论文检索

Yuxuan Wan, Tianqing Fang, Zaitang LI, Yintong Huo, Wenxuan Wang, Haitao Mi, Dong Yu, Michael R. Lyu

Recent advances in Deep Research Agents (DRAs) are transforming automated knowledge discovery and problem-solving.While the majority of existing efforts focus on enhancing policy capabilities via post-training, we propose an alternative paradigm: test-time self-evolving the agent’s ability by iteratively verifying the policy model’s outputs, guided by meticulously crafted rubrics. This approach gives rise to an inference-time scaling of verification, wherein an agent self-improves at test time by evaluating its generated answers to produce iterative feedback and refinements without any additional training. We derive the rubrics based on an automatically constructed DRA Failure Taxonomy, which systematically classifies agent failures into five major categories and thirteen sub-categories. We present DeepVerifier, a rubrics-based outcome reward verifier that leverages the asymmetry of verification and outperforms vanilla agent-as-judge and LLM judge baselines by 12%–48% in meta-evaluation F1 score. To enable practical test-time self-evolution, DeepVerifier integrates as a plug-and-play module during test-time inference. The verifier produces detailed rubric-based feedback, which is fed back to the agent for iterative bootstrapping—refining responses without additional training. This test-time scaling delivers 8%–11% accuracy gains on challenging subsets of GAIA and XBench-DeepResearch when powered by capable closed-source LLMs. Finally, to support open-source advancement, we release DeepVerifier-4K, a curated supervised fine-tuning dataset of 4,646 high-quality agent steps focused on DRA verification. These examples emphasize reflection and self-critique, enabling open models to develop robust verification capabilities.

Yurui Pan, Ke Xu, Bo Peng

Alignment of large language models (LLMs) typically relies on supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF), or more recently direct preference optimization (DPO). However, existing objectives largely ignore the global geometry and topology of the representation space: they operate on local token-level likelihoods or scalar preference scores, and do not explicitly constrain how hidden states move from a user prompt to an answer.We view generation as tracing a semantic trajectory in hidden space, and propose a topology-enhanced alignment framework that regularizes these trajectories using 0-dimensional persistent homology. First, at the SFT stage, we introduce a Trajectory Topology Loss (TTL). For each batch, we treat mean-pooled embeddings of prompts and gold answers as a mixed point cloud, run a Union-Find-based 0D persistent homology algorithm, and extract ”prompt–answer bridge” edges that connect previously disconnected components. TTL encourages the model’s actual update direction from prompt to answer to align with these topologically derived bridges, rather than with arbitrary or per-example directions.Second, at the RLHF/DPO stage, we propose Topological Preference Optimization (TPO). TPO constructs topic-specific semantic preference vectors from an offline pipeline and aligns the semantic improvement direction between rejected and chosen responses with these vectors in an intermediate hidden layer. We further introduce an exponential-moving-average-based dynamic weighting scheme to balance DPO and TPO losses, and also explore a fully topological variant that applies persistent homology on the chosen/rejected embedding cloud.We instantiate our methods on Qwen2.5-7B-Instruct and evaluate on UltraChat and Anthropic HH-RLHF. Across both SFT and DPO training, topology-enhanced objectives consistently outperform strong non-topological baselines (including per-example, nearest-neighbor, and random direction regularizers) on automatic preference metrics and LLM-judge evaluations, while maintaining or slightly improving toxicity. These results suggest that incorporating persistent homology and trajectory geometry is a promising and practical direction for more controllable LLM alignment.

Xidi Cai, Junhao Zheng, Jingye Li, Boyuan Li, Shaowei Zhang, Qianli Ma

Multimodal large language models (MLLMs) have achieved strong performance on challenging visual question answering benchmarks, yet their inference efficiency is severely constrained by the rapidly growing context. This growth stems from two primary sources: the large number of visual tokens required to encode images, and the accumulation of intermediate reasoning traces during autoregressive generation. To address these challenges, we propose LaT (**L**ook **a**nd **T**hink), the first modality-decoupled compression method that enables efficient multimodal inference. LaT structures reasoning into alternating looking and thinking steps, thereby explicitly signaling when visual grounding is required. Building on this design, LaT (1) evicts visual tokens whenever visual grounding is unnecessary, and (2) applies co-learning-guided compression after each completed step, mitigating the two sources of context growth respectively. Experimental results demonstrate that LaT reduces the average context length by up to 57%, while maintaining performance comparable to the standard MLLM baseline. The code will be publicly released.

Lingrui Mei, Shenghua Liu, Yiwei Wang, Baolong Bi, Yuyao Ge, Jun Wan, Yurong Wu, Xueqi Cheng

Large Language Models (LLMs) have made remarkable breakthroughs in reasoning, yet continue to struggle with hallucinations, logical errors, and inability to self-correct during complex multi-step tasks. Current approaches like chain-of-thought prompting offer limited reasoning capabilities that fail when precise step validation is required. We propose Environment Augmented Generation (EAG), a framework that enhances LLM reasoning through: (1) real-time environmental feedback validating each reasoning step, (2) dynamic branch exploration for investigating alternative solution paths when faced with errors, and (3) experience-based learning from successful reasoning trajectories. Unlike existing methods, EAG enables deliberate backtracking and strategic replanning through tight integration of execution feedback with branching exploration. Our a1-32B model achieves state-of-the-art performance among similar-sized models across all benchmarks, matching larger models like o1 on competition mathematics while outperforming comparable models by up to 24.4 percentage points. Analysis reveals EAG’s distinctive scaling pattern: initial token investment in environment interaction yields substantial long-term performance dividends, with advantages amplifying proportionally to task complexity.

Zhaolin Li, Jan Niehues

Automatic speech recognition (ASR) still covers only a small fraction of the world’s languages, mainly due to supervised data scarcity. In-context learning (ICL) with large language models (LLMs) addresses this problem, but prior work largely focuses on high-resource languages covered during training and text-only settings. This paper investigates whether speech LLMs can learn unseen languages with multimodal ICL (MICL), and how this learning can be used to improve ASR. We conduct experiments with two speech LLMs, Phi-4 and Qwen3-Omni, on three diverse endangered languages. Firstly, we find that MICL is effective for unseen languages, leveraging both speech and text modalities. We further show that cross-lingual transfer learning improves MICL efficiency on target languages without training on them. Moreover, we analyze attention patterns to interpret MICL mechanisms, and we observe layer-dependent preferences between audio and text context, with an overall bias towards text. Finally, we show that prompt-based ASR with speech LLMs performs poorly on unseen languages, motivating a simple ASR system that combines a stronger acoustic model with a speech LLM via MICL-based selection of acoustic hypotheses. Results show that MICL consistently improves ASR performance, and that cross-lingual transfer learning matches or outperforms corpus-trained language models without using target-language data. Our code is publicly available

Xingle Xu, Fanheng Kong, Dexian Cai, Shi Feng, Xiaocui Yang, Daling Wang, Yifei Zhang

Multimodal Large Language Models (MLLMs) integrate visual encoders with Large Language Models (LLMs) and enable multimodal reasoning. However, for tasks that heavily rely on visual information, the model’s utilization of visual information remains unstable, which leads to reasoning failures. Prior works mainly strengthen multimodal reasoning by improving representation alignment or increasing computation. However, these methods do not explicitly characterize the differences in visual demands across tasks, making it difficult for the model to decide where and how strongly to attend to visual information. Consequently, visual attention allocation becomes a key factor that affects multimodal reasoning. To address these, we propose RATION, an entropy-driven task-adaptive visual attention allocation framework. First, we use a task routing strategy to infer the task type of each sample and identify the key layers. We use visual attention entropy as a control signal to dynamically allocate attention according to task demands. Experiments show that RATION achieves consistent performance gains across diverse reasoning tasks, datasets, and models, providing a clear direction toward more reliable multimodal reasoning.

Taewook Hwang, Inbum Heo, Sung Jun Lee, Sangkeun Jung

Vision-Language Models (VLMs) have demonstrated remarkable performance in document understanding tasks; however, VLMs also suffer from hallucinations inherited from LLMs. While prior work has focused on reasoning-stage hallucinations, the role of visual perception remains underexplored. In this work, we define perceptual hallucination as the phenomenon where VLMs generate information as if perceived, despite absent or damaged visual evidence. To analyze this, we construct DocHallu, a benchmark of 2,671 original–damaged image pairs across three tasks, available at https://huggingface.co/datasets/IB99/DocHallu. Experiments reveal that perceptual hallucination occurs across all models, with higher rates for numerical content than textual content. Activation patching analysis suggests that hallucinations are strongly associated with errors introduced in the vision encoder, which can subsequently propagate and become amplified through the text decoding process. We also demonstrate that LLM-based post-hoc filtering can reduce hallucination exposure by 36% on average, with reductions of up to 88%. This work extends VLM hallucination research by defining, analyzing, and verifying perceptual hallucination in document understanding.

Nikolay Banar, Ehsan Lotfi, Jens Van Nooten, Cristina Arhiliuc, Marija Kliocaite, Walter Daelemans

Recently, embedding resources, including models, benchmarks, and datasets, have been widely released to support a variety of languages. However, the Dutch language remains underrepresented, typically comprising only a small fraction of the published multilingual resources. To address this gap and encourage the further development of Dutch embeddings, we introduce new resources for their evaluation and generation. First, we introduce the Massive Text Embedding Benchmark for Dutch (MTEB-NL), which includes both existing Dutch datasets and newly created ones, covering a wide range of tasks. Second, we provide a training dataset compiled from available Dutch retrieval datasets, complemented with synthetic data generated by large language models to expand task coverage beyond retrieval. Finally, we release a series of E5-NL compact yet efficient embedding models that demonstrate strong performance across multiple tasks. We make our resources publicly available through the Hugging Face Hub and the MTEB package.

Lingyi Kong, Zhuo Liu, Zhanghao Hu, Qilong Qiu, Yutao Yang, Jingjing Xue, Zheng Wang, Lin Gui, Feiping Nie

Adversarial attacks have attracted growing attention across domains, including natural language processing (NLP). Character-level adversarial attacks preserve semantics, but they have received less attention because the discrete operations they use are costly and inefficient. Challenging these beliefs, we introduce two adaptively learnable matrices that transform discrete choices into continuous representations, enabling automatic one-shot multi-position, multi-character insertion. To optimize the two learnable matrices, we propose OSCR-Attack, an end-to-end framework based on gradient-based optimization, with a conflict resolution strategy that maps the optimized continuous distributions back into discrete insertion operations. Extensive experiments on three benchmarks with three open-source large language models (LLMs) show that OSCR-Attack improves attack success rate (ASR) by up to 21.45% points and accelerates the attack by up to 3.66 times compared to recent baselines.

Yangryeol Park, Kunhui Lee, Hanback Choi, Cheoneum Park, Donghyeon Jeon, Inho Kang, Seung-Hoon Na

Masked diffusion language models (MDLMs) enable efficient parallel decoding but are limited by a monotonic unmasking policy, where committed tokens cannot be revised. While remasking-based methods mitigate early errors, they mainly intervene during generation. In this work, we study post-hoc refinement of a completed draft and find that naive correction often fails because of contextual lock-in, a phenomenon in which local error patterns become self-reinforcing. To address this, we propose PURE (Post-hoc Unlocking and REfinement), a training-free inference algorithm for two-phase decoding. PURE profiles confidence dynamics during drafting to identify unstable regions via an instability score (\Delta_i), then unlocks them through deterministic window masking and stochastic leftward relaxation. On reasoning benchmarks, PURE substantially improves accuracy when applied to LLaDA-8B-Instruct, including a gain of +12.9 points over the baseline on GSM8K. These gains require only a small refinement budget, yielding a favorable compute-quality trade-off for discrete diffusion decoding.

Yifan Song, Xingjian Tao, Zhicheng Yang, Yihong Luo, Jing Tang

Graph-based Retrieval-Augmented Generation (GraphRAG) enhances LLMs by structuring corpus into graphs to facilitate multi-hop reasoning. While recent lightweight approaches reduce indexing costs by leveraging Named Entity Recognition (NER), they rely strictly on structural co-occurrence, failing to capture latent semantic connections between disjoint entities. To address this, we propose EHRAG, a lightweight RAG framework that constructs a hypergraph capturing both structure and semantic level relationships, employing a hybrid structural-semantic retrieval mechanism. Specifically, EHRAG constructs structural hyperedges based on sentence-level co-occurrence with lightweight entity extraction and semantic hyperedges by clustering entity text embeddings, ensuring the hypergraph encompasses both structural and semantic information. For retrieval, EHRAG performs a structure-semantic hybrid diffusion with topic-aware scoring and personalized pagerank (PPR) refinement to identify the top-k relevant documents. Experiments on four datasets show that EHRAG outperforms state-of-the-art baselines while maintaining linear indexing complexity and zero token consumption for construction. Code is available at https://github.com/yfsong00/EHRAG.

Chen Qian, Yimeng Wang, Yu Chen, Lingfei Wu, Andreas Stathopoulos

Explainable AI (XAI) in high-stakes domains should help stakeholders trust and verify system outputs. Yet Chain-of-Thought methods reason before concluding, and logical gaps or hallucinations can yield conclusions that do not reliably align with their rationale. Thus, we propose “Result → Justify”, which constrains the output communication to present a conclusion before its structured justification. We introduce SEF (Structured Explainability Framework), operationalizing professional conventions (e.g., CREAC, BLUF) via six metrics for structure and grounding. Experiments across four tasks in three domains validate this approach: all six metrics correlate with correctness (r=0.20–0.42; p<0.001), and SEF achieves 83.9% accuracy (+5.3 over CoT). These results suggest structured justification can improve verifiability and may also improve reliability. Code is available at https://github.com/cqian03/SEF.

Kang Liu, YongKang Liu, Xiaocui Yang, Peidong Wang, Wen Zhang, Shi Feng, Yifei Zhang, Daling Wang

Large Reasoning Models (LRMs) often suffer from overthinking, a phenomenon in which redundant reasoning steps are generated after a correct solution has already been reached. Existing early reasoning exit methods primarily rely on output-level heuristics or trained probing models to skip redundant reasoning steps, thereby mitigating overthinking. However, these approaches typically require additional rollout computation or externally labeled datasets. In this paper, we propose NEAT, a Neuron-based Early reAsoning exiT framework that monitors neuron-level activation dynamics to enable training-free early exits, without introducing any additional test-time computation. NEAT identifies exit-associated neurons and tracks their activation patterns during reasoning to dynamically trigger early exit or suppress reflection, thereby reducing unnecessary reasoning while preserving solution quality. Experiments on four reasoning benchmarks across six models with different scales and architectures show that, for each model, NEAT achieves an average token reduction of 22% to 28% when averaged over the four benchmarks, while maintaining accuracy.

Yijie Lu, Manman Zhao, Tianjie Ju, Zihe Yan, Xinbei Ma, Yuan Guo, Daizong Ding, Gongshen Liu, Zhuosheng Zhang

Autonomous GUI agents are inherently vulnerable to Environmental Injection Attacks (EIAs). However, existing red-teaming methods face a trade-off between requiring target-specific knowledge and incurring prohibitive computational costs. More fundamentally, a key question remains: \textit{what factors determine attack success?} To answer this, we first analyze two dimensions: visual appearance (e.g., position, size, color) and semantic content. We find that semantic content dominates, while visual variations have negligible impact. Leveraging this insight, we introduce EVA, a framework that evolves payloads exclusively on the semantic dimension via a discovery-deployment pipeline. Experiments demonstrate that EVA significantly outperforms baselines, achieving 59% to 85% average Attack Success Rate (ASR) while evolving benign seeds into successful attacks within 1.18 to 1.71 iterations. This rapid convergence suggests a dense semantic attack space within the model’s latent space. Whenever an input falls into this space, the agent becomes inherently vulnerable, exposing a fundamental alignment flaw in current multimodal representations.

Weiyang Guo, Zesheng Shi, Liye Zhao, Jiayuan Ma, Zeen Zhu, Junxian He, Min Zhang, Jing Li

While Large Language Models (LLMs) have demonstrated significant potential in Tool-Integrated Reasoning (TIR), existing training paradigms face significant limitations: Zero-RL suffers from inefficient exploration and mode degradation due to a lack of prior guidance, while SFT-then-RL is limited by high data costs and capability plateaus caused by low-entropy collapse. To address these challenges, we propose E3-TIR (Enhanced Experience Exploitation), a warm-up paradigm for the early stages of agent training. Specifically, we formulate training as the dynamic integration of three experience types: Expert Prefixes, Expert Guided, and Self-Exploration. By executing diverse branching exploration around expert “anchors’’ and employing a mix policy optimization mechanism, we effectively mitigate distribution shifts and resolve optimization conflicts arising from shared prefixes. Our method dynamically adapts the model’s knowledge boundaries, effectively balancing exploration diversity with training efficiency. Experimental results demonstrate that E3-TIR achieves a 6% performance improvement over traditional paradigms on tool-use tasks, while requiring less than 10% of the synthetic data. Furthermore, in terms of ROI—a comprehensive metric integrating performance, data cost, and training efficiency—we achieve a 1.46 gain compared to baselines.

Yibo Yan, Shen Wang, Jiahao Huo, Jingheng Ye, Zhendong Chu, Xuming Hu, Philip S. Yu, Carla P Gomes, Bart Selman, Qingsong Wen

Scientific reasoning, the process through which humans apply logic, evidence, and critical thinking to explore and interpret scientific phenomena, is essential in advancing knowledge reasoning across diverse fields. However, despite significant progress, current scientific reasoning models still struggle with generalization across domains and often fall short of multimodal perception. Multimodal Large Language Models (MLLMs), which integrate text, images, and other modalities, present an exciting opportunity to overcome these limitations and enhance scientific reasoning. Therefore, **this position paper argues that MLLMs can significantly advance scientific reasoning across disciplines such as mathematics, physics, chemistry, and biology**. We highlight the current state of MLLM applications in scientific reasoning, noting their ability to integrate and reason over diverse data types. However, challenges such as multimodal alignment, data diversity, and reasoning depth remain obstacles to achieving their full potential. To address these challenges, we propose actionable suggestions in the near future. Overall, our work offers a novel perspective on MLLM integration with scientific reasoning, providing the LLM community with valuable insights for achieving Artificial General Intelligence (AGI).

Zhigan Zhang, Jie Sui

The rapid spread of fake news on digital platforms presents significant societal challenges, demanding detection methods capable of addressing intricate manipulation strategies. Existing methods predominantly rely on monolithic verification approaches, which fail to decompose the complex blend of factual inaccuracies, logical fallacies, and propaganda techniques in modern misinformation. To address this gap, we propose PARD, a Protocol-Adaptive Role-Specific multi-agent framework that decomposes verification into factual, logical, and contextual dimensions. PARD dynamically selects the optimal interaction protocol from a library that includes Round-Robin Discussion, Point-Counterpoint Dialogue, and Cross-Examination to tailor the reasoning process for more effective detection. We evaluate PARD with LIAR-RAW+, an extended version of the LIAR-RAW dataset enriched with fine-grained factual, logical, and contextual annotations. Experimental results demonstrate that PARD consistently outperforms baseline methods in both predictive accuracy and explanatory quality, supported by its efficient dynamic governance mechanism

Xuan Feng, Bo An, Tianlong Gu, Liang Chang, Fengrui Hao, Peipeng Yu, Shuai Zhao

Bias in Large Language Models (LLMs) poses significant risks to trustworthiness, manifesting primarily as stereotypical biases (e.g., gender or racial stereotypes) and structural biases (e.g., lexical overlap or position preferences). However, prior paradigms typically address these in isolation, often mitigating one at the expense of exacerbating the other. To address this, we conduct a systematic exploration of these reasoning failures and identify a primary inducement: the latent spurious feature correlations within the input that drive these erroneous reasoning shortcuts. Driven by these findings, we introduce Causal-Contrastive Preference Optimization (C2PO), a unified alignment framework designed to tackle these specific failures by simultaneously discovering and suppressing these correlations directly within the optimization process. Specifically, C2PO leverages causal counterfactual signals to isolate bias-inducing features from valid reasoning paths, and employs a fairness-sensitive preference update mechanism to dynamically evaluate logit-level contributions and suppress shortcut features. Extensive experiments across multiple benchmarks covering stereotypical bias (BBQ, Unqover), structural bias (MNLI, HANS, Chatbot, MT-Bench), out-of-domain fairness (StereoSet, WinoBias), and general utility (MMLU, GSM8K) demonstrate that C2PO effectively mitigates stereotypical and structural biases while preserving robust general reasoning capabilities.

Xingle Xu, YongKang Liu, Dexian Cai, Shi Feng, Xiaocui Yang, Daling Wang, Yifei Zhang

Multimodal Sentiment Analysis aims to integrate information from various modalities to make complementary predictions. However, it often struggles with irrelevant or misleading visual and auditory information. Most existing approaches treat entire modality as an independent unit for feature enhancement or denoising, which often suppresses redundant noise at the cost of weakening critical information. To address this challenge, we propose MoLAN, a unified ModaLity-aware noise dynAmic editiNg framework. Specifically, MoLAN performs modality-aware block partitioning by dividing the features of each modality into multiple blocks. Each block is then dynamically assigned a distinct denoising strength based on its noise level and semantic relevance, enabling fine-grained noise suppression while preserving essential multimodal information. Notably, MoLAN is a unified and flexible framework that can be seamlessly integrated into a wide range of multimodal models. Building upon this framework, we further introduce MoLAN+, a new multimodal sentiment analysis approach. Experiments across five models and four datasets demonstrate the broad effectiveness of the MoLAN framework. Extensive evaluations show that MoLAN+ achieves the state-of-the-art performance.

Xiangning Yu, Yuwei Guo, Yuqi Hou, Xiao Xue, Qun Ma

LLM-empowered agent simulations are increasingly used to study social emergence, yet the micro-to-macro causal mechanisms behind macro outcomes often remain unclear. This is challenging because emergence arises from intertwined agent interactions and meso-level feedback and nonlinearity, making generative mechanisms hard to disentangle. To this end, we introduce CAMO, an automated Causal discovery framework from Micro behaviors to Macro Emergence in LLM agent simulations. CAMO converts mechanistic hypotheses into computable factors grounded in simulation records and learns a compact causal representation centered on an emergent target . CAMO outputs a computable Markov boundary and a minimal upstream explanatory subgraph, yielding interpretable causal chains and actionable intervention levers. It also uses simulator-internal counterfactual probing to orient ambiguous edges and revise hypotheses when evidence contradicts the current view. Experiments across four emergent settings demonstrate the promise of CAMO.[The code is available at an anonymous link: <https://anonymous.4open.science/r/CAMO-0E6C/>.]