus how much is due to finite-sample limitations. We propose a novel framework to address this problem by considering a confidence set around the empirical observational distribution and obtaining the intersection of causal effect bounds for all distributions in this confidence set. This allows us to distinguish the part of the interval that can be reduced by collecting more samples, which we call sample uncertainty, from the part that can only be reduced by observing more variables, such as latent confounders or instrumental variables, but not with more data, which we call non-ID uncertainty. The upper and lower bounds to this intersection are obtained by solving min-max and max-min problems with neural causal models by searching over all distributions that the dataset might have been sampled from, and all SCMs that entail the corresponding distribution. We demonstrate via extensive experiments on synthetic and real-world datasets that our algorithm can determine when collecting more samples will not help determine the best action. This can guide practitioners to collect more variables or lean towards a randomized study for best action identification.

Method: Builds on Fenchel-Young framework to situate Softmax within broader surrogate family, analyzes gradient dynamics for convergence behaviors, introduces systematic bias-variance decomposition for approximate methods with convergence guarantees, and provides per-epoch complexity analysis.

Result: Extensive experiments show strong alignment between theoretical consistency, convergence properties, and empirical performance, establishing principled foundation for loss selection in large-class applications.

Conclusion: The paper offers practical guidance for loss function selection in large-class machine learning by establishing connections between theoretical properties (consistency, convergence) and empirical performance, with explicit trade-offs between effectiveness and efficiency.

Abstract: The Softmax loss is one of the most widely employed surrogate objectives for classification and ranking tasks. To elucidate its theoretical properties, the Fenchel-Young framework situates it as a canonical instance within a broad family of surrogates. Concurrently, another line of research has addressed scalability when the number of classes is exceedingly large, in which numerous approximations have been proposed to retain the benefits of the exact objective while improving efficiency. Building on these two perspectives, we present a principled investigation of the Softmax-family losses. We examine whether different surrogates achieve consistency with classification and ranking metrics, and analyze their gradient dynamics to reveal distinct convergence behaviors. We also introduce a systematic bias-variance decomposition for approximate methods that provides convergence guarantees, and further derive a per-epoch complexity analysis, showing explicit trade-offs between effectiveness and efficiency. Extensive experiments on a representative task demonstrate a strong alignment between consistency, convergence, and empirical performance. Together, these results establish a principled foundation and offer practical guidance for loss selections in large-class machine learning applications.

Method: Introduces physics-informed Müntz-Szász Networks (MSN-PINN) using power-law basis functions with trainable scaling exponents. The model outputs both solutions and scaling structures, with constraint-aware training to encode physical requirements like boundary condition compatibility.

Result: Achieves single-exponent recovery with 1-5% error under noise and sparse sampling, recovers corner singularity exponents for 2D Laplace equation with 0.009% error, matches classical Kondrat’ev results, and reaches 100% success rate on 40-configuration wedge benchmark with 0.022% mean error.

Conclusion: MSN-PINN combines neural network expressiveness with asymptotic analysis interpretability, producing learned parameters with direct physical meaning and significantly improving accuracy over naive training approaches.

Abstract: Physical systems near singularities, interfaces, and critical points exhibit power-law scaling, yet standard neural networks leave the governing exponents implicit. We introduce physics-informed M"untz-Sz’asz Networks (MSN-PINN), a power-law basis network that treats scaling exponents as trainable parameters. The model outputs both the solution and its scaling structure. We prove identifiability, or unique recovery, and show that, under these conditions, the squared error between learned and true exponents scales as $O(|μ- α|^2)$. Across experiments, MSN-PINN achieves single-exponent recovery with 1–5% error under noise and sparse sampling. It recovers corner singularity exponents for the two-dimensional Laplace equation with 0.009% error, matches the classical result of Kondrat’ev (1967), and recovers forcing-induced exponents in singular Poisson problems with 0.03% and 0.05% errors. On a 40-configuration wedge benchmark, it reaches a 100% success rate with 0.022% mean error. Constraint-aware training encodes physical requirements such as boundary condition compatibility and improves accuracy by three orders of magnitude over naive training. By combining the expressiveness of neural networks with the interpretability of asymptotic analysis, MSN-PINN produces learned parameters with direct physical meaning.

Method: OSNIP generalizes linear kernel geometry to high-dimensional LLM latent spaces, defining an “Obfuscated Semantic Null Space” that preserves semantics while enforcing near-orthogonality to original embeddings. It injects perturbations that project embeddings into this space using key-dependent stochastic mapping for individualized user trajectories.

Result: Evaluations on 12 generative and classification benchmarks show state-of-the-art performance, sharply reducing attack success rates while maintaining strong model utility under strict security constraints.

Conclusion: OSNIP provides an effective lightweight client-side encryption framework for privacy-preserving LLM inference that balances privacy protection with model utility through geometric perturbation techniques.

Abstract: We propose Obfuscated Semantic Null space Injection for Privacy (OSNIP), a lightweight client-side encryption framework for privacy-preserving LLM inference. Generalizing the geometric intuition of linear kernels to the high-dimensional latent space of LLMs, we formally define the ``Obfuscated Semantic Null Space’’, a high-dimensional regime that preserves semantic fidelity while enforcing near-orthogonality to the original embedding. By injecting perturbations that project the original embedding into this space, OSNIP ensures privacy without any post-processing. Furthermore, OSNIP employs a key-dependent stochastic mapping that synthesizes individualized perturbation trajectories unique to each user. Evaluations on 12 generative and classification benchmarks show that OSNIP achieves state-of-the-art performance, sharply reducing attack success rates while maintaining strong model utility under strict security constraints.

Method: The authors analyze how the geometry of learned embeddings controls predictive performance. They develop a theoretical framework showing population risk can be bounded by: (1) intrinsic dimension of embedding distribution (affecting convergence rate in Wasserstein distance), and (2) sensitivity of downstream mapping from embeddings to predictions (characterized by Lipschitz constants).

Result: The theory yields embedding-dependent error bounds that don’t rely on parameter counts. At the final embedding layer, architectural sensitivity vanishes and the bound is dominated by embedding dimension, explaining its strong empirical correlation with generalization performance. Experiments across architectures and datasets validate the theory.

Conclusion: Generalization in deep networks can be better understood through embedding geometry rather than parameter counts. The embedding dimension emerges as a key factor controlling generalization performance, providing new diagnostics and theoretical insights.

Abstract: Deep neural networks often generalize well despite heavy over-parameterization, challenging classical parameter-based analyses. We study generalization from a representation-centric perspective and analyze how the geometry of learned embeddings controls predictive performance for a fixed trained model. We show that population risk can be bounded by two factors: (i) the intrinsic dimension of the embedding distribution, which determines the convergence rate of empirical embedding distribution to the population distribution in Wasserstein distance, and (ii) the sensitivity of the downstream mapping from embeddings to predictions, characterized by Lipschitz constants. Together, these yield an embedding-dependent error bound that does not rely on parameter counts or hypothesis class complexity. At the final embedding layer, architectural sensitivity vanishes and the bound is dominated by embedding dimension, explaining its strong empirical correlation with generalization performance. Experiments across architectures and datasets validate the theory and demonstrate the utility of embedding-based diagnostics.

Method: Proposes MetaDrug with two-level meta-adaptation: self-adaptation (uses patient’s own medical events as support sets to capture temporal dependencies) and peer-adaptation (uses similar visits from peer patients to enrich representations). Includes uncertainty quantification module to rank support visits and filter unrelated information for adaptation consistency.

Result: MetaDrug consistently outperforms state-of-the-art medication recommendation methods on cold-start patients across MIMIC-III and Acute Kidney Injury (AKI) datasets.

Conclusion: MetaDrug effectively addresses patient cold-start problem in medication recommendation through multi-level meta-learning with uncertainty awareness, demonstrating superior performance over existing methods.

Abstract: Large-scale Electronic Health Record (EHR) databases have become indispensable in supporting clinical decision-making through data-driven treatment recommendations. However, existing medication recommender methods often struggle with a user (i.e., patient) cold-start problem, where recommendations for new patients are usually unreliable due to the lack of sufficient prescription history for patient profiling. While prior studies have utilized medical knowledge graphs to connect medication concepts through pharmacological or chemical relationships, these methods primarily focus on mitigating the item cold-start issue and fall short in providing personalized recommendations that adapt to individual patient characteristics. Meta-learning has shown promise in handling new users with sparse interactions in recommender systems. However, its application to EHRs remains underexplored due to the unique sequential structure of EHR data. To tackle these challenges, we propose MetaDrug, a multi-level, uncertainty-aware meta-learning framework designed to address the patient cold-start problem in medication recommendation. MetaDrug proposes a novel two-level meta-adaptation mechanism, including self-adaptation, which adapts the model to new patients using their own medical events as support sets to capture temporal dependencies; and peer-adaptation, which adapts the model using similar visits from peer patients to enrich new patient representations. Meanwhile, to further improve meta-adaptation outcomes, we introduce an uncertainty quantification module that ranks the support visits and filters out the unrelated information for adaptation consistency. We evaluate our approach on the MIMIC-III and Acute Kidney Injury (AKI) datasets. Experimental results on both datasets demonstrate that MetaDrug consistently outperforms state-of-the-art medication recommendation methods on cold-start patients.

Method: Trained 300 models with over 10,000 experiments, rigorously controlling compute budgets while independently varying model size, number of training tokens, and molecular representation to systematically investigate scaling behavior.

Result: Demonstrated clear scaling laws in molecular models for both pretraining and downstream transfer, revealed substantial impact of molecular representation on performance, and explained previously observed inconsistencies in scaling behavior for molecular generation.

Conclusion: Molecular language models do follow predictable scaling laws, with representation choice being a critical factor; the study provides the largest library of molecular language models to date for future research.

Abstract: Molecular generative models, often employing GPT-style language modeling on molecular string representations, have shown promising capabilities when scaled to large datasets and model sizes. However, it remains unclear and subject to debate whether these models adhere to predictable scaling laws under fixed computational budgets, which is a crucial understanding for optimally allocating resources between model size, data volume, and molecular representation. In this study, we systematically investigate the scaling behavior of molecular language models across both pretraining and downstream tasks. We train 300 models and conduct over 10,000 experiments, rigorously controlling compute budgets while independently varying model size, number of training tokens, and molecular representation. Our results demonstrate clear scaling laws in molecular models for both pretraining and downstream transfer, reveal the substantial impact of molecular representation on performance, and explain previously observed inconsistencies in scaling behavior for molecular generation. Additionally, we publicly release the largest library of molecular language models to date to facilitate future research and development. Code and models are available at https://github.com/SZU-ADDG/MLM-Scaling.

Method: Proposes a unified definition of behavior style and develops SCIQL framework that leverages offline goal-conditioned RL techniques (hindsight relabeling and value learning) combined with a novel Gated Advantage Weighted Regression mechanism to optimize both task performance and style alignment.

Result: SCIQL achieves superior performance on both style alignment and task objectives compared to prior offline methods, as demonstrated through experiments.

Conclusion: The proposed SCIQL framework successfully addresses the challenge of learning style-conditioned policies in offline RL by providing a unified style definition and effective optimization mechanism that balances style preservation with task performance.

Abstract: We study offline reinforcement learning of style-conditioned policies using explicit style supervision via subtrajectory labeling functions. In this setting, aligning style with high task performance is particularly challenging due to distribution shift and inherent conflicts between style and reward. Existing methods, despite introducing numerous definitions of style, often fail to reconcile these objectives effectively. To address these challenges, we propose a unified definition of behavior style and instantiate it into a practical framework. Building on this, we introduce Style-Conditioned Implicit Q-Learning (SCIQL), which leverages offline goal-conditioned RL techniques, such as hindsight relabeling and value learning, and combine it with a new Gated Advantage Weighted Regression mechanism to efficiently optimize task performance while preserving style alignment. Experiments demonstrate that SCIQL achieves superior performance on both objectives compared to prior offline methods. Code, datasets and visuals are available in: https://sciql-iclr-2026.github.io/.

Method: The authors establish mathematical correspondences between sparse attention mechanisms and compact kernel regression. They show normalized ReLU and sparsemax attention arise from Epanechnikov kernel regression under different normalizations, and demonstrate that α-entmax attention corresponds to various bounded-support kernels (Epanechnikov, biweight, triweight) with α = 1 + 1/n. They validate their framework with Memory Mosaics, a kernel-regression-based variant of transformers.

Result: The paper shows that kernel-based sparse attention achieves competitive performance on language modeling, in-context learning, and length generalization tasks. The unified perspective explains how sparsity naturally emerges from kernel design and provides principled alternatives to heuristic top-k attention mechanisms.

Conclusion: The work establishes a formal kernel-theoretic foundation for sparse attention mechanisms, showing they correspond to compact kernel regression. This provides a principled framework for designing attention mechanisms beyond heuristic approaches, with practical applications in transformer architectures.

Abstract: Recent work has revealed a link between self-attention mechanisms in transformers and test-time kernel regression via the Nadaraya-Watson estimator, with standard softmax attention corresponding to a Gaussian kernel. However, a kernel-theoretic understanding of sparse attention mechanisms is currently missing. In this paper, we establish a formal correspondence between sparse attention and compact (bounded support) kernels. We show that normalized ReLU and sparsemax attention arise from Epanechnikov kernel regression under fixed and adaptive normalizations, respectively. More generally, we demonstrate that widely used kernels in nonparametric density estimation – including Epanechnikov, biweight, and triweight – correspond to $α$-entmax attention with $α= 1 + \frac{1}{n}$ for $n \in \mathbb{N}$, while the softmax/Gaussian relationship emerges in the limit $n \to \infty$. This unified perspective explains how sparsity naturally emerges from kernel design and provides principled alternatives to heuristic top-$k$ attention and other associative memory mechanisms. Experiments with a kernel-regression-based variant of transformers – Memory Mosaics – show that kernel-based sparse attention achieves competitive performance on language modeling, in-context learning, and length generalization tasks, offering a principled framework for designing attention mechanisms.

Method: EntQuant decouples numerical precision from storage cost using entropy coding, allowing compression of large models (70B parameters) in under 30 minutes. It achieves data-dependent method performance with data-free method speed and universality.

Result: State-of-the-art results on standard evaluation sets and models, while retaining functional performance on complex benchmarks with instruction-tuned models. Maintains modest inference overhead despite extreme compression.

Conclusion: EntQuant successfully bridges the gap between data-free and data-dependent compression methods, enabling practical extreme compression with both speed and fidelity, making large model deployment more accessible.

Abstract: Post-training compression is currently divided into two contrasting regimes. On the one hand, fast, data-free, and model-agnostic methods (e.g., NF4 or HQQ) offer maximum accessibility but suffer from functional collapse at extreme bit-rates below 4 bits. On the other hand, techniques leveraging calibration data or extensive recovery training achieve superior fidelity but impose high computational constraints and face uncertain robustness under data distribution shifts. We introduce EntQuant, the first framework to unite the advantages of these distinct paradigms. By matching the performance of data-dependent methods with the speed and universality of data-free techniques, EntQuant enables practical utility in the extreme compression regime. Our method decouples numerical precision from storage cost via entropy coding, compressing a 70B parameter model in less than 30 minutes. We demonstrate that EntQuant does not only achieve state-of-the-art results on standard evaluation sets and models, but also retains functional performance on more complex benchmarks with instruction-tuned models, all at modest inference overhead.

Method: CFPO replaces heuristic clipping with a convex quadratic penalty derived from Total Variation divergence constraints, creating an everywhere-differentiable objective that enforces stable policy updates without hard boundaries.

Result: In reasoning tasks, CFPO matches clipping-based methods on downstream benchmarks while extending stable training. In alignment, it mitigates verbosity exploitation, reduces capability degradation, and achieves competitive instruction-following performance.

Conclusion: CFPO is a promising drop-in alternative to clipping-based methods for LLM post-training, requiring only a one-line code change and no additional hyperparameters.

Abstract: Reinforcement learning has become central to post-training large language models, yet dominant algorithms rely on clipping mechanisms that introduce optimization issues at scale, including zero-gradient regions, reward hacking, and training instability. We propose Clipping-Free Policy Optimization (CFPO), which replaces heuristic clipping with a convex quadratic penalty derived from Total Variation divergence constraints, yielding an everywhere-differentiable objective that enforces stable policy updates without hard boundaries. We evaluate CFPO across both reasoning and alignment settings. In reasoning, CFPO matches clipping-based methods on downstream benchmarks while extending the stable training regime. In alignment, CFPO mitigates verbosity exploitation and reduces capability degradation, while achieving competitive instruction-following performance. CFPO requires only a one-line code change and no additional hyperparameters. Our results suggest that CFPO is a promising drop-in alternative to clipping-based methods for LLM post-training.

Method: Proposes MS-EDEN (Micro-Scaled EDEN), a novel unbiased quantization routine for micro-scaled formats with 2x lower quantization error than stochastic rounding. Integrates this into Quartet II, a fully-NVFP4 quantization scheme for linear layers that improves gradient estimation across all major matrix multiplications in forward and backward passes.

Result: Quartet II achieves consistently better gradient estimation, synergizes with recent NVFP4 training improvements, and validates on end-to-end LLM training up to 1.9B parameters on 38B tokens. Provides kernels for NVIDIA Blackwell GPUs with up to 4.2x speedup over BF16.

Conclusion: Quartet II with MS-EDEN improves state-of-the-art for quantized training in NVFP4 format, enabling more accurate fully-quantized LLM pre-training while leveraging hardware acceleration on NVIDIA Blackwell GPUs.

Abstract: The NVFP4 lower-precision format, supported in hardware by NVIDIA Blackwell GPUs, promises to allow, for the first time, end-to-end fully-quantized pre-training of massive models such as LLMs. Yet, existing quantized training methods still sacrifice some of the representation capacity of this format in favor of more accurate unbiased quantized gradient estimation by stochastic rounding (SR), losing noticeable accuracy relative to standard FP16 and FP8 training. In this paper, improve the state of the art for quantized training in NVFP4 via a novel unbiased quantization routine for micro-scaled formats, called MS-EDEN, that has more than 2x lower quantization error than SR. We integrate it into a novel fully-NVFP4 quantization scheme for linear layers, called Quartet II. We show analytically that Quartet II achieves consistently better gradient estimation across all major matrix multiplications, both on the forward and on the backward passes. In addition, our proposal synergizes well with recent training improvements aimed specifically at NVFP4. We further validate Quartet II on end-to-end LLM training with up to 1.9B parameters on 38B tokens. We provide kernels for execution on NVIDIA Blackwell GPUs with up to 4.2x speedup over BF16. Our code is available at https://github.com/IST-DASLab/Quartet-II .

Method: Proposes a cascaded approach: 1) Generate low-resolution version of tabular rows (categorical features + coarse categorical representation of numerical features), 2) Use this information to guide a high-resolution flow matching model via novel guided conditional probability path and data-dependent coupling. The low-resolution representation explicitly accounts for discrete outcomes in numerical features.

Result: The model generates significantly more realistic samples and captures distributional details more accurately, with detection score increasing by 40%. Formally proves that the cascade tightens the transport cost bound.

Conclusion: The cascaded approach advances diffusion models for tabular data, particularly for mixed-type features, by leveraging low-resolution categorical guidance to improve high-resolution generation fidelity while formally optimizing transport costs.

Abstract: Advances in generative modeling have recently been adapted to tabular data containing discrete and continuous features. However, generating mixed-type features that combine discrete states with an otherwise continuous distribution in a single feature remains challenging. We advance the state-of-the-art in diffusion models for tabular data with a cascaded approach. We first generate a low-resolution version of a tabular data row, that is, the collection of the purely categorical features and a coarse categorical representation of numerical features. Next, this information is leveraged in the high-resolution flow matching model via a novel guided conditional probability path and data-dependent coupling. The low-resolution representation of numerical features explicitly accounts for discrete outcomes, such as missing or inflated values, and therewith enables a more faithful generation of mixed-type features. We formally prove that this cascade tightens the transport cost bound. The results indicate that our model generates significantly more realistic samples and captures distributional details more accurately, for example, the detection score increases by 40%.

Method: Proposes a latent flow-matching framework with equivariance regularization. Uses an equivariance loss to enforce consistency between transformed signals and their reconstructions, fine-tuning pre-trained autoencoder latent spaces with respect to time-series transformations like translation and amplitude scaling.

Result: Equivariance-regularized latent spaces improve generation quality while preserving computational advantages. Outperforms existing diffusion-based baselines on multiple real-world datasets in standard metrics, with orders-of-magnitude faster sampling.

Conclusion: Incorporating geometric inductive biases into latent generative models for time series provides practical benefits in both quality and efficiency, highlighting the value of equivariance-aware latent space design.

Abstract: Flow-based models have proven successful for time-series generation, particularly when defined in lower-dimensional latent spaces that enable efficient sampling. However, how to design latent representations with desirable equivariance properties for time-series generative modelling remains underexplored. In this work, we propose a latent flow-matching framework in which equivariance is explicitly encouraged through a simple regularisation of a pre-trained autoencoder. Specifically, we introduce an equivariance loss that enforces consistency between transformed signals and their reconstructions, and use it to fine-tune latent spaces with respect to basic time-series transformations such as translation and amplitude scaling. We show that these equivariance-regularised latent spaces improve generation quality while preserving the computational advantages of latent flow models. Experiments on multiple real-world datasets demonstrate that our approach consistently outperforms existing diffusion-based baselines in standard time-series generation metrics, while achieving orders-of-magnitude faster sampling. These results highlight the practical benefits of incorporating geometric inductive biases into latent generative models for time series.

Method: Hierarchical Shift Mixing (HSM) distributes pairwise token interactions across Transformer layers rather than computing them densely within each layer. It’s a general framework agnostic to specific mixing functions, enabling linear-time complexity. The approach includes simple HSM variants and hybrid architectures combining HSM with softmax attention.

Result: Simple HSM variants achieve performance close to softmax attention. Hybrid architectures combining HSM with softmax attention outperform GPT-style Transformer baselines while reducing computational cost during both training and inference.

Conclusion: HSM provides an effective framework for efficient token mixing in Transformers, offering linear-time alternatives to quadratic attention that maintain or improve performance while reducing computational costs.

Abstract: Since the introduction of the Transformer architecture for large language models, the softmax-based attention layer has faced increasing scrutinity due to its quadratic-time computational complexity. Attempts have been made to replace it with less complex methods, at the cost of reduced performance in most cases. We introduce Hierarchical Shift Mixing (HSM), a general framework for token mixing that distributes pairwise token interactions across Transformer layers rather than computing them densely within each layer. HSM enables linear-time complexity while remaining agnostic to the specific mixing function. We show that even simple HSM variants achieve performance close to softmax attention, and that hybrid architectures combining HSM with softmax attention can outperform a GPT-style Transformer baseline while reducing computational cost during both training and inference.

Method: Proposes OptiMAG, an Unbalanced Optimal Transport-based regularization framework using Fused Gromov-Wasserstein distance to guide cross-modal structural consistency within local neighborhoods. Includes KL divergence penalty for adaptive handling of cross-modal inconsistencies. Can be integrated as a drop-in regularizer into existing multimodal graph models.

Result: OptiMAG consistently outperforms baselines across multiple tasks including graph-centric tasks (node classification, link prediction) and multimodal-centric generation tasks (graph2text, graph2image).

Conclusion: The framework effectively mitigates structural-semantic conflicts in multimodal graphs and can be seamlessly integrated into existing models to improve performance on both graph analysis and multimodal generation tasks.

Abstract: Multimodal Attributed Graphs (MAGs) have been widely adopted for modeling complex systems by integrating multi-modal information, such as text and images, on nodes. However, we identify a discrepancy between the implicit semantic structure induced by different modality embeddings and the explicit graph structure. For instance, neighbors in the explicit graph structure may be close in one modality but distant in another. Since existing methods typically perform message passing over the fixed explicit graph structure, they inadvertently aggregate dissimilar features, introducing modality-specific noise and impeding effective node representation learning. To address this, we propose OptiMAG, an Unbalanced Optimal Transport-based regularization framework. OptiMAG employs the Fused Gromov-Wasserstein distance to explicitly guide cross-modal structural consistency within local neighborhoods, effectively mitigating structural-semantic conflicts. Moreover, a KL divergence penalty enables adaptive handling of cross-modal inconsistencies. This framework can be seamlessly integrated into existing multimodal graph models, acting as an effective drop-in regularizer. Experiments demonstrate that OptiMAG consistently outperforms baselines across multiple tasks, ranging from graph-centric tasks (e.g., node classification, link prediction) to multimodal-centric generation tasks (e.g., graph2text, graph2image). The source code will be available upon acceptance.

Method: Proposes Matterhorn with two key innovations: 1) M-TTFS encoding that reassigns silent state to most frequent membrane potential and uses ‘dead zone’ strategy to maximize sparsity, reducing spike movement energy; 2) Memristive synapse unit using compute-in-memory technology to eliminate weight access overhead.

Result: On GLUE benchmark, Matterhorn achieves new state-of-the-art, surpassing existing SNNs by 1.42% in average accuracy while delivering 2.31× improvement in energy efficiency.

Conclusion: Matterhorn demonstrates that comprehensive hardware-aware SNN design can significantly improve both accuracy and energy efficiency for LLM inference, addressing critical data movement and weight access bottlenecks.

Abstract: Spiking neural networks (SNNs) have emerged as a promising candidate for energy-efficient LLM inference. However, current energy evaluations for SNNs primarily focus on counting accumulate operations, and fail to account for real-world hardware costs such as data movement, which can consume nearly 80% of the total energy. In this paper, we propose Matterhorn, a spiking transformer that integrates a novel masked time-to-first-spike (M-TTFS) encoding method to reduce spike movement and a memristive synapse unit (MSU) to eliminate weight access overhead. M-TTFS employs a masking strategy that reassigns the zero-energy silent state (a spike train of all 0s) to the most frequent membrane potential rather than the lowest. This aligns the coding scheme with the data distribution, minimizing spike movement energy without information loss. We further propose a `dead zone’ strategy that maximizes sparsity by mapping all values within a given range to the silent state. At the hardware level, the MSU utilizes compute-in-memory (CIM) technology to perform analog integration directly within memory, effectively removing weight access costs. On the GLUE benchmark, Matterhorn establishes a new state-of-the-art, surpassing existing SNNs by 1.42% in average accuracy while delivering a 2.31 times improvement in energy efficiency.

Method: Transform time series into Quantile Graphs (QG) using complex network mappings, then reconstruct synthetic data via inverse mapping while preserving statistical and structural properties.

Result: The quantile graph-based approach produces synthetic data that preserves original properties and offers competitive performance compared to state-of-the-art GAN methods.

Conclusion: The framework provides an interpretable and effective alternative for synthetic time series generation using complex network mappings.

Abstract: Time series data are essential for a wide range of applications, particularly in developing robust machine learning models. However, access to high-quality datasets is often limited due to privacy concerns, acquisition costs, and labeling challenges. Synthetic time series generation has emerged as a promising solution to address these constraints. In this work, we present a framework for generating synthetic time series by leveraging complex networks mappings. Specifically, we investigate whether time series transformed into Quantile Graphs (QG) – and then reconstructed via inverse mapping – can produce synthetic data that preserve the statistical and structural properties of the original. We evaluate the fidelity and utility of the generated data using both simulated and real-world datasets, and compare our approach against state-of-the-art Generative Adversarial Network (GAN) methods. Results indicate that our quantile graph-based methodology offers a competitive and interpretable alternative for synthetic time series generation.

Method: Treats propositions as instances of parameterized predicates rather than discrete symbols, allowing policies to learn shared structure across related propositions. Proposes a novel architecture that embeds and composes predicates to represent LTL specifications.

Result: Demonstrates successful zero-shot generalization to novel propositions and tasks across challenging environments, achieving generalization across both LTL formula structures and proposition vocabularies.

Conclusion: PlatoLTRL enables more flexible generalization in multi-task RL by handling both compositional and parametric generalization through parameterized predicate representations.

Abstract: A central challenge in multi-task reinforcement learning (RL) is to train generalist policies capable of performing tasks not seen during training. To facilitate such generalization, linear temporal logic (LTL) has recently emerged as a powerful formalism for specifying structured, temporally extended tasks to RL agents. While existing approaches to LTL-guided multi-task RL demonstrate successful generalization across LTL specifications, they are unable to generalize to unseen vocabularies of propositions (or “symbols”), which describe high-level events in LTL. We present PlatoLTL, a novel approach that enables policies to zero-shot generalize not only compositionally across LTL formula structures, but also parametrically across propositions. We achieve this by treating propositions as instances of parameterized predicates rather than discrete symbols, allowing policies to learn shared structure across related propositions. We propose a novel architecture that embeds and composes predicates to represent LTL specifications, and demonstrate successful zero-shot generalization to novel propositions and tasks across challenging environments.

Method: Proposes a regularization-based calibration method that enforces multivariate calibration during training using pre-rank functions. Introduces a novel PCA-based pre-rank that projects predictions onto principal directions of the predictive distribution.

Result: The approach substantially improves multivariate pre-rank calibration without compromising predictive accuracy. The PCA pre-rank reveals dependence-structure misspecifications not detected by existing pre-ranks, as shown in simulation studies and experiments on 18 real-world multi-output regression datasets.

Conclusion: The proposed regularization method effectively improves multivariate calibration during training, and the PCA pre-rank provides valuable diagnostics for detecting dependence-structure issues in multivariate probabilistic predictions.

Abstract: The goal of probabilistic prediction is to issue predictive distributions that are as informative as possible, subject to being calibrated. Despite substantial progress in the univariate setting, achieving multivariate calibration remains challenging. Recent work has introduced pre-rank functions, scalar projections of multivariate forecasts and observations, as flexible diagnostics for assessing specific aspects of multivariate calibration, but their use has largely been limited to post-hoc evaluation. We propose a regularization-based calibration method that enforces multivariate calibration during training of multivariate distributional regression models using pre-rank functions. We further introduce a novel PCA-based pre-rank that projects predictions onto principal directions of the predictive distribution. Through simulation studies and experiments on 18 real-world multi-output regression datasets, we show that the proposed approach substantially improves multivariate pre-rank calibration without compromising predictive accuracy, and that the PCA pre-rank reveals dependence-structure misspecifications that are not detected by existing pre-ranks.

Method: Theoretical analysis shows policy smoothness depends on the critic’s differential geometry. PAVE (Policy-Aware Value-field Equalization) treats the critic as a scalar field and stabilizes its induced action-gradient field by minimizing Q-gradient volatility while preserving local curvature.

Result: PAVE achieves smoothness and robustness comparable to policy-side regularization methods while maintaining competitive task performance, without requiring actor network modifications.

Conclusion: Critic-centric regularization via PAVE effectively addresses policy oscillation by smoothing the learning signal at its source, providing a principled alternative to direct policy regularization.

Abstract: Policies learned via continuous actor-critic methods often exhibit erratic, high-frequency oscillations, making them unsuitable for physical deployment. Current approaches attempt to enforce smoothness by directly regularizing the policy’s output. We argue that this approach treats the symptom rather than the cause. In this work, we theoretically establish that policy non-smoothness is fundamentally governed by the differential geometry of the critic. By applying implicit differentiation to the actor-critic objective, we prove that the sensitivity of the optimal policy is bounded by the ratio of the Q-function’s mixed-partial derivative (noise sensitivity) to its action-space curvature (signal distinctness). To empirically validate this theoretical insight, we introduce PAVE (Policy-Aware Value-field Equalization), a critic-centric regularization framework that treats the critic as a scalar field and stabilizes its induced action-gradient field. PAVE rectifies the learning signal by minimizing the Q-gradient volatility while preserving local curvature. Experimental results demonstrate that PAVE achieves smoothness and robustness comparable to policy-side smoothness regularization methods, while maintaining competitive task performance, without modifying the actor.

Method: Extends VSPYCT with a single-tree Bayesian model where each leaf has a GP predictor. Uses Bayesian oblique splits for uncertainty-aware partitioning and GP leaves for local functional modeling. Includes efficient inference combining posterior sampling of splits with GP predictions, and a gating mechanism for extrapolation.

Result: Experiments on benchmark regression tasks show improved predictive performance compared to standard variational oblique trees, with substantial gains in extrapolation scenarios.

Conclusion: The proposed Bayesian tree with GP leaves addresses limitations of standard decision trees in regression, providing better extrapolation capabilities and uncertainty calibration.

Abstract: Decision trees are widely used due to their interpretability and efficiency, but they struggle in regression tasks that require reliable extrapolation and well-calibrated uncertainty. Piecewise-constant leaf predictions are bounded by the training targets and often become overconfident under distribution shift. We propose a single-tree Bayesian model that extends VSPYCT by equipping each leaf with a GP predictor. Bayesian oblique splits provide uncertainty-aware partitioning of the input space, while GP leaves model local functional behaviour and enable principled extrapolation beyond the observed target range. We present an efficient inference and prediction scheme that combines posterior sampling of split parameters with \gls{gp} posterior predictions, and a gating mechanism that activates GP-based extrapolation when inputs fall outside the training support of a leaf. Experiments on benchmark regression tasks show improvements in the predictive performance compared to standard variational oblique trees, and substantial performance gains in extrapolation scenarios.

Method: Projects momentum onto tangent space of model parameters and constrains it on a rotational Oblique manifold, creating a novel optimizer that bridges the performance gap between manifold optimization and modern optimizers.

Result: Extensive experiments on LLaMA and Qwen3 models show Mano consistently and significantly outperforms AdamW and Muon with less memory consumption and computational complexity, expanding the Pareto frontier in space and time efficiency.

Conclusion: Mano demonstrates that manifold optimization can be effectively applied to large-scale LLM training, offering superior performance and efficiency compared to state-of-the-art optimizers.

Abstract: While large language models (LLMs) have emerged as a significant advancement in artificial intelligence, the hardware and computational costs for training LLMs are also significantly burdensome. Among the state-of-the-art optimizers, AdamW relies on diagonal curvature estimates and ignores structural properties, while Muon applies global spectral normalization at the expense of losing curvature information. In this study, we restriked manifold optimization methods for training LLMs, which may address both optimizers’ limitations, while conventional manifold optimization methods have been largely overlooked due to the poor performance in large-scale model optimization. By innovatively projecting the momentum onto the tangent space of model parameters and constraining it on a rotational Oblique manifold, we propose a novel, powerful, and efficient optimizer Mano that is the first to bridge the performance gap between manifold optimization and modern optimizers. Extensive experiments on the LLaMA and Qwen3 models demonstrate that Mano consistently and significantly outperforms AdamW and Muon even with less memory consumption and computational complexity, respectively, suggesting an expanded Pareto frontier in terms of space and time efficiency.

Method: Uses spectral energy entropy to evaluate matrix importance, supports rank pruning and expansion under global budget, and employs zero-impact initialization for newly added singular directions to ensure stability.

Result: Extensive experiments show FlexLoRA consistently outperforms state-of-the-art baselines across benchmarks.

Conclusion: FlexLoRA provides a more principled solution for PEFT by addressing granularity, flexibility, and stability limitations of existing methods.

Abstract: Large pre-trained models achieve remarkable success across diverse domains, yet fully fine-tuning incurs prohibitive computational and memory costs. Parameter-efficient fine-tuning (PEFT) has thus become a mainstream paradigm. Among them, Low-Rank Adaptation (LoRA) introduces trainable low-rank matrices and shows strong performance, nevertheless, its fixed-rank design limits flexibility. Dynamic rank allocation methods mitigate this issue by pruning redundant directions; however, they often rely on heuristic, element-level metrics that globally sort rank directions without matrix-wise distinction, and they lack mechanisms to expand capacity in layers requiring additional adaptation. To overcome these limitations, we propose FlexLoRA, an entropy-guided flexible low-rank adaptation framework that (i) evaluates matrix importance via spectral energy entropy, (ii) supports rank pruning and expansion under a global budget, and (iii) employs zero-impact initialization for newly added singular directions to ensure stability. By addressing granularity, flexibility, and stability limitations, FlexLoRA provides a more principled solution for PEFT. Extensive experiments show that FlexLoRA consistently outperforms state-of-the-art baselines across benchmarks. Codes are available at https://github.com/Chongjie-Si/Subspace-Tuning.

Method: Proposes Continuous Constraint Interpolation (CCI) framework that unifies three constraint families along a common spectrum with a single interpolation parameter, and develops Automatic Constraint Policy Optimization (ACPO) algorithm with Lagrangian dual updates for adaptive constraint adaptation.

Result: Achieves state-of-the-art performance on D4RL and NeoRL2 benchmarks across diverse domains, demonstrating robust gains and effective constraint adaptation.

Conclusion: CCI provides a unified framework for understanding offline RL constraints, enabling principled combinations and smooth transitions between constraint types, with ACPO offering practical adaptive optimization.

Abstract: Offline Reinforcement Learning (RL) relies on policy constraints to mitigate extrapolation error, where both the constraint form and constraint strength critically shape performance. However, most existing methods commit to a single constraint family: weighted behavior cloning, density regularization, or support constraints, without a unified principle that explains their connections or trade-offs. In this work, we propose Continuous Constraint Interpolation (CCI), a unified optimization framework in which these three constraint families arise as special cases along a common constraint spectrum. The CCI framework introduces a single interpolation parameter that enables smooth transitions and principled combinations across constraint types. Building on CCI, we develop Automatic Constraint Policy Optimization (ACPO), a practical primal–dual algorithm that adapts the interpolation parameter via a Lagrangian dual update. Moreover, we establish a maximum-entropy performance difference lemma and derive performance lower bounds for both the closed-form optimal policy and its parametric projection. Experiments on D4RL and NeoRL2 demonstrate robust gains across diverse domains, achieving state-of-the-art performance overall.

Method: Leverages DC structure to smooth non-smooth regularizers using Moreau envelopes, redistributes concave part to data fidelity, and develops DC-LA (proximal Langevin algorithm) with convergence guarantees under distant dissipativity.

Result: Establishes convergence in q-Wasserstein distance for all q ∈ ℕ*, improves upon previous non-log-concave sampling results, and demonstrates effectiveness in synthetic settings and real-world CT uncertainty quantification.

Conclusion: DC-LA provides a theoretically sound and practically effective sampling method for distributions with DC regularizers, enabling reliable uncertainty quantification in inverse problems like computed tomography.

Abstract: We study a sampling problem whose target distribution is $π\propto \exp(-f-r)$ where the data fidelity term $f$ is Lipschitz smooth while the regularizer term $r=r_1-r_2$ is a non-smooth difference-of-convex (DC) function, i.e., $r_1,r_2$ are convex. By leveraging the DC structure of $r$, we can smooth out $r$ by applying Moreau envelopes to $r_1$ and $r_2$ separately. In line of DC programming, we then redistribute the concave part of the regularizer to the data fidelity and study its corresponding proximal Langevin algorithm (termed DC-LA). We establish convergence of DC-LA to the target distribution $π$, up to discretization and smoothing errors, in the $q$-Wasserstein distance for all $q \in \mathbb{N}^*$, under the assumption that $V$ is distant dissipative. Our results improve previous work on non-log-concave sampling in terms of a more general framework and assumptions. Numerical experiments show that DC-LA produces accurate distributions in synthetic settings and reliably provides uncertainty quantification in a real-world Computed Tomography application.

Method: Uses Convolutional Sparse Autoencoder (CSAE) for temporal feature extraction from raw sEMG signals, with few-shot transfer learning for subject adaptation and incremental learning for gesture expansion.

Result: Achieved 94.3% F1-score on 6-class gestures, improved unseen subject performance from 35.1% to 92.3% with few-shot learning, and expanded to 10-class with 90.0% F1-score via incremental learning.

Conclusion: Proposes scalable, efficient framework for affordable prosthetic systems with minimal sensor requirements and computational overhead.

Abstract: Reliable control of myoelectric prostheses is often hindered by high inter-subject variability and the clinical impracticality of high-density sensor arrays. This study proposes a deep learning framework for accurate gesture recognition using only two surface electromyography (sEMG) channels. The method employs a Convolutional Sparse Autoencoder (CSAE) to extract temporal feature representations directly from raw signals, eliminating the need for heuristic feature engineering. On a 6-class gesture set, our model achieved a multi-subject F1-score of 94.3% $\pm$ 0.3%. To address subject-specific differences, we present a few-shot transfer learning protocol that improved performance on unseen subjects from a baseline of 35.1% $\pm$ 3.1% to 92.3% $\pm$ 0.9% with minimal calibration data. Furthermore, the system supports functional extensibility through an incremental learning strategy, allowing for expansion to a 10-class set with a 90.0% $\pm$ 0.2% F1-score without full model retraining. By combining high precision with minimal computational and sensor overhead, this framework provides a scalable and efficient approach for the next generation of affordable and adaptive prosthetic systems.

Method: Proposes Scalable Topology-Preserving Graph Coarsening (STPGC) using concepts from algebraic topology: graph strong collapse and graph edge collapse. Includes three algorithms: GStrongCollapse, GEdgeCollapse, and NeighborhoodConing, which eliminate dominated nodes/edges while preserving topological features. Also develops approximate algorithms to accelerate GNN training and proves STPGC preserves GNN receptive field.

Result: Experiments on node classification with GNNs demonstrate STPGC’s efficiency and effectiveness. The method achieves scalable graph coarsening while maintaining topological features that are important for GNN performance.

Conclusion: STPGC provides a scalable solution for topology-preserving graph coarsening that maintains GNN performance while addressing the exponential time complexity of previous methods, making it practical for large-scale graph applications.

Abstract: Graph coarsening reduces the size of a graph while preserving certain properties. Most existing methods preserve either spectral or spatial characteristics. Recent research has shown that preserving topological features helps maintain the predictive performance of graph neural networks (GNNs) trained on the coarsened graph but suffers from exponential time complexity. To address these problems, we propose Scalable Topology-Preserving Graph Coarsening (STPGC) by introducing the concepts of graph strong collapse and graph edge collapse extended from algebraic topology. STPGC comprises three new algorithms, GStrongCollapse, GEdgeCollapse, and NeighborhoodConing based on these two concepts, which eliminate dominated nodes and edges while rigorously preserving topological features. We further prove that STPGC preserves the GNN receptive field and develop approximate algorithms to accelerate GNN training. Experiments on node classification with GNNs demonstrate the efficiency and effectiveness of STPGC.

Method: The authors analyze the non-normal dynamics of the Sinkhorn fixed-point map to reveal a theoretical thermodynamic speed limit. They propose Efficient PH-ASC, an adaptive scheduling algorithm that monitors inference process stability by enforcing a linear stability law, decoupling expensive spectral diagnostics from training to reduce overhead from O(N³) to amortized O(1).

Result: The proposed method provides a stable approach to annealing in differentiable matching layers by preventing premature mode collapse. The implementation and interactive demo are made available, offering practical tools for structural prediction tasks using entropy-regularized Optimal Transport.

Conclusion: The paper provides theoretical insight into the instability of annealing in differentiable matching layers and offers a practical solution through adaptive scheduling that maintains stability while reducing computational overhead, making entropy-regularized Optimal Transport more reliable for structural prediction.

Abstract: Differentiable matching layers, often implemented via entropy-regularized Optimal Transport, serve as a critical approximate inference mechanism in structural prediction. However, recovering discrete permutations via annealing $ε\to 0$ is notoriously unstable. We identify a fundamental mechanism for this failure: \textbf{Premature Mode Collapse}. By analyzing the non-normal dynamics of the Sinkhorn fixed-point map, we reveal a theoretical \textbf{thermodynamic speed limit}. Under standard exponential cooling, the shift in the target posterior ($O(1)$) outpaces the contraction rate of the inference operator, which degrades as $O(1/ε)$. This mismatch inevitably forces the inference trajectory into spurious local basins. To address this, we propose \textbf{Efficient PH-ASC}, an adaptive scheduling algorithm that monitors the stability of the inference process. By enforcing a linear stability law, we decouple expensive spectral diagnostics from the training loop, reducing overhead from $O(N^3)$ to amortized $O(1)$. Our implementation and interactive demo are available at https://github.com/xxx0438/torch-sinkhorn-asc and https://huggingface.co/spaces/leon0923/torch-sinkhorn-asc-demo. bounded away from zero in generic training dynamics unless the feature extractor converges unrealistically fast.

Method: Proposes Environment-Conditioned Tail Reweighting for Total Variation Invariant Risk Minimization (ECTR), which augments TV-based invariant learning with environment-conditioned tail reweighting to jointly address both types of distribution shift. Also extends to scenarios without explicit environment annotations through latent environment inference via minimax formulation.

Result: Experiments across regression, tabular, time-series, and image classification benchmarks under mixed distribution shifts demonstrate consistent improvements in both worst-environment and average OOD performance.

Conclusion: ECTR provides a unified framework that makes environment-level invariance and within-environment robustness complementary under mixed distribution shifts, effectively addressing both correlation and diversity shifts for better OOD generalization.

Abstract: Out-of-distribution (OOD) generalization remains challenging when models simultaneously encounter correlation shifts across environments and diversity shifts driven by rare or hard samples. Existing invariant risk minimization (IRM) methods primarily address spurious correlations at the environment level, but often overlook sample-level heterogeneity within environments, which can critically impact OOD performance. In this work, we propose \emph{Environment-Conditioned Tail Reweighting for Total Variation Invariant Risk Minimization} (ECTR), a unified framework that augments TV-based invariant learning with environment-conditioned tail reweighting to jointly address both types of distribution shift. By integrating environment-level invariance with within-environment robustness, the proposed approach makes these two mechanisms complementary under mixed distribution shifts. We further extend the framework to scenarios without explicit environment annotations by inferring latent environments through a minimax formulation. Experiments across regression, tabular, time-series, and image classification benchmarks under mixed distribution shifts demonstrate consistent improvements in both worst-environment and average OOD performance.

Method: Proposes a density-aware conditional graph generation framework using Wasserstein GANs (WGAN) that replaces random sampling with a learnable distance-based edge predictor. Nodes are embedded into a latent space where proximity correlates with edge likelihood. A differentiable edge predictor determines pairwise relationships from node embeddings, while a density-aware selection mechanism adaptively controls edge density to match class-specific sparsity distributions.

Result: Experiments on benchmark datasets show the method produces graphs with superior structural coherence and class-consistent connectivity compared to existing baselines. The learned edge predictor captures complex relational patterns beyond simple heuristics, generating graphs whose density and topology closely match real structural distributions.

Conclusion: The framework demonstrates improved training stability and controllable synthesis, making it effective for realistic graph generation and data augmentation. The approach successfully captures complex structural dependencies in graph data.

Abstract: Generating realistic graph-structured data is challenging due to discrete structures, variable sizes, and class-specific connectivity patterns that resist conventional generative modelling. While recent graph generation methods employ generative adversarial network (GAN) frameworks to handle permutation invariance and irregular topologies, they typically rely on random edge sampling with fixed probabilities, limiting their capacity to capture complex structural dependencies between nodes. We propose a density-aware conditional graph generation framework using Wasserstein GANs (WGAN) that replaces random sampling with a learnable distance-based edge predictor. Our approach embeds nodes into a latent space where proximity correlates with edge likelihood, enabling the generator to learn meaningful connectivity patterns. A differentiable edge predictor determines pairwise relationships directly from node embeddings, while a density-aware selection mechanism adaptively controls edge density to match class-specific sparsity distributions observed in real graphs. We train the model using a WGAN with gradient penalty, employing a GCN-based critic to ensure generated graphs exhibit realistic topology and align with target class distributions. Experiments on benchmark datasets demonstrate that our method produces graphs with superior structural coherence and class-consistent connectivity compared to existing baselines. The learned edge predictor captures complex relational patterns beyond simple heuristics, generating graphs whose density and topology closely match real structural distributions. Our results show improved training stability and controllable synthesis, making the framework effective for realistic graph generation and data augmentation. Source code is publicly available at https://github.com/ava-12/Density_Aware_WGAN.git.

Method: Introduces new analytical tools for Nash regret analysis and proposes FairLinBandit, a generic algorithmic framework that works as a meta-algorithm on top of any linear bandit strategy. Instantiates the framework using Phased Elimination and Upper Confidence Bound algorithms.

Result: Achieves order-optimal Nash regret bound in linear bandits and proves that both instantiations achieve sublinear p-means regret for the entire range of p. Extensive experiments on real-world datasets show consistent outperformance over state-of-the-art baselines.

Conclusion: The paper resolves the open problem of suboptimal Nash regret bounds in linear bandits and introduces a comprehensive framework for studying fairness-utility trade-offs through p-means regret, with practical algorithms that outperform existing methods.

Abstract: Nash regret has recently emerged as a principled fairness-aware performance metric for stochastic multi-armed bandits, motivated by the Nash Social Welfare objective. Although this notion has been extended to linear bandits, existing results suffer from suboptimality in ambient dimension $d$, stemming from proof techniques that rely on restrictive concentration inequalities. In this work, we resolve this open problem by introducing new analytical tools that yield an order-optimal Nash regret bound in linear bandits. Beyond Nash regret, we initiate the study of $p$-means regret in linear bandits, a unifying framework that interpolates between fairness and utility objectives and strictly generalizes Nash regret. We propose a generic algorithmic framework, FairLinBandit, that works as a meta-algorithm on top of any linear bandit strategy. We instantiate this framework using two bandit algorithms: Phased Elimination and Upper Confidence Bound, and prove that both achieve sublinear $p$-means regret for the entire range of $p$. Extensive experiments on linear bandit instances generated from real-world datasets demonstrate that our methods consistently outperform the existing state-of-the-art baseline.

Method: dgMARK steers the unmasking order toward positions where high-reward candidate tokens satisfy a simple parity constraint induced by a binary hash. It works plug-and-play with common decoding strategies (confidence, entropy, margin-based ordering) and has a one-step lookahead variant. Detection uses elevated parity-matching statistics with a sliding-window detector for robustness.

Result: The method provides effective watermarking for dLLMs that is robust to post-editing operations including insertion, deletion, substitution, and paraphrasing.

Conclusion: dgMARK demonstrates that the unmasking order sensitivity in practical dLLMs can be leveraged for watermarking without explicit probability reweighting, offering a new approach for discrete diffusion models.

Abstract: We propose dgMARK, a decoding-guided watermarking method for discrete diffusion language models (dLLMs). Unlike autoregressive models, dLLMs can generate tokens in arbitrary order. While an ideal conditional predictor would be invariant to this order, practical dLLMs exhibit strong sensitivity to the unmasking order, creating a new channel for watermarking. dgMARK steers the unmasking order toward positions whose high-reward candidate tokens satisfy a simple parity constraint induced by a binary hash, without explicitly reweighting the model’s learned probabilities. The method is plug-and-play with common decoding strategies (e.g., confidence, entropy, and margin-based ordering) and can be strengthened with a one-step lookahead variant. Watermarks are detected via elevated parity-matching statistics, and a sliding-window detector ensures robustness under post-editing operations including insertion, deletion, substitution, and paraphrasing.

Method: Train a tabular foundation model (ExplainerPFN) on synthetic datasets from random structural causal models, supervised using exact/near-exact Shapley values. Once trained, it predicts feature attributions without model access, gradients, or example explanations.

Result: ExplainerPFN achieves performance competitive with few-shot surrogate explainers using 2-10 SHAP examples, showing high fidelity to SHAP values with as few as two reference observations.

Conclusion: Zero-shot Shapley value estimation is feasible without model access, enabling interpretability in real-world deployments where direct model access is restricted.

Abstract: Computing the importance of features in supervised classification tasks is critical for model interpretability. Shapley values are a widely used approach for explaining model predictions, but require direct access to the underlying model, an assumption frequently violated in real-world deployments. Further, even when model access is possible, their exact computation may be prohibitively expensive. We investigate whether meaningful Shapley value estimations can be obtained in a zero-shot setting, using only the input data distribution and no evaluations of the target model. To this end, we introduce ExplainerPFN, a tabular foundation model built on TabPFN that is pretrained on synthetic datasets generated from random structural causal models and supervised using exact or near-exact Shapley values. Once trained, ExplainerPFN predicts feature attributions for unseen tabular datasets without model access, gradients, or example explanations. Our contributions are fourfold: (1) we show that few-shot learning-based explanations can achieve high fidelity to SHAP values with as few as two reference observations; (2) we propose ExplainerPFN, the first zero-shot method for estimating Shapley values without access to the underlying model or reference explanations; (3) we provide an open-source implementation of ExplainerPFN, including the full training pipeline and synthetic data generator; and (4) through extensive experiments on real and synthetic datasets, we show that ExplainerPFN achieves performance competitive with few-shot surrogate explainers that rely on 2-10 SHAP examples.

Method: VaR-CPO uses the one-sided Chebyshev inequality to create a tractable surrogate for the non-differentiable VaR constraint based on the first two moments of cost return. It extends the trust-region framework of Constrained Policy Optimization (CPO) to provide rigorous worst-case bounds.

Result: Empirical results show VaR-CPO achieves zero constraint violations during training in feasible environments, a property that baseline methods fail to uphold. The method is sample efficient and conservative.

Conclusion: VaR-CPO provides a practical solution for safe reinforcement learning with rigorous theoretical guarantees for both policy improvement and constraint violation bounds during training.

Abstract: We introduce the Value-at-Risk Constrained Policy Optimization algorithm (VaR-CPO), a sample efficient and conservative method designed to optimize Value-at-Risk (VaR) constraints directly. Empirically, we demonstrate that VaR-CPO is capable of safe exploration, achieving zero constraint violations during training in feasible environments, a critical property that baseline methods fail to uphold. To overcome the inherent non-differentiability of the VaR constraint, we employ the one-sided Chebyshev inequality to obtain a tractable surrogate based on the first two moments of the cost return. Additionally, by extending the trust-region framework of the Constrained Policy Optimization (CPO) method, we provide rigorous worst-case bounds for both policy improvement and constraint violation during the training process.

Method: Proposes a causal model that explicitly captures both anomaly types by treating outliers as latent interventions on latent (“true”) and observed (“measured”) variables. Shows these are identifiable and develops a maximum likelihood estimation approach for practical implementation.

Result: The method matches state-of-the-art performance in root cause localization while additionally enabling accurate classification of anomaly types. It remains robust even when the causal DAG (Directed Acyclic Graph) is unknown.

Conclusion: The proposed causal modeling approach successfully distinguishes between measurement errors and mechanism shifts in anomaly detection, providing both root cause localization and anomaly type classification with robustness to unknown causal structures.

Abstract: Root cause analysis of anomalies aims to identify those features that cause the deviation from the normal process. Existing methods ignore, however, that anomalies can arise through two fundamentally different processes: measurement errors, where data was generated normally but one or more values were recorded incorrectly, and mechanism shifts, where the causal process generating the data changed. While measurement errors can often be safely corrected, mechanistic anomalies require careful consideration. We define a causal model that explicitly captures both types by treating outliers as latent interventions on latent (“true”) and observed (“measured”) variables. We show that they are identifiable, and propose a maximum likelihood estimation approach to put this to practice. Experiments show that our method matches state-of-the-art performance in root cause localization, while it additionally enables accurate classification of anomaly types, and remains robust even when the causal DAG is unknown.

Method: Train a Divide-and-Conquer CoT model where the LLM acts as a director identifying parallelizable subtasks, spawns workers to execute them, using SFT initialization followed by multi-stage RL with data filtering to recover accuracy.

Result: DC-CoT achieves similar accuracy to DeepScaleR-1.5B-Preview on benchmarks like AIME 2024 and HMMT 2025 while reducing longest path length by 35-40%.

Conclusion: The DC-CoT approach successfully reduces LLM reasoning latency through parallel execution while maintaining accuracy, offering a practical solution for efficient mathematical reasoning.

Abstract: Long chain-of-thought reasoning (Long CoT) is now fundamental to state-of-the-art LLMs, especially in mathematical reasoning. However, LLM generation is highly sequential, and long CoTs lead to a high latency. We propose to train Divide-and-Conquer CoT (DC-CoT) to reduce the latency. With DC-CoT, the model can act as a director that identifies distinct subtasks that can be performed in parallel in its reasoning process, and then spawns workers to execute the subtasks. Our goal is to achieve high accuracy, with a low longest path length, which is a theoretical measure of the latency needed for the response. We start with a long CoT base model (DeepScaleR-1.5B-Preview), and first use SFT with a small curated demonstration set to initialize its ability to spawn workers in a certain format. Because SFT degrades the accuracy significantly, we design a multi-stage RL algorithm, with various data filtering strategies, to recover the accuracy while decreasing the longest path length. Across several benchmarks including AIME 2024 and HMMT 2025, DC-CoT achieves similar accuracy as DeepScaleR-1.5B-Preview while decreasing longest path length by 35-40%. Our code, SFT dataset and models are publicly available at https://github.com/amahankali10/DC_CoT_RL_for_Low_Latency_CoT_with_Parallel_Reasoning.

Method: Proposes Reinforcement Learning with Relative Rewards (RLRR) framework that shifts from absolute scoring to relative ranking. Introduces Ranking Reward Model, a listwise preference model tailored for group-based optimization to directly generate relative rankings.

Result: RLRR yields consistent performance improvements over standard group-based baselines across reasoning benchmarks and open-ended generation tasks by mitigating signal sparsity and reward instability.

Conclusion: Relative ranking-based reward shaping effectively addresses limitations of absolute scoring in group-based RL for LLMs, providing more robust optimization signals for both reasoning and open-ended generation tasks.

Abstract: Reinforcement learning has become a cornerstone for enhancing the reasoning capabilities of Large Language Models, where group-based approaches such as GRPO have emerged as efficient paradigms that optimize policies by leveraging intra-group performance differences. However, these methods typically rely on absolute numerical rewards, introducing intrinsic limitations. In verifiable tasks, identical group evaluations often result in sparse supervision, while in open-ended scenarios, the score range instability of reward models undermines advantage estimation based on group means. To address these limitations, we propose Reinforcement Learning with Relative Rewards (RLRR), a framework that shifts reward shaping from absolute scoring to relative ranking. Complementing this framework, we introduce the Ranking Reward Model, a listwise preference model tailored for group-based optimization to directly generate relative rankings. By transforming raw evaluations into robust relative signals, RLRR effectively mitigates signal sparsity and reward instability. Experimental results demonstrate that RLRR yields consistent performance improvements over standard group-based baselines across reasoning benchmarks and open-ended generation tasks.

Method: Proposes Evolutionary Forecasting (EF) as a unified generative framework where DF is a special case, enabling models trained on short horizons to outperform those trained directly on long horizons through better gradient optimization

Result: A single EF model surpasses task-specific DF ensembles across standard benchmarks and shows robust asymptotic stability in extreme extrapolation scenarios

Conclusion: EF represents a paradigm shift from passive Static Mapping to autonomous Evolutionary Reasoning in Long-term Time Series Forecasting, offering computational efficiency and superior performance

Abstract: The prevailing Direct Forecasting (DF) paradigm dominates Long-term Time Series Forecasting (LTSF) by forcing models to predict the entire future horizon in a single forward pass. While efficient, this rigid coupling of output and evaluation horizons necessitates computationally prohibitive re-training for every target horizon. In this work, we uncover a counter-intuitive optimization anomaly: models trained on short horizons-when coupled with our proposed Evolutionary Forecasting (EF) paradigm-significantly outperform those trained directly on long horizons. We attribute this success to the mitigation of a fundamental optimization pathology inherent in DF, where conflicting gradients from distant futures cripple the learning of local dynamics. We establish EF as a unified generative framework, proving that DF is merely a degenerate special case of EF. Extensive experiments demonstrate that a singular EF model surpasses task-specific DF ensembles across standard benchmarks and exhibits robust asymptotic stability in extreme extrapolation. This work propels a paradigm shift in LTSF: moving from passive Static Mapping to autonomous Evolutionary Reasoning.

Method: Uses B-spline interpolation to construct conditional paths across observations, exploiting B-spline smoothness and stability to learn complex dynamics while meeting multi-marginal requirements. Jointly models paths via spline-based flow matching.

Result: Strong improvements over existing baselines across various deterministic/stochastic dynamical systems of varying complexity and cellular trajectory inference tasks. Demonstrates superior performance in modeling complex dynamics.

Conclusion: SplineFlow provides theoretically grounded flow matching for dynamical systems using B-spline interpolation, effectively capturing higher-order dynamics from irregular observations while maintaining stability and meeting multi-marginal constraints.

Abstract: Flow matching is a scalable generative framework for characterizing continuous normalizing flows with wide-range applications. However, current state-of-the-art methods are not well-suited for modeling dynamical systems, as they construct conditional paths using linear interpolants that may not capture the underlying state evolution, especially when learning higher-order dynamics from irregular sampled observations. Constructing unified paths that satisfy multi-marginal constraints across observations is challenging, since naïve higher-order polynomials tend to be unstable and oscillatory. We introduce SplineFlow, a theoretically grounded flow matching algorithm that jointly models conditional paths across observations via B-spline interpolation. Specifically, SplineFlow exploits the smoothness and stability of B-spline bases to learn the complex underlying dynamics in a structured manner while ensuring the multi-marginal requirements are met. Comprehensive experiments across various deterministic and stochastic dynamical systems of varying complexity, as well as on cellular trajectory inference tasks, demonstrate the strong improvement of SplineFlow over existing baselines. Our code is available at: https://github.com/santanurathod/SplineFlow.

Method: Proposed a taxonomy of regularization methods into four categories: data-based strategies, architecture strategies, training strategies, and loss function strategies. Conducted empirical comparison of various regularization techniques on classification tasks using ten numerical and image datasets applied to MLP and CNN architectures.

Result: Regularization effectiveness is dataset-dependent: regularization terms improved performance only on numeric datasets, while batch normalization improved performance only on image datasets. No universal regularization benefit was found across all datasets.

Conclusion: Generalization is crucial for ML, but regularization effects are context-dependent. Understanding connections between techniques and their dataset-specific impacts is essential for appropriate practical application.

Abstract: Despite huge successes on a wide range of tasks, neural networks are known to sometimes struggle to generalise to unseen data. Many approaches have been proposed over the years to promote the generalisation ability of neural networks, collectively known as regularisation techniques. These are used as common practice under the assumption that any regularisation added to the pipeline would result in a performance improvement. In this study, we investigate whether this assumption holds in practice. First, we provide a broad review of regularisation techniques, including modern theories such as double descent. We propose a taxonomy of methods under four broad categories, namely: (1) data-based strategies, (2) architecture strategies, (3) training strategies, and (4) loss function strategies. Notably, we highlight the contradictions and correspondences between the approaches in these broad classes. Further, we perform an empirical comparison of the various regularisation techniques on classification tasks for ten numerical and image datasets applied to the multi-layer perceptron and convolutional neural network architectures. Results show that the efficacy of regularisation is dataset-dependent. For example, the use of a regularisation term only improved performance on numeric datasets, whereas batch normalisation improved performance on image datasets only. Generalisation is crucial to machine learning; thus, understanding the effects of applying regularisation techniques, and considering the connections between them is essential to the appropriate use of these methods in practice.

Method: Proposes discretized categorical actors that represent each action dimension with a distribution over bins, creating a policy objective resembling cross-entropy loss. Builds on supervised learning advances to add regularized actor networks while keeping critic design unchanged.

Result: Simply replacing standard actor networks with discretized regularized actors yields consistent performance gains and achieves state-of-the-art results across diverse continuous-control benchmarks.

Conclusion: Policy representation is a crucial design choice for on-policy optimization, and discretized categorical actors with regularization provide a superior alternative to standard Gaussian actors for continuous control tasks.

Abstract: On-policy deep reinforcement learning remains a dominant paradigm for continuous control, yet standard implementations rely on Gaussian actors and relatively shallow MLP policies, often leading to brittle optimization when gradients are noisy and policy updates must be conservative. In this paper, we revisit policy representation as a first-class design choice for on-policy optimization. We study discretized categorical actors that represent each action dimension with a distribution over bins, yielding a policy objective that resembles a cross-entropy loss. Building on architectural advances from supervised learning, we further propose regularized actor networks, while keeping critic design fixed. Our results show that simply replacing the standard actor network with our discretized regularized actor yields consistent gains and achieve the state-of-the-art performance across diverse continuous-control benchmarks.

Method: Introduces Calibration Aware Token-level Training Objective (CATTO) that aligns predicted confidence with empirical prediction correctness, combinable with preference optimization objectives. Also introduces Confidence@k for test-time scaling using calibrated token probabilities.

Result: CATTO reduces Expected Calibration Error by 2.22%-7.61% in-distribution and 1.46%-10.44% out-of-distribution vs DPO, and maintains/improves multiple-choice QA accuracy on five datasets.

Conclusion: CATTO effectively improves LLM calibration without sacrificing task accuracy, providing better confidence estimation for token-level predictions.

Abstract: Large language models (LLMs) often make accurate next token predictions but their confidence in these predictions can be poorly calibrated: high-confidence predictions are frequently wrong, and low-confidence predictions may be correct. This miscalibration is exacerbated by preference-based alignment methods breaking the link between predictive probability and correctness. We introduce a Calibration Aware Token-level Training Objective (CATTO), a calibration-aware objective that aligns predicted confidence with empirical prediction correctness, which can be combined with the original preference optimization objectives. Empirically, CATTO reduces Expected Calibration Error (ECE) by 2.22%-7.61% in-distribution and 1.46%-10.44% out-of-distribution compared to direct preference optimization (DPO), and by 0.22%-1.24% in-distribution and 1.23%-5.07% out-of-distribution compared to the strongest DPO baseline. This improvement in confidence does not come at a cost of losing task accuracy, where CATTO maintains or slightly improves multiple-choice question-answering accuracy on five datasets. We also introduce Confidence@k, a test-time scaling mechanism leveraging calibrated token probabilities for Bayes-optimal selection of output tokens.

Method: DCR derives the exact distribution of non-conformity scores by modeling absolute ranks of calibration items as following Negative Hypergeometric distributions conditional on their relative ranks. This allows for more precise determination of conformal thresholds compared to using conservative bounds.

Result: Extensive experiments show DCR reduces average prediction set size by up to 36% compared to baseline methods while maintaining valid coverage guarantees. The method achieves improved efficiency without sacrificing theoretical coverage properties.

Conclusion: DCR provides a more efficient conformal prediction framework for ranking models by leveraging exact score distributions, offering practical improvements for uncertainty quantification in real-world ranking applications while maintaining theoretical guarantees.

Abstract: Quantifying uncertainty is critical for the safe deployment of ranking models in real-world applications. Recent work offers a rigorous solution using conformal prediction in a full ranking scenario, which aims to construct prediction sets for the absolute ranks of test items based on the relative ranks of calibration items. However, relying on upper bounds of non-conformity scores renders the method overly conservative, resulting in substantially large prediction sets. To address this, we propose Distribution-informed Conformal Ranking (DCR), which produces efficient prediction sets by deriving the exact distribution of non-conformity scores. In particular, we find that the absolute ranks of calibration items follow Negative Hypergeometric distributions, conditional on their relative ranks. DCR thus uses the rank distribution to derive non-conformity score distribution and determine conformal thresholds. We provide theoretical guarantees that DCR achieves improved efficiency over the baseline while ensuring valid coverage under mild assumptions. Extensive experiments demonstrate the superiority of DCR, reducing average prediction set size by up to 36%, while maintaining valid coverage.

Method: STGAT combines drift-aware temporal embeddings and temporal self-attention to capture corrupted time evolution at individual devices, and uses graph attention to model spatial propagation of timing errors. A curvature-regularized latent representation geometrically separates normal clock evolution from anomalies.

Result: STGAT achieves 95.7% accuracy on energy IoT telemetry with controlled timing perturbations, outperforming recurrent, transformer, and graph-based baselines with significant improvements (d > 1.8, p < 0.001). It reduces detection delay by 26%, achieving a 2.3-time-step delay while maintaining stable performance under various timing anomalies.

Conclusion: STGAT effectively detects temporal anomalies in distributed IoT energy systems by modeling both temporal distortion and inter-device consistency, addressing critical vulnerabilities in time-sensitive cyber-physical systems.

Abstract: The integrity of time in distributed Internet of Things (IoT) devices is crucial for reliable operation in energy cyber-physical systems, such as smart grids and microgrids. However, IoT systems are vulnerable to clock drift, time-synchronization manipulation, and timestamp discontinuities, such as the Year 2038 (Y2K38) Unix overflow, all of which disrupt temporal ordering. Conventional anomaly-detection models, which assume reliable timestamps, fail to capture temporal inconsistencies. This paper introduces STGAT (Spatio-Temporal Graph Attention Network), a framework that models both temporal distortion and inter-device consistency in energy IoT systems. STGAT combines drift-aware temporal embeddings and temporal self-attention to capture corrupted time evolution at individual devices, and uses graph attention to model spatial propagation of timing errors. A curvature-regularized latent representation geometrically separates normal clock evolution from anomalies caused by drift, synchronization offsets, and overflow events. Experimental results on energy IoT telemetry with controlled timing perturbations show that STGAT achieves 95.7% accuracy, outperforming recurrent, transformer, and graph-based baselines with significant improvements (d > 1.8, p < 0.001). Additionally, STGAT reduces detection delay by 26%, achieving a 2.3-time-step delay while maintaining stable performance under overflow, drift, and physical inconsistencies.

Method: Theoretical analysis through the lens of local curvature of sequence-level policy gradient, showing std normalization implements adaptive gradient scaling. Empirical analysis on GSM8K and MATH benchmarks to identify training phases.

Result: GRPO enjoys strictly improved convergence rate over unnormalized REINFORCE, with gains characterized by average within-prompt reward standard deviation. Three training phases identified: early acceleration (high variance/orthogonality), stable transition, and late-stage where orthogonality loss limits gains.

Conclusion: Provides principled understanding of when std normalization helps in GRPO, offering broader insights for critic-free RL algorithm design in language model reasoning.

Abstract: Reinforcement learning (RL) has become a key driver of language model reasoning. Among RL algorithms, Group Relative Policy Optimization (GRPO) is the de facto standard, avoiding the need for a critic by using per-prompt baselines and variance normalization. Yet why and when this normalization helps remains unclear. In this work, we provide an explanation through the lens of local curvature of the sequence-level policy gradient: standard deviation normalization implements an adaptive gradient. Theoretically, under mild conditions, GRPO enjoys a strictly improved convergence rate over unnormalized REINFORCE, with gains characterized by the average within-prompt reward standard deviation across prompts and iterations. Empirically, our analysis on GSM8K and MATH benchmarks reveals three distinct training phases governed by the interplay between feature orthogonality and reward variance: (I) an early acceleration phase where high variance and orthogonality favor adaptive scaling; (II) a relatively stable transition phase; and (III) a late-stage regime where the loss of orthogonality limits further gains. Together, these results provide a principled account of when std normalization helps in GRPO, and offer broader insights into the design of critic-free RL algorithms.

Method: Implemented a deep reinforcement learning framework for hourly medication dosing under partial observability using data from 47,144 ICU stays in MIMIC-IV database. Trained policies to prescribe opioids, propofol, benzodiazepines, and dexmedetomidine with two different objectives: (1) reduce pain only, and (2) jointly reduce pain and mortality.

Result: Both policies were associated with lower pain, but actions from the pain-only policy were positively correlated with mortality, while actions from the joint pain-mortality reduction policy were negatively correlated with mortality. This demonstrates that including long-term outcomes in the objective function leads to safer treatment policies.

Conclusion: Valuing long-term outcomes (like mortality) is critical for developing safer treatment policies in ICU pain management, even when short-term goals (pain reduction) remain the primary objective. This highlights the importance of appropriate reward design in reinforcement learning for healthcare applications.

Abstract: Pain management in intensive care usually involves complex trade-offs between therapeutic goals and patient safety, since both inadequate and excessive treatment may induce serious sequelae. Reinforcement learning can help address this challenge by learning medication dosing policies from retrospective data. However, prior work on sedation and analgesia has optimized for objectives that do not value patient survival while relying on algorithms unsuitable for imperfect information settings. We investigated the risks of these design choices by implementing a deep reinforcement learning framework to suggest hourly medication doses under partial observability. Using data from 47,144 ICU stays in the MIMIC-IV database, we trained policies to prescribe opioids, propofol, benzodiazepines, and dexmedetomidine according to two goals: reduce pain or jointly reduce pain and mortality. We found that, although the two policies were associated with lower pain, actions from the first policy were positively correlated with mortality, while those proposed by the second policy were negatively correlated. This suggests that valuing long-term outcomes could be critical for safer treatment policies, even if a short-term goal remains the primary objective.

Method: Develops a computationally cheap, mathematically justified distributional modification that perturbs data distributions in a constraint-aware way. The perturbation creates a new distribution with support matching ambient space dimension while implicitly incorporating underlying manifold geometry.

Result: The approach consistently enables data distribution recovery and stable sampling for both diffusion models and normalizing flows across several representative tasks, as demonstrated through theoretical analyses and empirical evidence.

Conclusion: The proposed constraint-aware perturbation method effectively addresses fundamental limitations of generative models in equality-constrained settings, providing a flexible and mathematically sound solution for scientific applications.

Abstract: Generative models have enjoyed widespread success in a variety of applications. However, they encounter inherent mathematical limitations in modeling distributions where samples are constrained by equalities, as is frequently the setting in scientific domains. In this work, we develop a computationally cheap, mathematically justified, and highly flexible distributional modification for combating known pitfalls in equality-constrained generative models. We propose perturbing the data distribution in a constraint-aware way such that the new distribution has support matching the ambient space dimension while still implicitly incorporating underlying manifold geometry. Through theoretical analyses and empirical evidence on several representative tasks, we illustrate that our approach consistently enables data distribution recovery and stable sampling with both diffusion models and normalizing flows.

Method: Proposes SPICE (Submodular Penalized Information Conflict-aware Selection) that formalizes gradient conflicts via ε-decomposition, quantifying deviation from ideal submodularity. The method maximizes Fisher information while penalizing misalignment between per-sample gradients, supports early stopping and proxy models for efficiency.

Result: SPICE selects subsets with higher log-determinant information than original criteria. Across 8 benchmarks with LLaMA2-7B and Qwen2-7B, using only 10% of data, SPICE matches or exceeds 6 methods including full-data tuning, achieving performance improvements with substantially lower training cost.

Conclusion: Addressing gradient conflicts is crucial for effective data selection in instruction tuning. SPICE provides a practical, efficient solution that achieves strong performance with minimal data, making large-scale instruction tuning more accessible.

Abstract: Information-based data selection for instruction tuning is compelling: maximizing the log-determinant of the Fisher information yields a monotone submodular objective, enabling greedy algorithms to achieve a $(1-1/e)$ approximation under a cardinality budget. In practice, however, we identify alleviating gradient conflicts, misalignment between per-sample gradients, is a key factor that slows down the decay of marginal log-determinant information gains, thereby preventing significant loss of information. We formalize this via an $\varepsilon$-decomposition that quantifies the deviation from ideal submodularity as a function of conflict statistics, yielding data-dependent approximation factors that tighten as conflicts diminish. Guided by this analysis, we propose SPICE, a conflict-aware selector that maximizes information while penalizing misalignment, and that supports early stopping and proxy models for efficiency. Empirically, SPICE selects subsets with higher log-determinant information than original criteria, and these informational gains translate into performance improvements: across 8 benchmarks with LLaMA2-7B and Qwen2-7B, SPICE uses only 10% of the data, yet matches or exceeds 6 methods including full-data tuning. This achieves performance improvements with substantially lower training cost.

Method: Proposes Behemoth, a fully synthetic data generation framework that allows controlled study of model editing. Demonstrates the framework using simple tabular data to explore editing techniques and their effectiveness.

Result: The framework reveals surprising findings about model editing, including that restricting update rank can sometimes result in more effective updates, echoing real-world observations.

Conclusion: Synthetic frameworks like Behemoth provide valuable insights into model editing that are difficult to obtain from real-world LLMs, helping understand the relationship between training data and editing effectiveness.

Abstract: As artificial neural networks, and specifically large language models, have improved rapidly in capabilities and quality, they have increasingly been deployed in real-world applications, from customer service to Google search, despite the fact that they frequently make factually incorrect or undesirable statements. This trend has inspired practical and academic interest in model editing, that is, in adjusting the weights of the model to modify its likely outputs for queries relating to a specific fact or set of facts. This may be done either to amend a fact or set of facts, for instance, to fix a frequent error in the training data, or to suppress a fact or set of facts entirely, for instance, in case of dangerous knowledge. Multiple methods have been proposed to do such edits. However, at the same time, it has been shown that such model editing can be brittle and incomplete. Moreover the effectiveness of any model editing method necessarily depends on the data on which the model is trained, and, therefore, a good understanding of the interaction of the training data distribution and the way it is stored in the network is necessary and helpful to reliably perform model editing. However, working with large language models trained on real-world data does not allow us to understand this relationship or fully measure the effects of model editing. We therefore propose Behemoth, a fully synthetic data generation framework. To demonstrate the practical insights from the framework, we explore model editing in the context of simple tabular data, demonstrating surprising findings that, in some cases, echo real-world results, for instance, that in some cases restricting the update rank results in a more effective update. The code is available at https://github.com/IST-DASLab/behemoth.git.

Method: 1) Generate LLM reasoning traces, 2) truncate at fixed token-percentiles, 3) inject partial traces back into models to measure induced answer distributions via next-token probabilities. Applied to Qwen3 and gpt-oss models on GPQA Diamond and MMLU-Pro benchmarks.

Result: Accuracy and decision commitment consistently increase with more reasoning tokens; gains driven by relevant content rather than context length or generic reasoning style; stronger models can backtrack from incorrect partial traces; immediate answers often anchored in weaker model’s incorrect responses.

Conclusion: Trajectory probing provides diagnostics for efficient and safer deployment of reasoning models, informing practical trace-handling and monitoring policies that improve reliability without assuming intermediate tokens are inherently faithful explanations.

Abstract: Large language models (LLMs) increasingly solve difficult problems by producing “reasoning traces” before emitting a final response. However, it remains unclear how accuracy and decision commitment evolve along a reasoning trajectory, and whether intermediate trace segments provide answer-relevant information beyond generic length or stylistic effects. Here, we propose a protocol to systematically probe the trajectories of reasoning traces in LLMs by 1) generating a model’s reasoning trace, 2) truncating it at fixed token-percentiles, and 3) injecting each partial trace back into the model (or a different model) to measure the induced distribution over answer choices via next-token probabilities. We apply this protocol to the open-source Qwen3-4B/-8B/-14B and gpt-oss-20b/-120b models across the multiple-choice GPQA Diamond and MMLU-Pro benchmarks. We find that accuracy and decision commitment consistently increase as the percentage of provided reasoning tokens grows. These gains are primarily driven by relevant content in the model generation rather than context length or generic “reasoning style” effects. Stronger models often backtrack successfully from incorrect partial traces, but immediate answers often remain anchored in the weaker model’s incorrect response. More broadly, we show that trajectory probing provides diagnostics for efficient and safer deployment of reasoning models as the measurements can inform practical trace-handling and monitoring policies that improve reliability without assuming intermediate tokens are inherently faithful explanations.

Method: Grammar-based approach that segments unlabeled trajectories into skills and induces hierarchical structure over them. Works with high-dimensional, pixel-based environments without requiring action labels or rewards.

Result: Method produces more structured and semantically meaningful hierarchies than existing baselines. Discovered hierarchies accelerate and stabilize learning on downstream RL tasks.

Conclusion: Grammar-based approach enables unsupervised discovery of hierarchical skill structures in complex pixel-based environments, improving downstream RL performance.

Abstract: We consider the problem of unsupervised skill segmentation and hierarchical structure discovery in reinforcement learning. While recent approaches have sought to segment trajectories into reusable skills or options, most rely on action labels, rewards, or handcrafted annotations, limiting their applicability. We propose a method that segments unlabelled trajectories into skills and induces a hierarchical structure over them using a grammar-based approach. The resulting hierarchy captures both low-level behaviours and their composition into higher-level skills. We evaluate our approach in high-dimensional, pixel-based environments, including Craftax and the full, unmodified version of Minecraft. Using metrics for skill segmentation, reuse, and hierarchy quality, we find that our method consistently produces more structured and semantically meaningful hierarchies than existing baselines. Furthermore, as a proof of concept for utility, we demonstrate that these discovered hierarchies accelerate and stabilise learning on downstream reinforcement learning tasks.

Method: Two-step approach: 1) Train an ensemble of MLPs to execute local instructions of a single node, 2) Use trained MLP ensemble as update function within GNN during inference. Leverages Neural Tangent Kernel theory to show local instructions can be learned from small training sets.

Result: Proves exact learnability for graph algorithms without error and with high probability. Establishes rigorous learnability result for LOCAL model of distributed computation. Demonstrates positive learnability for message flooding, BFS, DFS, and Bellman-Ford algorithms.

Conclusion: Graph neural networks can learn to execute graph algorithms exactly under bounded-degree and finite-precision constraints, providing theoretical foundation for algorithmic learning in GNNs.

Abstract: Understanding what graph neural networks can learn, especially their ability to learn to execute algorithms, remains a central theoretical challenge. In this work, we prove exact learnability results for graph algorithms under bounded-degree and finite-precision constraints. Our approach follows a two-step process. First, we train an ensemble of multi-layer perceptrons (MLPs) to execute the local instructions of a single node. Second, during inference, we use the trained MLP ensemble as the update function within a graph neural network (GNN). Leveraging Neural Tangent Kernel (NTK) theory, we show that local instructions can be learned from a small training set, enabling the complete graph algorithm to be executed during inference without error and with high probability. To illustrate the learning power of our setting, we establish a rigorous learnability result for the LOCAL model of distributed computation. We further demonstrate positive learnability results for widely studied algorithms such as message flooding, breadth-first and depth-first search, and Bellman-Ford.

Method: Analyzes stochastic linear bandits with parameter noise model. Provides upper bounds using explore-exploit algorithms and lower bounds via information-theoretic arguments. Specifically examines ℓ_p unit balls with p ≤ 2 and their dual norms.

Result: Shows regret upper bound of Õ(√(dT log(K/δ)σ²_max)) and lower bound of Ω̃(d√(Tσ²_max)). For ℓ_p unit balls, achieves minimax regret Θ̃(√(dTσ²_q)) where σ²_q ≤ 4, significantly better than classic additive noise model’s d√T.

Conclusion: Parameter noise model enables substantially better regret bounds than additive noise model, especially for structured action sets. Simple explore-exploit algorithms can achieve near-optimal performance in this setting.

Abstract: We study the stochastic linear bandits with parameter noise model, in which the reward of action $a$ is $a^\top θ$ where $θ$ is sampled i.i.d. We show a regret upper bound of $\widetilde{O} (\sqrt{d T \log (K/δ) σ^2_{\max})}$ for a horizon $T$, general action set of size $K$ of dimension $d$, and where $σ^2_{\max}$ is the maximal variance of the reward for any action. We further provide a lower bound of $\widetildeΩ (d \sqrt{T σ^2_{\max}})$ which is tight (up to logarithmic factors) whenever $\log (K) \approx d$. For more specific action sets, $\ell_p$ unit balls with $p \leq 2$ and dual norm $q$, we show that the minimax regret is $\widetildeΘ (\sqrt{dT σ^2_q)}$, where $σ^2_q$ is a variance-dependent quantity that is always at most $4$. This is in contrast to the minimax regret attainable for such sets in the classic additive noise model, where the regret is of order $d \sqrt{T}$. Surprisingly, we show that this optimal (up to logarithmic factors) regret bound is attainable using a very simple explore-exploit algorithm.

Method: Proposes a Transformer-based mechanism using parallel embedding streams to isolate each interchangeable token’s contribution, combined with aggregated attention for structured information sharing across streams.

Result: Experimental results confirm theoretical guarantees and show substantial performance gains on open-vocabulary tasks requiring generalization to novel symbols.

Conclusion: The proposed method provides a principled way to handle interchangeable tokens in neural architectures, enabling better generalization to unseen symbols while maintaining semantic invariance.

Abstract: Current neural architectures lack a principled way to handle interchangeable tokens, i.e., symbols that are semantically equivalent yet distinguishable, such as bound variables. As a result, models trained on fixed vocabularies often struggle to generalize to unseen symbols, even when the underlying semantics remain unchanged. We propose a novel Transformer-based mechanism that is provably invariant to the renaming of interchangeable tokens. Our approach employs parallel embedding streams to isolate the contribution of each interchangeable token in the input, combined with an aggregated attention mechanism that enables structured information sharing across streams. Experimental results confirm the theoretical guarantees of our method and demonstrate substantial performance gains on open-vocabulary tasks that require generalization to novel symbols.

Method: SPAN adapts the low rank KHRONOS framework by integrating a learnable preprocessing layer with a separable tensor product B-spline basis for function approximation in RL.

Result: SPAN achieves 30-50% improvement in sample efficiency and 1.3-9 times higher success rates across benchmarks compared to MLP baselines, with superior anytime performance and robustness to hyperparameter variations.

Conclusion: SPAN is a viable, high-performance alternative for learning intrinsically efficient policies in resource-limited RL settings, offering better sample efficiency and success rates than traditional MLPs.

Abstract: Deep reinforcement learning (RL) is increasingly deployed in resource-constrained environments, yet the go-to function approximators - multilayer perceptrons (MLPs) - are often parameter-inefficient due to an imperfect inductive bias for the smooth structure of many value functions. This mismatch can also hinder sample efficiency and slow policy learning in this capacity-limited regime. Although model compression techniques exist, they operate post-hoc and do not improve learning efficiency. Recent spline-based separable architectures - such as Kolmogorov-Arnold Networks (KANs) - have been shown to offer parameter efficiency but are widely reported to exhibit significant computational overhead, especially at scale. In seeking to address these limitations, this work introduces SPAN (SPline-based Adaptive Networks), a novel function approximation approach to RL. SPAN adapts the low rank KHRONOS framework by integrating a learnable preprocessing layer with a separable tensor product B-spline basis. SPAN is evaluated across discrete (PPO) and high-dimensional continuous (SAC) control tasks, as well as offline settings (Minari/D4RL). Empirical results demonstrate that SPAN achieves a 30-50% improvement in sample efficiency and 1.3-9 times higher success rates across benchmarks compared to MLP baselines. Furthermore, SPAN demonstrates superior anytime performance and robustness to hyperparameter variations, suggesting it as a viable, high performance alternative for learning intrinsically efficient policies in resource-limited settings.

Method: Proposes MeshGraphNet-Transformer (MGN-T) that integrates a physics-attention Transformer as a global processor with MeshGraphNets’ geometric inductive bias. The Transformer updates all nodal states simultaneously while preserving node and edge attributes, enabling direct capture of long-range physical interactions without deep message-passing stacks or hierarchical meshes.

Result: MGN-T successfully handles industrial-scale meshes for impact dynamics where standard MGN fails. It accurately models self-contact, plasticity, and multivariate outputs including internal plastic variables. Outperforms state-of-the-art approaches on classical benchmarks with higher accuracy and practical efficiency using fewer parameters.

Conclusion: MGN-T provides an effective solution for efficient physics simulation on high-resolution meshes by combining the global modeling capabilities of Transformers with geometric inductive biases, enabling industrial-scale applications.

Abstract: We present MeshGraphNet-Transformer (MGN-T), a novel architecture that combines the global modeling capabilities of Transformers with the geometric inductive bias of MeshGraphNets, while preserving a mesh-based graph representation. MGN-T overcomes a key limitation of standard MGN, the inefficient long-range information propagation caused by iterative message passing on large, high-resolution meshes. A physics-attention Transformer serves as a global processor, updating all nodal states simultaneously while explicitly retaining node and edge attributes. By directly capturing long-range physical interactions, MGN-T eliminates the need for deep message-passing stacks or hierarchical, coarsened meshes, enabling efficient learning on high-resolution meshes with varying geometries, topologies, and boundary conditions at an industrial scale. We demonstrate that MGN-T successfully handles industrial-scale meshes for impact dynamics, a setting in which standard MGN fails due message-passing under-reaching. The method accurately models self-contact, plasticity, and multivariate outputs, including internal, phenomenological plastic variables. Moreover, MGN-T outperforms state-of-the-art approaches on classical benchmarks, achieving higher accuracy while maintaining practical efficiency, using only a fraction of the parameters required by competing baselines.

Method: The authors propose a variational framework that views transformer layers as iterations of an optimization algorithm. Self-attention implements gradient steps on an interaction energy, while MLP layers correspond to gradient updates of a potential energy. Standard transformers emerge as vanilla gradient descent on the composite objective using Lie-Trotter splitting.

Result: As a proof of concept, the authors introduce a Nesterov-style accelerated transformer that preserves the same attention and MLP oracles. This architecture consistently outperforms a nanoGPT baseline on TinyStories and OpenWebText datasets.

Conclusion: The optimization-theoretic perspective on transformers enables principled architectural design and can translate into practical performance gains, as demonstrated by the accelerated transformer variant.

Abstract: We propose a variational framework that interprets transformer layers as iterations of an optimization algorithm acting on token embeddings. In this view, self-attention implements a gradient step of an interaction energy, while MLP layers correspond to gradient updates of a potential energy. Standard GPT-style transformers emerge as vanilla gradient descent on the resulting composite objective, implemented via Lie–Trotter splitting between these two energy functionals. This perspective enables principled architectural design using classical optimization ideas. As a proof of concept, we introduce a Nesterov-style accelerated transformer that preserves the same attention and MLP oracles. The resulting architecture consistently outperforms a nanoGPT baseline on TinyStories and OpenWebText, demonstrating that optimization-theoretic insights can translate into practical gains.

Method: TriSpec proposes a ternary speculative decoding framework that introduces a lightweight proxy model. The system has three verification states: approve (easy tokens verified by proxy), reject (incorrect tokens), and uncertain (requires full target model verification). This reduces target model invocations by only using it for uncertain tokens.

Result: Experiments on Qwen3 and DeepSeek-R1-Distill-Qwen/LLaMA families show TriSpec achieves up to 35% speedup over standard speculative decoding, with up to 50% fewer target model invocations while maintaining comparable accuracy. It can be integrated with state-of-the-art SD methods like EAGLE-3 for further improvements.

Conclusion: TriSpec successfully addresses the verification cost bottleneck in speculative decoding through its ternary framework with lightweight proxy verification, achieving significant speedups while maintaining accuracy, representing an important advancement in LLM inference efficiency.

Abstract: Inference efficiency in Large Language Models (LLMs) is fundamentally limited by their serial, autoregressive generation, especially as reasoning becomes a key capability and response sequences grow longer. Speculative decoding (SD) offers a powerful solution, providing significant speed-ups through its lightweight drafting and parallel verification mechanism. While existing work has nearly saturated improvements in draft effectiveness and efficiency, this paper advances SD from a new yet critical perspective: the verification cost. We propose TriSpec, a novel ternary SD framework that, at its core, introduces a lightweight proxy to significantly reduce computational cost by approving easily verifiable draft sequences and engaging the full target model only when encountering uncertain tokens. TriSpec can be integrated with state-of-the-art SD methods like EAGLE-3 to further reduce verification costs, achieving greater acceleration. Extensive experiments on the Qwen3 and DeepSeek-R1-Distill-Qwen/LLaMA families show that TriSpec achieves up to 35% speedup over standard SD, with up to 50% fewer target model invocations while maintaining comparable accuracy.

Method: TEON generalizes Muon by modeling gradients as structured higher-order tensors rather than independent matrices per layer. This enables orthogonalization across layers. The paper develops a practical instantiation with theoretical convergence guarantees and evaluates various approximate SVD schemes.

Result: TEON consistently improves training and validation perplexity across GPT-style (130M-774M) and LLaMA-style (60M-1B) models. It shows strong robustness under different approximate SVD schemes and demonstrates improved convergence over layer-wise Muon.

Conclusion: TEON provides a principled tensor-based generalization of Muon that captures cross-layer gradient relationships, leading to better optimization performance and convergence guarantees for large language model pre-training.

Abstract: The Muon optimizer has demonstrated strong empirical performance in pre-training large language models by performing matrix-level gradient (or momentum) orthogonalization in each layer independently. In this work, we propose TEON, a principled generalization of Muon that extends orthogonalization beyond individual layers by modeling the gradients of a neural network as a structured higher-order tensor. We present TEON’s improved convergence guarantee over layer-wise Muon, and further develop a practical instantiation of TEON based on the theoretical analysis with corresponding ablation. We evaluate our approach on two widely adopted architectures: GPT-style models, ranging from 130M to 774M parameters, and LLaMA-style models, ranging from 60M to 1B parameters. Experimental results show that TEON consistently improves training and validation perplexity across model scales and exhibits strong robustness under various approximate SVD schemes.

Method: Uses Implicit Function Theorem to establish mapping between data space and weight representation space. Implements hypernetwork framework that maps instance-specific embeddings to INR weights, applied to 2D and 3D datasets.

Result: Achieves competitive performance with existing baselines on downstream classification tasks across 2D and 3D datasets. Provides theoretical foundation for understanding weight space representations.

Conclusion: Establishes theoretical framework for Weight Space Learning using Implicit Function Theorem, offering foundation for future investigations into network weights as data representations.

Abstract: Weight Space Learning (WSL), which frames neural network weights as a data modality, is an emerging field with potential for tasks like meta-learning or transfer learning. Particularly, Implicit Neural Representations (INRs) provide a convenient testbed, where each set of weights determines the corresponding individual data sample as a mapping from coordinates to contextual values. So far, a precise theoretical explanation for the mechanism of encoding semantics of data into network weights is still missing. In this work, we deploy the Implicit Function Theorem (IFT) to establish a rigorous mapping between the data space and its latent weight representation space. We analyze a framework that maps instance-specific embeddings to INR weights via a shared hypernetwork, achieving performance competitive with existing baselines on downstream classification tasks across 2D and 3D datasets. These findings offer a theoretical lens for future investigations into network weights.

Method: Transformed color-coded traffic maps into concentric ring-based descriptions to characterize traffic conditions. Used Partial Least Squares Regression to predict pollution levels based on these traffic intensity representations, optimizing with various training samples for best predictive performance.

Result: Developed a predictive model linking traffic intensity to air pollution levels, achieving optimized performance through training sample variations. The method provides insights into the relationship between specific pollutants and traffic patterns.

Conclusion: The workflow successfully demonstrates how traffic data can be used for hyper-local, dynamic air quality forecasting, with potential applicability to other cities beyond Mexico City.

Abstract: Air pollution is a chronic problem in large cities worldwide and awareness is rising as the long-term health implications become clearer. Vehicular traffic has been identified as a major contributor to poor air quality. In a lot of cities the publicly available air quality measurements and forecasts are coarse-grained both in space and time. However, in general, real-time traffic intensity data is openly available in various forms and is fine-grained. In this paper, we present an in-depth study of pollution sensor measurements combined with traffic data from Mexico City. We analyse and model the relationship between traffic intensity and air quality with the aim to provide hyper-local, dynamic air quality forecasts. We developed an innovative method to represent traffic intensities by transforming simple colour-coded traffic maps into concentric ring-based descriptions, enabling improved characterisation of traffic conditions. Using Partial Least Squares Regression, we predict pollution levels based on these newly defined traffic intensities. The model was optimised with various training samples to achieve the best predictive performance and gain insights into the relationship between pollutants and traffic. The workflow we have designed is straightforward and adaptable to other contexts, like other cities beyond the specifics of our dataset.

Method: Analyzes fairness of crowdsourced aggregation methods within ε-fairness framework for Majority Vote and Optimal Bayesian aggregation. Derives upper bounds on fairness gaps, shows convergence properties, and generalizes a multiclass fairness post-processing algorithm from continuous to discrete setting to enforce strict demographic parity constraints.

Result: Theoretical analysis shows fairness gap of aggregated consensus converges exponentially fast to ground-truth under interpretable conditions. Experiments on synthetic and real datasets demonstrate effectiveness of the approach and corroborate theoretical insights.

Conclusion: The paper addresses the gap in fairness analysis for crowdsourced aggregation, providing theoretical guarantees and practical post-processing methods to enforce demographic parity fairness constraints in label aggregation systems.

Abstract: As acquiring reliable ground-truth labels is usually costly, or infeasible, crowdsourcing and aggregation of noisy human annotations is the typical resort. Aggregating subjective labels, though, may amplify individual biases, particularly regarding sensitive features, raising fairness concerns. Nonetheless, fairness in crowdsourced aggregation remains largely unexplored, with no existing convergence guarantees and only limited post-processing approaches for enforcing $\varepsilon$-fairness under demographic parity. We address this gap by analyzing the fairness s of crowdsourced aggregation methods within the $\varepsilon$-fairness framework, for Majority Vote and Optimal Bayesian aggregation. In the small-crowd regime, we derive an upper bound on the fairness gap of Majority Vote in terms of the fairness gaps of the individual annotators. We further show that the fairness gap of the aggregated consensus converges exponentially fast to that of the ground-truth under interpretable conditions. Since ground-truth itself may still be unfair, we generalize a state-of-the-art multiclass fairness post-processing algorithm from the continuous to the discrete setting, which enforces strict demographic parity constraints to any aggregation rule. Experiments on synthetic and real datasets demonstrate the effectiveness of our approach and corroborate the theoretical insights.

Method: SDG injects noise into entire historical interaction sequences and jointly reconstructs all interaction embeddings through conditional denoising process. Uses cross-attention denoising decoder to guide destination sequence reconstruction, optimized end-to-end.

Result: Extensive experiments on various temporal graph benchmarks show SDG consistently achieves state-of-the-art performance in temporal link prediction.

Conclusion: SDG provides a novel generative approach to temporal link prediction that captures interaction distributions and uncertainty better than discriminative models.

Abstract: Temporal link prediction in dynamic graphs is a fundamental problem in many real-world systems. Existing temporal graph neural networks mainly focus on learning representations of historical interactions. Despite their strong performance, these models are still purely discriminative, producing point estimates for future links and lacking an explicit mechanism to capture the uncertainty and sequential structure of future temporal interactions. In this paper, we propose SDG, a novel sequence-level diffusion framework that unifies dynamic graph learning with generative denoising. Specifically, SDG injects noise into the entire historical interaction sequence and jointly reconstructs all interaction embeddings through a conditional denoising process, thereby enabling the model to capture more comprehensive interaction distributions. To align the generative process with temporal link prediction, we employ a cross-attention denoising decoder to guide the reconstruction of the destination sequence and optimize the model in an end-to-end manner. Extensive experiments on various temporal graph benchmarks show that SDG consistently achieves state-of-the-art performance in the temporal link prediction task.

Method: Benchmark comparison of four approaches: 1) Traditional Bayesian inverse with forward regression model and MCMC sampling, 2) Conditional Generative Adversarial Networks (cGANs), 3) Invertible Neural Networks, 4) Conditional Flow Matching. Applied to gas turbine combustor design mapping 6 design parameters to 3 performance labels.

Result: Conditional Flow Matching consistently outperformed all competing approaches across multiple metrics evaluating accuracy of generated designs’ labels and diversity. Performance was also studied as a function of training dataset size.

Conclusion: Generative learning methods, particularly Conditional Flow Matching, show superior performance for inverse design problems compared to traditional Bayesian approaches, offering promising tools for engineering design applications.

Abstract: Generative learning generates high dimensional data based on low dimensional conditions, also called prompts. Therefore, generative learning algorithms are eligible for solving (Bayesian) inverse problems. In this article we compare a traditional Bayesian inverse approach based on a forward regression model and a prior sampled with the Markov Chain Monte Carlo method with three state of the art generative learning models, namely conditional Generative Adversarial Networks, Invertible Neural Networks and Conditional Flow Matching. We apply them to a problem of gas turbine combustor design where we map six independent design parameters to three performance labels. We propose several metrics for the evaluation of this inverse design approaches and measure the accuracy of the labels of the generated designs along with the diversity. We also study the performance as a function of the training dataset size. Our benchmark has a clear winner, as Conditional Flow Matching consistently outperforms all competing approaches.

Method: Proposes a guided stochastic sampling method that augments diffusion model sampling with physics-based guidance from PDE residuals and observational constraints. Embeds this in a Sequential Monte Carlo (SMC) framework to create a scalable generative PDE solver.

Result: The method produces solution fields with lower numerical error than existing state-of-the-art generative methods across multiple benchmark PDE systems, including multiphysics and interacting PDE systems.

Conclusion: The physics-guided stochastic sampling approach within an SMC framework provides an effective generative PDE solver that ensures physical admissibility while achieving superior accuracy compared to existing methods.

Abstract: We introduce a guided stochastic sampling method that augments sampling from diffusion models with physics-based guidance derived from partial differential equation (PDE) residuals and observational constraints, ensuring generated samples remain physically admissible. We embed this sampling procedure within a new Sequential Monte Carlo (SMC) framework, yielding a scalable generative PDE solver. Across multiple benchmark PDE systems as well as multiphysics and interacting PDE systems, our method produces solution fields with lower numerical error than existing state-of-the-art generative methods.

Method: Decoupled Diffusion Inverse Solver (DDIS) uses unconditional diffusion for coefficient prior learning and a neural operator for explicit forward PDE modeling. Includes Decoupled Annealing Posterior Sampling (DAPS) to prevent over-smoothing.

Result: State-of-the-art performance under sparse observation: 11% improvement in l2 error and 54% improvement in spectral error on average. With only 1% data, maintains 40% advantage in l2 error over joint models.

Conclusion: Decoupled design enables superior data efficiency and physics-informed learning, theoretically avoids guidance attenuation, and empirically outperforms joint models in inverse PDE problems with limited data.

Abstract: We propose a data-efficient, physics-aware generative framework in function space for inverse PDE problems. Existing plug-and-play diffusion posterior samplers represent physics implicitly through joint coefficient-solution modeling, requiring substantial paired supervision. In contrast, our Decoupled Diffusion Inverse Solver (DDIS) employs a decoupled design: an unconditional diffusion learns the coefficient prior, while a neural operator explicitly models the forward PDE for guidance. This decoupling enables superior data efficiency and effective physics-informed learning, while naturally supporting Decoupled Annealing Posterior Sampling (DAPS) to avoid over-smoothing in Diffusion Posterior Sampling (DPS). Theoretically, we prove that DDIS avoids the guidance attenuation failure of joint models when training data is scarce. Empirically, DDIS achieves state-of-the-art performance under sparse observation, improving $l_2$ error by 11% and spectral error by 54% on average; when data is limited to 1%, DDIS maintains accuracy with 40% advantage in $l_2$ error compared to joint models.

Method: FC-KAN combines outputs from B-splines, wavelets, and radial basis functions using various combination methods including sum, element-wise product, quadratic/cubic representations, concatenation, and linear transformations. The approach applies these combinations through element-wise operations on low-dimensional data.

Result: Two FC-KAN variants (B-splines + DoG and B-splines + linear transformations as quadratic functions) outperformed MLPs and other KANs (BSRBF-KAN, EfficientKAN, FastKAN, FasterKAN) on MNIST and Fashion-MNIST datasets across 5 independent training runs.

Conclusion: FC-KAN demonstrates that combining mathematical functions can improve KAN performance, suggesting this approach could guide future KAN architecture design for better results.

Abstract: In this paper, we introduce FC-KAN, a Kolmogorov-Arnold Network (KAN) that leverages combinations of popular mathematical functions such as B-splines, wavelets, and radial basis functions on low-dimensional data through element-wise operations. We explore several methods for combining the outputs of these functions, including sum, element-wise product, the addition of sum and element-wise product, representations of quadratic and cubic functions, concatenation, linear transformation of the concatenated output, and others. In our experiments, we compare FC-KAN with a multi-layer perceptron network (MLP) and other existing KANs, such as BSRBF-KAN, EfficientKAN, FastKAN, and FasterKAN, on the MNIST and Fashion-MNIST datasets. Two variants of FC-KAN, which use a combination of outputs from B-splines and Difference of Gaussians (DoG) and from B-splines and linear transformations in the form of a quadratic function, outperformed overall other models on the average of 5 independent training runs. We expect that FC-KAN can leverage function combinations to design future KANs. Our repository is publicly available at: https://github.com/hoangthangta/FC_KAN.

Method: Represents text as interpretable knowledge graphs with facts as triples. Analyzes factual consistency across multiple LLM responses using sampling-based approach without external resources or training data. Computes fine-grained hallucination scores at fact level.

Result: Competitive performance with leading sentence-level sampling methods while providing more detailed interpretable insights. Achieves 35.5% increase in factual content for hallucination correction vs baseline (sentence-level SelfCheckGPT only 10.6%). Introduces FavaMultiSamples dataset for evaluating sampling-based methods.

Conclusion: FactSelfCheck enables fine-grained fact-level hallucination detection with superior correction capabilities. Granular approach provides more precise identification and correction of hallucinated content. Contributes new dataset to research community.

Abstract: Large Language Models (LLMs) frequently generate hallucinated content, posing significant challenges for applications where factuality is crucial. While existing hallucination detection methods typically operate at the sentence level or passage level, we propose FactSelfCheck, a novel zero-resource black-box sampling-based method that enables fine-grained fact-level detection. Our approach represents text as interpretable knowledge graphs consisting of facts in the form of triples, providing clearer insights into content factuality than traditional approaches. Through analyzing factual consistency across multiple LLM responses, we compute fine-grained hallucination scores without requiring external resources or training data. Our evaluation demonstrates that FactSelfCheck performs competitively with leading sentence-level sampling-based methods while providing more detailed and interpretable insights. Most notably, our fact-level approach significantly improves hallucination correction, achieving a 35.5% increase in factual content compared to the baseline, while sentence-level SelfCheckGPT yields only a 10.6% improvement. The granular nature of our detection enables more precise identification and correction of hallucinated content. Additionally, we contribute FavaMultiSamples, a novel dataset that addresses a gap in the field by providing the research community with a second dataset for evaluating sampling-based methods.

Method: NeUQI efficiently determines near-optimal initialization for uniform quantization by simplifying the joint optimization of scale and zero-point parameters. It derives zero-point for a given scale, reducing the problem to scale-only optimization.

Result: NeUQI consistently outperforms existing methods on LLaMA and Qwen families across various settings and tasks. When combined with lightweight distillation, it even surpasses PV-tuning, a more resource-intensive method.

Conclusion: NeUQI provides an effective solution for quantization parameter initialization that improves LLM deployment efficiency on consumer hardware while maintaining performance.

Abstract: Large language models (LLMs) achieve impressive performance across domains but face significant challenges when deployed on consumer-grade GPUs or personal devices such as laptops, due to high memory consumption and inference costs. Post-training quantization (PTQ) of LLMs offers a promising solution that reduces their memory footprint and decoding latency. In practice, PTQ with uniform quantization representation is favored due to its efficiency and ease of deployment, as uniform quantization is widely supported by mainstream hardware and software libraries. Recent studies on low-bit uniform quantization have led to noticeable improvements in post-quantization model performance; however, they mainly focus on quantization methodologies, while the initialization of quantization parameters remains underexplored and still relies on the conventional Min-Max formula. In this work, we identify the limitations of the Min-Max formula, move beyond its constraints, and propose NeUQI, a method that efficiently determines near-optimal initialization for uniform quantization. Our NeUQI simplifies the joint optimization of the scale and zero-point by deriving the zero-point for a given scale, thereby reducing the problem to a scale-only optimization. Benefiting from the improved quantization parameters, our NeUQI consistently outperforms existing methods in the experiments with the LLaMA and Qwen families on various settings and tasks. Furthermore, when combined with a lightweight distillation strategy, NeUQI even achieves superior performance to PV-tuning, a considerably more resource-intensive method.

Method: Thoughtbubbles modifies transformers to learn parallel adaptive computation in latent space by learning to fork or delete residual streams. Tokens requiring more computation form “bubbles” of cloned residuals in the network middle, learned purely through language modeling loss during pretraining.

Result: Using half the training budget, Thoughtbubbles outperforms standard decoder LMs and non-adaptive parallel computation approaches on perplexity and zero-shot evaluations across model sizes (150M to 1.9B). Achieves competitive GSM8K results with half the baseline’s token budget.

Conclusion: Thoughtbubbles enables models to learn adaptive computation during pretraining, paving the way for unified train-time and test-time scaling behaviors through implicit parallel computation in latent space.

Abstract: Current approaches for scaling inference-time compute in transformers train them to emit explicit chain-of-thought tokens before producing an answer. While these methods are powerful, they are limited because they cannot be applied during pretraining and rely solely on serially-generated, natural-language verbalization. In this work, we propose Thoughtbubbles, a transformer variant that natively performs parallel adaptive computation in latent space by learning to fork or delete residual streams. Thus, tokens requiring more computation can form a “bubble” of cloned residuals in the middle of the network. Crucially, this behavior is learned during pretraining with only language modeling loss. Using half of the training budget, Thoughtbubbles outperforms the perplexity and zero-shot evals of both standard decoder LMs and those using non-adaptive parallel computation approaches. These results hold across model sizes from 150M to 1.9B. Thoughtbubbles achieves competitive GSM8K results using half of the baseline’s token budget. The implicit nature of our method enables models to begin learning adaptive computation at pretraining time, paving the way to unified train-time and test-time scaling behaviors.

Method: Theoretical analysis showing GRPO’s implicit contrastive objective, connecting it to DPO, and empirical validation with minimal 2-rollout configuration (2-GRPO).

Result: 2-GRPO retains 98.1% performance of 16-GRPO while requiring only 12.5% of rollouts and 21% training time, demonstrating group size is not the key factor.

Conclusion: GRPO’s effectiveness stems from contrastive learning principles, not group size, offering new perspectives for efficient LLM post-training algorithm design.

Abstract: Group Relative Policy Optimization (GRPO) has emerged as a prominent reinforcement learning algorithm for post-training Large Language Models. Different from critic-based methods such as PPO, GRPO estimates the advantage function using group-level statistics to reduce the variance of policy gradient estimators. While the prevailing view attributes GRPO’s effectiveness to large group sizes for accurate advantage estimation, we propose a different perspective. We demonstrate that the efficacy of GRPO stems from its implicit contrastive objective in the optimization, which helps reduce variance via the control variate method. This perspective establishes a fundamental connection between GRPO and DPO, wherein group size influences only the Monte Carlo estimators of the contrastive objective. To validate this, we investigate the minimal two-rollout case (2-GRPO), a configuration permissible under the contrastive framework but typically considered insufficient for reward normalization. We provide a rigorous theoretical analysis of 2-GRPO and empirically validate its effectiveness: 2-GRPO retains 98.1% of the performance of 16-GRPO, while requiring only 12.5% of the rollouts and 21% of the training time. This study offers a new perspective for future algorithm design in LLM post-training.

Method: ACE (Agentic Context Engineering) treats contexts as evolving playbooks that accumulate, refine, and organize strategies through a modular process of generation, reflection, and curation. It prevents collapse with structured, incremental updates that preserve detailed knowledge and scales with long-context models.

Result: ACE consistently outperforms strong baselines: +10.6% on agents and +8.6% on finance benchmarks, while significantly reducing adaptation latency and rollout cost. On the AppWorld leaderboard, ACE matches the top-ranked production-level agent on overall average and surpasses it on the harder test-challenge split using a smaller open-source model.

Conclusion: Comprehensive, evolving contexts enable scalable, efficient, and self-improving LLM systems with low overhead, demonstrating that ACE can adapt effectively without labeled supervision by leveraging natural execution feedback.

Abstract: Large language model (LLM) applications such as agents and domain-specific reasoning increasingly rely on context adaptation – modifying inputs with instructions, strategies, or evidence, rather than weight updates. Prior approaches improve usability but often suffer from brevity bias, which drops domain insights for concise summaries, and from context collapse, where iterative rewriting erodes details over time. Building on the adaptive memory introduced by Dynamic Cheatsheet, we introduce ACE (Agentic Context Engineering), a framework that treats contexts as evolving playbooks that accumulate, refine, and organize strategies through a modular process of generation, reflection, and curation. ACE prevents collapse with structured, incremental updates that preserve detailed knowledge and scale with long-context models. Across agent and domain-specific benchmarks, ACE optimizes contexts both offline (e.g., system prompts) and online (e.g., agent memory), consistently outperforming strong baselines: +10.6% on agents and +8.6% on finance, while significantly reducing adaptation latency and rollout cost. Notably, ACE could adapt effectively without labeled supervision and instead by leveraging natural execution feedback. On the AppWorld leaderboard, ACE matches the top-ranked production-level agent on the overall average and surpasses it on the harder test-challenge split, despite using a smaller open-source model. These results show that comprehensive, evolving contexts enable scalable, efficient, and self-improving LLM systems with low overhead.

Method: Proposes SPECS: (1) generates introspective preference data pairs via self-distillation without external teachers, (2) uses preference-based training (DPO-like) to learn shallow surface-form criteria (format/structure/style), and (3) hands off to RL for deep reasoning.

Result: Improves MEGA-Bench by 4.1% and MathVista by 12.2%, reduces in-distribution “stuckness,” improves exploration, stabilizes training, and raises performance ceiling across multiple multimodal benchmarks.

Conclusion: Decoupling multimodal learning into format-focused preference training followed by RL for reasoning yields better generalization and performance than SFT-based cold starts.

Abstract: Reinforcement learning (RL) with verifiable rewards has recently catalyzed a wave of “MLLM-r1” approaches that bring RL to vision language models. Most representative paradigms begin with a cold start, typically employing supervised fine-tuning (SFT), to initialize the policy before RL. However, SFT-based cold start adopts the reasoning paradigm intertwined with task solution and output format, which may induce instruction-style overfitting, weakens out-of-distribution generalization, and ultimately affects downstream RL. We revisit the cold start along two views, its training method and data construction, and introduce the Generalization Factor (GF) coefficient to quantify the generalization capability under different methods. Our empirical study finds that preference-based training methods (e.g. DPO) generalizes better than SFT-based methods in cold start. Motivated by this, we propose SPECS-a Self-distilled, Preference-based Cold Start framework that decouples multimodal learning: (1) generates introspective preference data pairs via self-distillation, avoiding reliance on larger teachers or manual annotation; (2) performs preference-based training to learn, focusing on shallow, transferable surface-form criteria (format, structure, style) rather than memorizing content; and (3) hands off to RL with verifiable rewards for deep reasoning results. Experimental results across multiple multimodal benchmarks show that our decoupling learning framework yields consistent performance gains over strong baselines, improving MEGA-Bench by 4.1% and MathVista by 12.2%. Additional experiments indicate that SPECS contributes to reducing in-distribution “stuckness,” improving exploration, stabilizing training, and raising the performance ceiling. Project Page: https://kwen-chen.github.io/SPECS-VL/

Method: Theoretical analysis reveals connection between LTR and CL, showing weights converge to bounded neighborhood of Head-only weights. Proposes CLTR approach using standard off-the-shelf CL methods to sequentially learn Head and Tail classes without forgetting Head knowledge.

Result: Extensive experiments on CIFAR100-LT, CIFAR10-LT, ImageNet-LT, and Caltech256 validate theoretical predictions, achieving strong results across various LTR benchmarks.

Conclusion: The work bridges gap between LTR and CL, providing principled way to tackle imbalanced data challenges with standard existing CL strategies, showing CLTR mitigates gradient saturation and improves Tail learning while maintaining Head performance.

Abstract: Real-world datasets often exhibit long-tailed distributions, where a few dominant “Head” classes have abundant samples while most “Tail” classes are severely underrepresented, leading to biased learning and poor generalization for the Tail. We present a theoretical framework that reveals a previously undescribed connection between Long-Tailed Recognition (LTR) and Continual Learning (CL), the process of learning sequential tasks without forgetting prior knowledge. Our analysis demonstrates that, for models trained on imbalanced datasets, the weights converge to a bounded neighborhood of those trained exclusively on the Head, with the bound scaling as the inverse square root of the imbalance factor. Leveraging this insight, we introduce Continual Learning for Long-Tailed Recognition (CLTR), a principled approach that employs standard off-the-shelf CL methods to address LTR problems by sequentially learning Head and Tail classes without forgetting the Head. Our theoretical analysis further suggests that CLTR mitigates gradient saturation and improves Tail learning while maintaining strong Head performance. Extensive experiments on CIFAR100-LT, CIFAR10-LT, ImageNet-LT, and Caltech256 validate our theoretical predictions, achieving strong results across various LTR benchmarks. Our work bridges the gap between LTR and CL, providing a principled way to tackle imbalanced data challenges with standard existing CL strategies.

Method: Proposes Geometric-disentanglement Unlearning (GU) - a lightweight projection method based on theoretical insight that retain loss is locally invariant if and only if update direction is orthogonal to subspace spanned by retain gradients. GU can be plug-and-play with existing gradient-based unlearning methods.

Result: Experiments on TOFU, MUSE, and WMDP-cyber show GU strengthens forgetting while reducing retain drift. When added to SimNPO, achieves up to 62% improved forgetting Extraction Strength (ES) and 31% higher retain ES.

Conclusion: GU provides a theoretically grounded solution to mitigate forget-retain side effects in LLM unlearning, offering improved performance with theoretical guarantees while being lightweight and compatible with existing methods.

Abstract: Large language models (LLMs) can internalize private or harmful content, motivating unlearning that removes a forget set while preserving retaining knowledge. However, forgetting updates often cause collateral degradation on retaining knowledge, creating a persistent trade-off. Existing LLM unlearning methods are often heuristic, and other theoretical approaches rely on offline feature constructions that do not capture update-time forget-retain interaction in LLMs. To address this limitation, we aim to develop an LLM unlearning method that reduces the forget-retain trade-off with theoretical guarantees. We take a first-principles view by formalizing “no side effects” as local retain invariance under small parameter updates, and prove an equivalence under optimizer-induced geometry: the retain loss is locally invariant if and only if the update direction is orthogonal to the subspace spanned by retain gradients. Based on the insight, we propose Geometric-disentanglement Unlearning (GU), a lightweight and theoretically grounded projection that can be plug-and-play to existing gradient-based unlearning methods to mitigate forget-retain side effects. Experiments on TOFU, MUSE, and WMDP-cyber show that GU strengthens forgetting while reducing retain drift. When added to SimNPO, it achieves up to 62% improved forgetting Extraction Strength (ES) and 31% higher retain ES. We open-sourced our code in https://github.com/Lemutisme/Geometric-Unlearning.

Method: Developed a PyTorch-native library with low-coupling design, implementing state-of-the-art CP algorithms, enabling CP-specific training algorithms, online prediction, and GPU-accelerated batch processing. The library includes about 16k lines of code with 100% unit test coverage.

Result: TorchCP achieves up to 90% reduction in inference time on large datasets through GPU acceleration, supports various DL models including LLMs, and provides comprehensive uncertainty quantification capabilities with full GPU scalability.

Conclusion: TorchCP successfully bridges the gap between conformal prediction and modern deep learning, providing researchers and practitioners with a scalable, efficient tool for uncertainty quantification across cutting-edge applications including large language models.

Abstract: Conformal prediction (CP) is a powerful statistical framework that generates prediction intervals or sets with guaranteed coverage probability. While CP algorithms have evolved beyond traditional classifiers and regressors to sophisticated deep learning models like deep neural networks (DNNs), graph neural networks (GNNs), and large language models (LLMs), existing CP libraries often lack the model support and scalability for large-scale deep learning (DL) scenarios. This paper introduces TorchCP, a PyTorch-native library designed to integrate state-of-the-art CP algorithms into DL techniques, including DNN-based classifiers/regressors, GNNs, and LLMs. Released under the LGPL-3.0 license, TorchCP comprises about 16k lines of code, validated with 100% unit test coverage and detailed documentation. Notably, TorchCP enables CP-specific training algorithms, online prediction, and GPU-accelerated batch processing, achieving up to 90% reduction in inference time on large datasets. With its low-coupling design, comprehensive suite of advanced methods, and full GPU scalability, TorchCP empowers researchers and practitioners to enhance uncertainty quantification across cutting-edge applications.

Method: DDL replaces identity shortcuts with learnable rank-1 perturbations: A(X) = I - β(X)k(X)k(X)⊤, where k(X) is a unit direction and β(X) is a scalar gate. This enables interpolation between identity (β=0), orthogonal projection (β=1), and Householder reflection (β=2).

Result: Replacing Transformer residual additions with DDL improves validation loss, perplexity, and downstream evaluation accuracy on language modeling tasks, with larger gains in expanded-state settings.

Conclusion: DDL provides a principled way to make residual shortcuts learnable and state-dependent while maintaining stable training, offering better modeling of complex hidden state transitions in deep networks.

Abstract: The effectiveness of deep residual networks hinges on the identity shortcut connection. While this mechanism alleviates the vanishing-gradient problem, it also has a strictly additive inductive bias on feature transformations, limiting the network’s ability to model complex hidden state transitions. In this paper, we introduce \textbf{Deep Delta Learning (DDL)}, which generalizes the shortcut from a fixed identity map to a learnable, state-dependent linear operator. The resulting Delta Operator is a rank-1 perturbation of the identity, $\mathbf{A}(\mathbf{X}) = \mathbf{I}- β(\mathbf{X})\mathbf{k} (\mathbf{X}) \mathbf{k} (\mathbf{X})^\top$, parameterized by a unit direction $\mathbf{k}(\mathbf{X})$ and a scalar gate $β(\mathbf{X})$. We provide a spectral analysis showing that $β(\mathbf{X})$ continuously interpolates the shortcut between identity ($β=0$), orthogonal projection ($β=1$), and Householder reflection ($β=2$). Furthermore, we rewrite the residual update as a synchronized rank-1 delta write: $β$ scales both the removal of the current $\mathbf{k}$-component and the injection of the new $\mathbf{k}$-component. This unification enables explicit control of the shortcut spectrum along a data-dependent direction while retaining stable training behavior. Empirically, replacing Transformer residual additions with DDL improves validation loss and perplexity, as well as downstream evaluation accuracy on language modeling tasks, with larger gains in the expanded-state setting.

Method: Anchored Posterior Sampling (APS) introduces two key innovations: 1) quantized expectation for gradient-like guidance in discrete embedding space, and 2) anchored remasking for adaptive decoding.

Result: APS achieves state-of-the-art performance among discrete diffusion samplers on both linear and nonlinear inverse problems across standard image benchmarks, with demonstrated applications in training-free stylization, text-guided editing, and improved question answering in large-scale diffusion language models.

Conclusion: APS overcomes limitations of existing discrete diffusion posterior sampling methods and shows strong performance across diverse applications including vision and language tasks.

Abstract: While continuous diffusion models have achieved remarkable success, discrete diffusion offers a unified framework for jointly modeling text and images. Beyond unification, discrete diffusion provides faster inference, finer control, and principled training-free guidance, making it well-suited for posterior sampling. Existing approaches to posterior sampling using discrete diffusion face severe challenges: derivative-free guidance yields sparse signals, continuous relaxations limit applicability, and split Gibbs samplers suffer from the curse of dimensionality. To overcome these limitations, we introduce Anchored Posterior Sampling (APS), built on two key innovations: quantized expectation for gradient-like guidance in discrete embedding space, and anchored remasking for adaptive decoding. APS achieves state-of-the-art performance among discrete diffusion samplers on both linear and nonlinear inverse problems across the standard image benchmarks. We demonstrate the generality of APS through training-free stylization and text-guided editing. We further apply APS to a large-scale diffusion language model, showing consistent improvement in question answering.

Method: Replaces the ResNet-style residual pathway in Post-LN Transformers with a Highway-style connection that preserves gradient flow through the residual branch, preventing signal vanishing from top to bottom layers. This enables stable training without specialized initialization or complex optimization tricks.

Result: Keel trains robustly at depths exceeding 1000 layers and consistently improves perplexity and depth-scaling characteristics over Pre-LN Transformers. Demonstrates that Post-LN with Highway connections provides effective foundation for deeply scalable LLMs.

Conclusion: Post-LN Transformers with Highway-style connections enable stable training at extreme depths, opening possibilities for infinite-depth architectures and addressing fundamental limitations in current LLM scaling approaches.

Abstract: Large language model (LLM) scaling is hitting a wall. Widening models yields diminishing returns, and extending context length does not improve fundamental expressivity. In contrast, depth scaling offers theoretically superior expressivity, yet current Transformer architectures struggle to train reliably at extreme depths. We revisit the Post-LayerNorm (Post-LN) formulation, whose instability at scale caused its replacement by Pre-LN in modern LLMs. We show that the central failure mode of Post-LN arises from the ResNet-style residual pathway, which introduces gradient vanishing in deep networks. We present Keel, a Post-LN Transformer that replaces this residual path with a Highway-style connection. This modification preserves the gradient flow through the residual branch, preventing signal vanishing from the top layers to the bottom. Unlike prior methods, Keel enables stable training at extreme depths without requiring specialized initialization or complex optimization tricks. Keel trains robustly at depths exceeding 1000 layers and consistently improves perplexity and depth-scaling characteristics over Pre-LN. These findings indicate that Post-LN, when paired with a Highway-style connection, provides a simple and effective foundation for building deeply scalable LLMs, opening the possibility for future infinite-depth architectures.

Method: Proposes posterior label smoothing that derives soft labels from posterior distribution conditioned on neighborhood labels. Estimates likelihood and prior from global graph statistics, making it adaptable to different graph properties.

Result: Evaluated on 10 benchmark datasets with 8 baseline models, showing consistent improvements in classification accuracy. Soft labels mitigate overfitting and pseudo-labeling refines global label statistics.

Conclusion: Posterior label smoothing is effective for transductive node classification across diverse graph types, improving generalization through regularization and better label statistics.

Abstract: Label smoothing is a widely studied regularization technique in machine learning. However, its potential for node classification in graph-structured data, spanning homophilic to heterophilic graphs, remains largely unexplored. We introduce posterior label smoothing, a novel method for transductive node classification that derives soft labels from a posterior distribution conditioned on neighborhood labels. The likelihood and prior distributions are estimated from the global statistics of the graph structure, allowing our approach to adapt naturally to various graph properties. We evaluate our method on 10 benchmark datasets using eight baseline models, demonstrating consistent improvements in classification accuracy. The following analysis demonstrates that soft labels mitigate overfitting during training, leading to better generalization performance, and that pseudo-labeling effectively refines the global label statistics of the graph. Our code is available at https://github.com/ml-postech/PosteL.

Method: Proposes SET-TIME network with: 1) common feature extractor for temporal/spatial features, 2) secondary time-correlation feature extractor for correlation between multiple time windows, 3) domain adaptation module for cross-subject detection capability

Result: Achieves 92.00% accuracy on PRED+CT dataset and 94.00% on MODMA dataset, outperforming state-of-the-art methods. Ablation experiments confirm effectiveness of all modules.

Conclusion: SET-TIME effectively explores intrinsic spatiotemporal information of EEG signals for MDD detection by incorporating electrode spatial topology and time-window transition information, improving cross-subject detection capability.

Abstract: Recently, researchers have begun to experiment with deep learning-based methods for detecting major depressive disor-der (MDD) using electroencephalogram (EEG) signals in search of a more objective means of diagnosis. However, exist-ing spatiotemporal feature extraction methods only consider the functional correlation between multiple electrodes and temporal correlation of EEG signals, ignoring the spatial posi-tion connection information between electrodes and the conti-nuity between time windows, which reduces the model’s fea-ture extraction capabilities. To address this issue, a Spatio-temporal fused network for MDD detection with Electrode spatial Topology and adjacent TIME-window transition in-formation (SET-TIME) is proposed in this study. SET-TIME is composed by a common feature extractor, a secondary time-correlation feature extractor, and a domain adaptation (DA) module, in which the former extractor is used to obtain the temporal and spatial features, and the latter extractor can mine the correlation between multiple time windows, and the DA module is adopted to enhance cross-subject detection ca-pability. The experimental results of 10-fold cross-validation show that the proposed SET-TIME method outperforms the state-of-the-art (SOTA) method by achieving MDD detection accuracies of 92.00% and 94.00% on the public datasets PRED+CT and MODMA, respectively. Ablation experiments demonstrate the effectiveness of the multiple modules in SET-TIME, which assist in MDD detection by exploring the intrin-sic spatiotemporal information of EEG signals.

Method: GLAM uses LLMs as policies in RL agents that interact with textual environments, leveraging online reinforcement learning to progressively update the LLM policy and improve performance on spatial/navigation tasks.

Result: The paper studies whether LLMs can boost sample efficiency for online RL learning, improve generalization, and examines the impact of online learning on various FLAN-T5 variants.

Conclusion: Functional grounding through online RL can align LLMs’ abstract knowledge with specific environments, potentially improving sample efficiency and generalization in decision-making tasks.

Abstract: Recent works successfully leveraged Large Language Models’ (LLM) abilities to capture abstract knowledge about world’s physics to solve decision-making problems. Yet, the alignment between LLMs’ knowledge and the environment can be wrong and limit functional competence due to lack of grounding. In this paper, we study an approach (named GLAM) to achieve this alignment through functional grounding: we consider an agent using an LLM as a policy that is progressively updated as the agent interacts with the environment, leveraging online Reinforcement Learning to improve its performance to solve goals. Using an interactive textual environment designed to study higher-level forms of functional grounding, and a set of spatial and navigation tasks, we study several scientific questions: 1) Can LLMs boost sample efficiency for online learning of various RL tasks? 2) How can it boost different forms of generalization? 3) What is the impact of online learning? We study these questions by functionally grounding several variants (size, architecture) of FLAN-T5.

Method: Developed as part of the PyTorch Ecosystem, NeuralOperator offers tools for training, deploying, and developing neural operator models with a simple user interface and gentle learning curve.

Result: A high-quality, tested open-source package that combines cutting-edge models with customizability for operator learning applications.

Conclusion: NeuralOperator provides a comprehensive framework for advancing operator learning research and applications through accessible, well-tested tools.

Abstract: We present NeuralOperator, an open-source Python library for operator learning. Neural operators generalize neural networks to maps between function spaces instead of finite-dimensional Euclidean spaces. They can be trained and inferenced on input and output functions given at various discretizations, satisfying a discretization convergence properties. Part of the official PyTorch Ecosystem, NeuralOperator provides all the tools for training and deploying neural operator models, as well as developing new ones, in a high-quality, tested, open-source package. It combines cutting-edge models and customizability with a gentle learning curve and simple user interface for newcomers.

Method: Repurposes knowledge distillation to extract benign knowledge and remove backdoors from poisoned SSL encoders. Evaluates different teacher architectures, student models, and loss functions to optimize the distillation process for backdoor removal.

Result: Distillation reduces attack success rate from 80.87% to 27.51% with only 6.35% accuracy drop. Best performance achieved with fine-tuned teacher networks, warm-up-based student training, and attention-based distillation losses.

Conclusion: Distillation is an effective defense against backdoor attacks in SSL, successfully removing malicious triggers while preserving model performance. The approach provides a practical solution for securing distributed pre-trained encoders.

Abstract: Self-Supervised Learning (SSL) has become a prominent paradigm for pre-training encoders to learning general-purpose representations from unlabeled data and releasing them on third-party platforms for broad downstream deep learning tasks. However, SSL is vulnerable to backdoor attacks, where an adversary may train and distribute poisoned pre-training encoders to contaminate the downstream models. In this paper, we study a defense mechanism based on distillation against poisoned encoders in SSL. Traditionally, distillation transfers knowledge from a pre-trained teacher model to a student model, enabling the student to replicate or refine the teacher’s learned representations. We repurpose distillation to extract benign knowledge and remove backdoors from a poisoned pre-trained encoder to produce a clean and reliable pre-trained model. We conduct extensive experiments to evaluate the effectiveness of distillation in mitigating backdoor attacks on pre-trained encoders. Based on two state-of-the-art backdoor attacks and four widely adopted image classification datasets, our results demonstrate that distillation reduces the attack success rate from 80.87% to 27.51%, with only a 6.35% drop in model accuracy. Furthermore, by comparing four teacher architectures, three student models, and six loss functions, we find that the distillation with fine-tuned teacher networks, warm-up-based student training, and attention-based distillation losses yield the best performance.

Method: Theoretical analysis of gradient heterogeneity and its impact on optimization, showing that sign-based methods like SignSGD are less sensitive than SGD. Investigates gradient heterogeneity origins in Transformers, particularly layer normalization placement. Experimental validation through fine-tuning Transformers in NLP and vision domains.

Result: Gradient heterogeneity degrades SGD convergence but affects sign-based methods less. Adam’s coordinate-wise normalization makes it behave like soft SignSGD. Post-LN Transformer architectures exhibit particularly pronounced gradient heterogeneity. Experiments confirm theoretical analysis across NLP and vision tasks.

Conclusion: Adam’s superior performance for Transformers stems from its reduced sensitivity to gradient heterogeneity via coordinate-wise normalization, which makes it behave like a soft sign-based method. Layer normalization placement significantly affects gradient heterogeneity in Transformer architectures.

Abstract: Transformers are difficult to optimize with stochastic gradient descent (SGD) and largely rely on adaptive optimizers such as Adam. Despite their empirical success, the reasons behind Adam’s superior performance over SGD remain poorly understood. In this study, we analyze the optimization of Transformer models through the lens of \emph{gradient heterogeneity}, defined as the variation in gradient norms across parameter blocks. We provide a theoretical analysis showing that gradient heterogeneity, together with Hessian heterogeneity, degrades the convergence of gradient-based methods such as SGD, while sign-based methods are substantially less sensitive to this effect. Adam’s coordinate-wise normalization makes its update directions depend mainly on gradient signs, so Adam can be interpreted as a soft variant of SignSGD. Our analysis uses the fact that SGD and SignSGD follow steepest descent directions under different norms, and derives upper bounds on the iteration complexity with implications for learning rate scaling in SignSGD. We further investigate the origin of gradient heterogeneity in Transformer architectures and show that it is strongly influenced by the placement of layer normalization, with Post-LN architectures exhibiting particularly pronounced heterogeneity. Experimental results from fine-tuning Transformers in both NLP and vision domains validate our theoretical analysis. Code is available at https://github.com/tom4649/gradient-heterogeneity.

Method: Controlled empirical analysis comparing prominent DLWP models including U-Net, Transformer, Graph Neural Network, and Fourier Neural Operator backbones. Evaluated on synthetic 2D incompressible Navier-Stokes data and real-world global weather dynamics from WeatherBench dataset.

Result: On synthetic data: FNO performed best. On WeatherBench: ConvLSTM and SwinTransformer excelled for short-to-mid-range forecasts. For long-range rollouts (up to 50 years): GraphCast and Spherical FNO showed superior stability and physical soundness due to spherical data representation.

Conclusion: Different architectures excel in different scenarios: FNO for synthetic data, ConvLSTM/SwinTransformer for short-to-mid real-world forecasts, and spherical representations (GraphCast/Spherical FNO) for long-term stability. Provides benchmark for future DLWP development.

Abstract: A large number of Deep Learning Weather Prediction (DLWP) architectures – based on various backbones, including U-Net, Transformer, Graph Neural Network, and Fourier Neural Operator (FNO) – have demonstrated their potential at forecasting atmospheric states. However, due to differences in training protocols, forecast horizons, and data choices, it remains unclear which (if any) of these methods and architectures are most suitable for weather forecasting and for future model development. Here, we step back and provide a detailed empirical analysis, under controlled conditions, comparing and contrasting the most prominent DLWP models, along with their backbones. We accomplish this by predicting synthetic two-dimensional incompressible Navier-Stokes and real-world global weather dynamics. On synthetic data, we observe favorable performance of FNO, while on the real-world WeatherBench dataset, our results demonstrate the suitability of ConvLSTM and SwinTransformer for short-to-mid-ranged forecasts. For long-ranged weather rollouts of up to 50 years, we observe superior stability and physical soundness in architectures that formulate a spherical data representation, i.e., GraphCast and Spherical FNO. The code is available at https://github.com/amazon-science/dlwp-benchmark.

Method: Proposes DML-IL algorithm that reformulates causal imitation learning as a Conditional Moment Restriction problem using trajectory histories as instruments. Uses instrumental variable regression to handle hidden confounders.

Result: DML-IL outperforms existing causal IL baselines on continuous state-action environments including Mujoco tasks. Theoretical upper bounds on imitation gap are provided.

Conclusion: The framework successfully addresses hidden confounder issues in imitation learning through instrumental variable methods, demonstrating practical improvements in complex continuous control tasks.

Abstract: We propose a general framework for causal Imitation Learning (IL) with hidden confounders, which subsumes several existing settings. Our framework accounts for two types of hidden confounders: (a) variables observed by the expert but not by the imitator, and (b) confounding noise hidden from both. By leveraging trajectory histories as instruments, we reformulate causal IL in our framework into a Conditional Moment Restriction (CMR) problem. We propose DML-IL, an algorithm that solves this CMR problem via instrumental variable regression, and upper bound its imitation gap. Empirical evaluation on continuous state-action environments, including Mujoco tasks, demonstrates that DML-IL outperforms existing causal IL baselines.

Method: Proposes PSDNorm that leverages Monge mapping and temporal context to normalize feature maps in deep learning models for signals. The method preserves temporal dependencies while handling distribution shifts, evaluated with U-Net and transformer architectures on sleep data from 10K subjects across 10 datasets.

Result: PSDNorm achieves state-of-the-art performance on unseen left-out datasets and demonstrates greater robustness to data scarcity compared to existing normalization methods.

Conclusion: PSDNorm effectively addresses distribution shift in biomedical signal processing by incorporating temporal context and Monge mapping, offering improved generalization and data efficiency for time-series applications.

Abstract: Distribution shift poses a significant challenge in machine learning, particularly in biomedical applications using data collected across different subjects, institutions, and recording devices, such as sleep data. While existing normalization layers, BatchNorm, LayerNorm and InstanceNorm, help mitigate distribution shifts, when applied over the time dimension they ignore the dependencies and auto-correlation inherent to the vector coefficients they normalize. In this paper, we propose PSDNorm that leverages Monge mapping and temporal context to normalize feature maps in deep learning models for signals. Evaluations with architectures based on U-Net or transformer backbones trained on 10K subjects across 10 datasets, show that PSDNorm achieves state-of-the-art performance on unseen left-out datasets while being more robust to data scarcity.

Method: Combines neural operator (approximates elastodynamics Green’s operator) with denoising diffusion probabilistic model. The diffusion model is trained to correct ground motion time series generated by the neural operator.

Result: Approach enhances realism of synthetic seismograms, improving frequency biases and Goodness-Of-Fit scores. Diffusion model mitigates mid-frequency spectral falloff observed in neural operator outputs. Method shows fast and cheap inference across different site and source conditions.

Conclusion: The AI physics-based approach combining neural operator with diffusion model effectively generates realistic synthetic earthquake ground motion, addressing limitations of traditional methods while maintaining computational efficiency.

Abstract: Nuclear reactor buildings must be designed to withstand the dynamic load induced by strong ground motion earthquakes. For this reason, their structural behavior must be assessed in multiple realistic ground shaking scenarios (e.g., the Maximum Credible Earthquake). However, earthquake catalogs and recorded seismograms may not always be available in the region of interest. Therefore, synthetic earthquake ground motion is progressively being employed, although with some due precautions: earthquake physics is sometimes not well enough understood to be accurately reproduced with numerical tools, and the underlying epistemic uncertainties lead to prohibitive computational costs related to model calibration. In this study, we propose an AI physics-based approach to generate synthetic ground motion, based on the combination of a neural operator that approximates the elastodynamics Green’s operator in arbitrary source-geology setups, enhanced by a denoising diffusion probabilistic model. The diffusion model is trained to correct the ground motion time series generated by the neural operator. Our results show that such an approach promisingly enhances the realism of the generated synthetic seismograms, with frequency biases and Goodness-Of-Fit (GOF) scores being improved by the diffusion model. This indicates that the latter is capable to mitigate the mid-frequency spectral falloff observed in the time series generated by the neural operator. Our method showcases fast and cheap inference in different site and source conditions.

Method: Develops StyleDDG, a decentralized DG algorithm where devices in a peer-to-peer network share style information inferred from their datasets. Provides systematic approach to analyzing style-based DG training in decentralized networks, casting existing centralized DG algorithms within the framework and modeling StyleDDG using their formalisms.

Result: Obtains analytical conditions for convergence guarantee of StyleDDG. Experimental results on popular DG datasets show significant improvements in accuracy across target domains with minimal communication overhead compared to baseline decentralized gradient methods.

Conclusion: StyleDDG successfully enables decentralized federated domain generalization through style sharing, with formal convergence guarantees and practical effectiveness demonstrated on benchmark datasets.

Abstract: Much of federated learning (FL) focuses on settings where local dataset statistics remain the same between training and testing. However, this assumption often does not hold in practice due to distribution shifts, motivating the development of domain generalization (DG) approaches that leverage source domain data to train models capable of generalizing to unseen target domains. In this paper, we are motivated by two major gaps in existing work on FL and DG: (1) the lack of formal mathematical analysis of DG objectives; and (2) DG research in FL being limited to the star-topology architecture. We develop Decentralized Federated Domain Generalization with Style Sharing ($\textit{StyleDDG}$), a decentralized DG algorithm which allows devices in a peer-to-peer network to achieve DG based on sharing style information inferred from their datasets. Additionally, we provide the first systematic approach to analyzing style-based DG training in decentralized networks. We cast existing centralized DG algorithms within our framework, and employ their formalisms to model $\textit{StyleDDG}$. We then obtain analytical conditions under which convergence of $\textit{StyleDDG}$ can be guaranteed. Through experiments on popular DG datasets, we demonstrate that $\textit{StyleDDG}$ can obtain significant improvements in accuracy across target domains with minimal communication overhead compared to baseline decentralized gradient methods.

Method: Leverages influence functions under semantic transformation by comparing influence distributions before and after semantic inversions to identify critical poisons. Introduces multi-transform ensemble approach that identifies examples with strong, unchanged influence across transformations.

Result: Achieves F1 scores between 79.5-95.2% with precision 66-100% on sentiment classification, significantly improving over single-transform methods. Generalizes to unseen transformation types with 86% F1 score. Removing 1-3% of detected poisons restores model performance to near-clean levels.

Conclusion: Demonstrates practicality of influence-based diagnostics for defending against instruction fine-tuning attacks in real-world LLM deployment. Method works across multiple models (T5-small, DeepSeek-Coder-1.3B) and tasks (sentiment classification, math reasoning).

Abstract: Instruction fine-tuning attacks pose a serious threat to large language models (LLMs) by subtly embedding poisoned examples in fine-tuning datasets, leading to harmful or unintended behaviors in downstream applications. Detecting such attacks is challenging because poisoned data is often indistinguishable from clean data, and prior knowledge of triggers or attack strategies is rarely available. We present a detection method that requires no prior knowledge of the attack. Our approach leverages influence functions under semantic transformation by comparing influence distributions before and after semantic inversions to identify critical poisons, defined as examples whose influence is strong and remains unchanged across transformations. We introduce a multi-transform ensemble approach that achieves F1 scores between 79.5 and 95.2 percent with precision between 66 and 100 percent on sentiment classification, significantly improving over single-transform methods. Our method generalizes to unseen transformation types with an F1 score of 86 percent through cross-category validation. We demonstrate effectiveness across multiple models, including T5-small and DeepSeek-Coder-1.3B, and across tasks such as sentiment classification and math reasoning. Removing a small fraction of detected poisons, between 1 and 3 percent of the data, restores model performance to near-clean levels. These results demonstrate the practicality of influence-based diagnostics for defending against instruction fine-tuning attacks in real-world large language model deployment. Artifact available at https://github.com/lijiawei20161002/Poison-Detection. Warning: this paper contains offensive data examples.

Method: SCARLET integrates synchronized soft-label caching to reuse cached soft-labels and an enhanced Entropy Reduction Aggregation (Enhanced ERA) mechanism to resolve instability in conventional temperature-based aggregation.

Result: SCARLET achieves up to 50% reduction in communication costs compared to existing methods while maintaining competitive accuracy, consistently outperforming state-of-the-art distillation-based FL methods.

Conclusion: SCARLET provides an efficient federated learning framework that significantly reduces communication overhead while maintaining performance, addressing key limitations in conventional and distillation-based FL approaches.

Abstract: Federated Learning (FL) enables collaborative model training across decentralized clients, enhancing privacy by keeping data local. Yet conventional FL, relying on frequent parameter-sharing, suffers from high communication overhead and limited model heterogeneity. Distillation-based FL approaches address these issues by sharing predictions (soft-labels, i.e., normalized probability distributions) instead, but they often involve redundant transmissions across communication rounds, reducing efficiency. We propose SCARLET, a novel framework integrating synchronized soft-label caching and an enhanced Entropy Reduction Aggregation (Enhanced ERA) mechanism. SCARLET minimizes redundant communication by reusing cached soft-labels, achieving up to 50% reduction in communication costs compared to existing methods while maintaining competitive accuracy. Enhanced ERA resolves the fundamental instability of conventional temperature-based aggregation, ensuring robust control and high performance in diverse client scenarios. Experimental evaluations demonstrate that SCARLET consistently outperforms state-of-the-art distillation-based FL methods in terms of accuracy and communication efficiency. The implementation of SCARLET is publicly available at https://github.com/kitsuyaazuma/SCARLET.

Method: SAINT represents multi-component actions as unordered sets and models their dependencies using self-attention conditioned on the global state. The architecture is permutation-invariant and compatible with standard policy optimization algorithms.

Result: SAINT consistently outperforms strong baselines across 18 distinct combinatorial environments in three task domains, including environments with up to 1.35 × 10¹⁸ possible actions.

Conclusion: SAINT provides an effective solution for combinatorial action spaces by capturing complex joint behavior through set-based representation and attention mechanisms, demonstrating superior sample efficiency and performance.

Abstract: The combinatorial structure of many real-world action spaces leads to exponential growth in the number of possible actions, limiting the effectiveness of conventional reinforcement learning algorithms. Recent approaches for combinatorial action spaces impose factorized or sequential structures over sub-actions, failing to capture complex joint behavior. We introduce the Sub-Action Interaction Network using Transformers (SAINT), a novel policy architecture that represents multi-component actions as unordered sets and models their dependencies via self-attention conditioned on the global state. SAINT is permutation-invariant, sample-efficient, and compatible with standard policy optimization algorithms. In 18 distinct combinatorial environments across three task domains, including environments with $1.35 \times 10^{18}$ possible actions, SAINT consistently outperforms strong baselines.

Method: Proposes Kalman Filter Enhanced Group Relative Policy Optimization (KRPO) that uses lightweight Kalman filtering to dynamically estimate latent reward baselines and uncertainty, replacing the naive group mean approach of GRPO without adding learned parameters.

Result: KRPO improves performance and shows more stable return curves compared to GRPO on math question answering and reasoning tasks, as measured by accuracies and rewards.

Conclusion: Kalman filtering provides a simple yet effective way to incorporate group-level uncertainty for advantage estimation, improving policy optimization in language models with dynamic reward signals.

Abstract: The advantage function is a central concept in RL that helps reduce variance in policy gradient estimates. Recently, for language modeling, Group Relative Policy Optimization (GRPO) was proposed to compute the advantage for each output by subtracting the mean reward, as the baseline, for all outputs in the group. However, it can lead to high variance when the reward advantage is inaccurately estimated. In this work, we propose Kalman Filter Enhanced Group Relative Policy Optimization (KRPO) model, by using lightweight Kalman filtering to dynamically estimate the latent reward baseline and uncertainty. This filtering technique replaces the naive group mean, enabling more adaptive advantage normalization. Our method does not require additional learned parameters over GRPO. This approach offers a simple yet effective way to incorporate group-level uncertainty for advantage estimation, improving policy optimization in settings where highly dynamic reward signals are difficult to model for language models. Through the accuracies and rewards obtained from math question answering and reasoning, we show that using a more adaptive advantage estimation model, KRPO can improve the performance and show more stable return curves upon GRPO. The code is available at https://github.com/billhhh/KRPO_LLMs_RL.

Method: Proposes a framework with dynamic adaptation of world model alongside policy under unified learning objective using maximin optimization solved via Stackelberg learning dynamics.

Result: Demonstrates state-of-the-art performance on twelve noisy D4RL MuJoCo tasks and three stochastic Tokamak Control tasks.

Conclusion: Joint optimization of world model and policy with Stackelberg dynamics improves robustness and performance in offline model-based RL.

Abstract: Offline reinforcement learning (RL) offers a powerful paradigm for data-driven control. Compared to model-free approaches, offline model-based RL (MBRL) explicitly learns a world model from a static dataset and uses it as a surrogate simulator, improving data efficiency and enabling potential generalization beyond the dataset support. However, most existing offline MBRL methods follow a two-stage training procedure: first learning a world model by maximizing the likelihood of the observed transitions, then optimizing a policy to maximize its expected return under the learned model. This objective mismatch results in a world model that is not necessarily optimized for effective policy learning. Moreover, we observe that policies learned via offline MBRL often lack robustness during deployment, and small adversarial noise in the environment can lead to significant performance degradation. To address these, we propose a framework that dynamically adapts the world model alongside the policy under a unified learning objective aimed at improving robustness. At the core of our method is a maximin optimization problem, which we solve by innovatively utilizing Stackelberg learning dynamics. We provide theoretical analysis to support our design and introduce computationally efficient implementations. We benchmark our algorithm on twelve noisy D4RL MuJoCo tasks and three stochastic Tokamak Control tasks, demonstrating its state-of-the-art performance.

Method: Two-stage framework: 1) During training, use sharpness-aware minimization to stabilize learning of noise-sensitive concepts. 2) During inference, rank concepts by predictive entropy and correct only the most uncertain ones, using uncertainty as a proxy for susceptibility to noise.