论文检索

Xin Yin, Zixiang Ding, Yiang Zhang, Qiang Wang, Rui Wang, Chao Ni, Zhe Cui

Large Language Models (LLMs) have achieved strong performance on many code-related tasks, yet they still struggle with repository-level scenarios where reasoning depends on long, noisy, and structurally complex contexts. While existing retrieval methods, including both similarity-based and graph-based approaches, can identify relevant code snippets, they often retrieve excessive contexts that intensify the "lost-in-the-middle" phenomenon and dilute model attention with redundant contexts. To address this, we present RepoDistill, a novel framework that integrates retrieval with learned budget allocation for fine-grained context compression. RepoDistill first employs a plug-and-play lightweight GraphRAG to retrieve context that follows logical flows. It then applies Compression-Aware Budget Allocation guided by Compression-Aware Policy Optimization, which formulates context management as a multi-step decision problem and learns allocation policies for contexts. Experiments show that RepoDistill outperforms baselines, achieving gains of up to +7.00 on SWE-QA, +24.4% on CoderEval, and +0.25 on LongCodeU. Furthermore, a compact 4B-parameter model trained with RepoDistill can serve as an effective context compressor for closed-source LLMs, reducing input tokens by up to 66% while maintaining comparable performance. We release our code at https://anonymous.4open.science/r/RepoDistill-12B0.

Zhipeng Qian, Zihan Liang, Yufei Ma, Ben Chen, Huangyu Dai, Yiwei Ma, Jiayi Ji, Chenyi Lei, Han Li, Xiaoshuai Sun

Composed Image Retrieval (CIR) enables users to search for target images using both a reference image and manipulation text, offering substantial advantages over single-modality retrieval systems. However, existing CIR methods suffer from representation space fragmentation: queries and targets comprise heterogeneous modalities and are processed by distinct encoders, forcing models to bridge misaligned representation spaces only through post-hoc alignment, which fundamentally limits retrieval performance. As evidenced by t-SNE visualization in Fig.(a), this architectural asymmetry manifests as three distinct, well-separated clusters in the feature space, directly demonstrating how heterogeneous modalities create fundamentally misaligned representation spaces from initialization. In this work, we propose CSMCIR, a unified representation framework that achieves efficient query-target alignment through three synergistic components. First, we introduce a Multi-level Chain-of-Thought (MCoT) prompting strategy that guides Multimodal Large Language Models to generate discriminative, semantically compatible captions for target images, establishing modal symmetry. Building upon this, we design a symmetric dual-tower architecture where both query and target sides utilize the identical shared-parameter Q-Former for cross-modal encoding, ensuring consistent feature representations and further reducing the alignment gap. Finally, this architectural symmetry enables an entropy-based, temporally dynamic Memory Bank strategy that provides high-quality negative samples while maintaining consistency with the evolving model state. Extensive experiments on four benchmark datasets demonstrate that our CSMCIR achieves state-of-the-art performance with superior training efficiency. Comprehensive ablation studies further validate the effectiveness of each proposed component.

Xingjian Diao, Zheyuan Liu, Chunhui Zhang, Weiyi Wu, Keyi Kong, Lin Shi, Kaize Ding, Soroush Vosoughi, Jiang Gui

Large Vision-Language Models (LVLMs) have exhibited strong reasoning capabilities through chain-of-thought mechanisms that generate step-by-step rationales. However, such slow-thinking approaches often lead to overthinking, where models produce excessively verbose responses even for simple queries, resulting in test-time inefficiency and even degraded accuracy. Prior work has attempted to mitigate this issue via adaptive reasoning strategies, but these methods largely overlook a fundamental bottleneck: visual perception failures. We argue that stable reasoning critically depends on low-level visual grounding, and that reasoning errors often originate from imperfect perception rather than insufficient deliberation. To address this limitation, we propose Gated Perception-Reasoning Optimization (GPRO), a meta-reasoning controller that dynamically routes computation among three decision paths at each generation step: a lightweight fast path, a slow perception path for re-examining visual inputs, and a slow reasoning path for internal self-reflection. To learn this distinction, we derive large-scale failure attribution supervision from approximately 790k samples, using teacher models to distinguish perceptual hallucinations from reasoning errors. We then train the controller with multi-objective reinforcement learning to optimize the trade-off between task accuracy and computational cost under uncertainty. Experiments on five benchmarks demonstrate that GPRO substantially improves both accuracy and efficiency, outperforming recent slow-thinking methods while generating significantly shorter responses.

Xinzhe Li

Test-time scaling improves large language models (LLMs) on long-horizon reasoning tasks by allocating more compute at inference. LLM inference via tree search (LITS) achieves strong performance but is highly inefficient. We propose Chain-in-Tree (CiT), a plug-in framework that decides when to branch during search instead of expanding at every step. CiT introduces lightweight Branching Necessity (BN) evaluations, including BN-DP (direct prompting) and BN-SC (self-consistency). Integrated into Tree of Thoughts, ReST-MCTS, and RAP, BN-DP reduces token generation, model calls, and runtime by 75-85% on GSM8K and Math500, with often negligible or no accuracy loss. BN-SC typically yields substantial savings (up to 80%) generally but shows instability in 1-4 out of 14 settings, caused by a small subset of examples that produce extremely long reasoning steps. We theoretically prove that BN-DP never increases policy invocations and release unified implementations applicable across LITS frameworks. The full codebase is publicly available at https://github.com/xinzhel/chain_in_tree.

Irina Tolstykh, Aleksandra Tsybina, Sergey Yakubson, Aleksandr Gordeev, Vladimir Dokholyan, Maksim Kuprashevich

With the increasing quality and spread of LLM assistants, the amount of generated content is growing rapidly. In many cases and tasks, such texts are already indistinguishable from those written by humans, and the quality of generation continues to increase. At the same time, detection methods are advancing more slowly than generation models, making it challenging to prevent misuse of generative AI technologies. We propose GigaCheck, a dual-strategy framework for AI-generated text detection. At the document level, we leverage the representation learning of fine-tuned LLMs to discern authorship with high data efficiency. At the span level, we introduce a novel structural adaptation that treats generated text segments as "objects." By integrating a DETR-like vision model with linguistic encoders, we achieve precise localization of AI intervals, effectively transferring the robustness of visual object detection to the textual domain. Experimental results across three classification and three localization benchmarks confirm the robustness of our approach. The shared fine-tuned backbone delivers strong accuracy in both scenarios, highlighting the generalization power of the learned embeddings. Moreover, we successfully demonstrate that visual detection architectures like DETR are not limited to pixel space, effectively generalizing to the localization of generated text spans. To ensure reproducibility and foster further research, we publicly release our source code.

Yuyan Zhou, Weiyu Chen, James Kwok

Diffusion-based Large Language Models (dLLMs) are emerging as a powerful alternative to traditional autoregressive models. These models learn to generate text by iteratively denoising masked sequences. In this work, we identify a critical problem in dLLMs: the model’s attention is wastefully expended on uninformative mask tokens, diluting its focus on meaningful context. We term this phenomenon “attention dilution”. We further show that this artifact is amplified by token-level noising, whereas models employing sequence-level noise exhibit a reduced effect. To resolve this problem, we introduce Truncated Block Generation, a novel sampling algorithm that not only mitigates attention dilution but also enables faster inference and flexible-length sequence generation. Extensive experiments validate our analysis and demonstrate the marked effectiveness of our proposed method in enhancing both the performance and efficiency of dLLMs.

Wenpeng Xing, Moran Fang, Guangtai Wang, Changting Lin, Meng Han

While Large Language Models (LLMs) have achieved remarkable performance, they remain vulnerable to jailbreak attacks that circumvent safety constraints. Existing strategies, ranging from heuristic prompt engineering to computationally intensive optimization, often face significant trade-offs between effectiveness and efficiency. In this work, we propose Contextual Representation Ablation (CRA), a novel inference-time intervention framework designed to dynamically silence model guardrails. Predicated on the geometric insight that refusal behaviors are mediated by specific low-rank subspaces within the model’s hidden states, CRA identifies and suppresses these refusal-inducing activation patterns during decoding without requiring expensive parameter updates or training. Empirical evaluation across multiple safety-aligned open-source LLMs demonstrates that CRA significantly outperforms baselines. By revealing that safety constraints can be surgically ablated from internal representations, our findings expose the intrinsic fragility of current alignment mechanisms and underscore the urgent need for more robust latent-space defenses.

Yi Chen, Yuying Ge, Rui Wang, Yixiao Ge, Junhao Cheng, Ying Shan, Xihui Liu

Recent reinforcement learning (RL) approaches, such as outcome-supervised GRPO, have advanced reasoning in Large Language Models (LLMs), yet their adaptation to multimodal LLMs (MLLMs) remains underexplored. Progress has been further limited by the lack of evaluation settings that jointly test perception and reasoning under controlled generalization challenges. To enable such analysis, we present **SEED-Bench-R1**, a structured testbed featuring real-world video tasks and hierarchical evaluation across in-distribution, cross-environment, and cross-environment-task scenarios. Our analysis reveals that standard outcome-supervised GRPO often yields "logical incoherence"—achieving correct answers through flawed reasoning—due to its exclusive focus on final-answer rewards and rigid KL penalties. To address this, we propose **GRPO-CARE**, a consistency-aware RL framework that eliminates KL penalties while introducing a two-tiered reward system: a base reward for accuracy and an adaptive bonus for consistency. This bonus, derived from a slowly evolving reference model through group-relative likelihood calibration, rewards reasoning paths that logically support the final answer without requiring expensive process supervision. Experiments on SEED-Bench-R1 show that GRPO-CARE consistently outperforms standard GRPO, achieving a 6.7% gain on the hardest evaluation level and a 24.5% increase in reasoning consistency. Moreover, models trained with GRPO-CARE transfer effectively to diverse video understanding and even language-only reasoning benchmarks, validating its robustness and generality.

Hui Huang, Julien Velcin, Yacine Kessaci

Long-document question answering is challenging because relevant evidence is often scattered across distant sections. Traditional long-document QA/RAG pipelines often suffer from context fragmentation, retrieving locally plausible but structurally misaligned passages. We present the Hierarchical Quantized Document Retriever (HQDR), a framework that aligns hierarchical graph representations with a universal token vocabulary and integrates explicit structure into retrieval. By grounding continuous structural features in a fixed, discrete semantic space, HQDR captures universal hierarchical patterns rather than overfitting to specific layouts. We further propose a hybrid scoring mechanism that decouples semantic matching from structural alignment. Extensive experiments on QASPER and Natural Questions demonstrate that HQDR achieves consistent gains over strong baselines and exhibits superior robustness when transferring between datasets with distinct structural characteristics.

Weiyu Ma, Yuqian Fu, Zecheng Zhang, Bernard Ghanem, Guohao Li

We introduce AVACraft — the first multimodal benchmark environment for complex decision-making in StarCraft II, supporting both traditional Multi-Agent Reinforcement Learning (MARL) and modern Vision-Language Model (VLM) paradigms. Existing StarCraft II environments like SMAC rely on abstract state representations that deviate from human perception and lack support for emerging VLM-based decision-making. AVACraft mitigates these limitations via a unified framework, which provides RGB visual inputs, natural language observations and structured state information, enabling systematic comparisons between training-based and zero-shot decision-making methods. Our benchmark features 21 carefully designed scenarios covering micromanagement, coordination and strategic planning, with standardized evaluation protocols for both paradigms. We establish comprehensive baselines using four MARL algorithms (IQL, QMIX, QTRAN, VDN) and multiple state-of-the-art VLMs (GPT-4o, Qwen-VL, etc.). Experimental results reveal their complementary strengths: MARL methods achieve up to 27.1% win rate after 1M training steps in complex scenarios, while VLMs deliver superior zero-shot performance (75–81% win rate) and human-aligned decision processes without any training. Systematic analysis (including expert human evaluation) also identifies key trade-offs between training efficiency, performance ceilings and interpretability across the two paradigms. Our implementation is available at https://anonymous.4open.science/r/VLM-Play-StarCraft2-70C4 .

Rui Sun, Jie Ding, Chenghua Gong, Tianjun Gu, Yihang Jiang, Juyuan Zhang, Liming Pan, Linyuan Lü

Optimizing communication topology in LLM–based multi-agent system is critical for enabling collective intelligence. Existing methods mainly rely on spatio-temporal interaction paradigms, where the sequential execution of multi-round dialogues incurs high latency and computation. Motivated by the recent insights that evaluation and debate mechanisms can improve problem-solving in multi-agent systems, we propose TopoDIM, a framework for one-shot Topology generation with Diverse Interaction Modes. Designed for decentralized execution to enhance adaptability and privacy, TopoDIM enables agents to autonomously construct heterogeneous communication without iterative coordination, achieving token efficiency and improved task performance. Experiments demonstrate that TopoDIM reduces total token consumption by 46.41% while improving average performance by 1.50% over state-of-the-art methods. Moreover, the framework exhibits strong adaptability in organizing communication among heterogeneous agents. Code is available at: https://github.com/Sundiasy/TopoDIM.

Xingjin Wang, Howe Tissue, Lu Wang, Linjing Li, Daniel Dajun Zeng

We observe that entropy in reinforcement learning functions analogously to the learning rate in LLMs. Maintaining stable entropy, as demonstrated in DAPO, helps stabilize RL training, while rapid entropy annealing (i.e., so-called entropy collapse) accelerates local performance improvement and enables faster convergence. We argue that these two processes are not antithetical, but can be effectively controlled and scheduled within a single training run, similar to learning rate scheduling. We propose Entropy Schduling (ES), which optimizes different pre-set goals (e.g. k in optimizing Pass@k) by controlling and scheduling entropy at each step of the RL process. We find that maintaining stable entropy early in training followed by entropy annealing achieves superior performance. Moreover, since stable-state entropy and annealed entropy exhibit distinctly different learning dynamics, curriculum learning can be seamlessly integrated to maximize model performance based on different entropy phases. We show that entropy scheduling is straightforward to implement and intuitive in design. Extensive experiments suggest that it delivers consistent and stable performance improvements across diverse models and algorithms.

Jiaxin Wu, Jie Zhao, Huangxinrong, Cuihong Zhang, Xiuzhu Wu

Whether the personality of LLMs can be intentionally reshaped remains controversial. Existing studies often limited to small models, argue for its immutability. Crucially, prior studies fail to uncover that different LLMs exhibit significant compliance divergence when exposed to personality-inducing contexts. To bridge this gap, we introduce Personality Induction Framework (PIF), which systematically reshapes the personality of different LLMs via multi-agent collaboration. Specifically, via Generator-Judge agents, PIF paraphrases MBTI questions to create semantically equivalent but expressively diverse inducing contexts, enabling LLMs to learn personality patterns instead of superficial token matching. Also, PIF achieves fine-grained personality modulation by controlling the intensity of inducing contexts. Extensive experiments on worldwide mainstream LLMs show that PIF reliably transforms their original personalities into desired target personalities. Notably, we find that the outputs of most Western LLMs behave like “Chameleons”, exhibiting high personality plasticity; whereas the outputs of most Eastern LLMs act as “Guardians”, manifesting pronounced cognitive resistance. Strikingly, extreme induction intensity (100%) triggers a counter-intuitive “Alignment Rebound” in Guardians, resulting in the opposite direction rather than compliance. These findings suggest that LLM personality is a dynamic equilibrium shaped by the trade-off between instruction compliance and cognitive resistance.

Xinming Wang, Jian Xu, Bin Yu, Sheng Lian, yi Chen, Boran Wang, Yingjian Zhu, Hongzhu Yi, Hong-Ming Yang, Han Hu 等

Large reasoning models (LRMs) show strong capabilities in complex reasoning, yet their marginal gains on evidence-dependent factual questions are limited. We find this limitation is partially attributable to a reasoning–answer hit gap, where the model identifies the correct facts during reasoning but fails to incorporate them into the final response, thereby reducing factual fidelity. To address this issue, we propose MR-ALIGN, a Meta-Reasoning informed alignment framework that enhances factuality without relying on external verifiers. MR-ALIGN quantifies state-transition probabilities along the model’s thinking process and constructs a transition-aware implicit reward that reinforces beneficial reasoning patterns while suppressing defective ones at the atomic thinking segments. This re-weighting reshapes token-level signals into probability-aware segment scores, encouraging coherent reasoning trajectories that are more conducive to factual correctness. Empirical evaluations across four factual QA datasets and one long-form factuality benchmark show that MR-ALIGN consistently improves accuracy and truthfulness while reducing misleading reasoning. These results highlight that aligning the reasoning process itself, rather than merely the outputs, is pivotal for advancing factuality in LRMs.

Guido Ivetta, Pietro Palombini, Sofía Martinelli, Marcos J Gomez, M Emilia Echeveste, Sunipa Dev, Vinodkumar Prabhakaran, Luciana Benotti

The evaluation of societal biases in NLP models is critically hindered by a geo-cultural gap. This leaves regions such as Latin America severely underserved, making it impossible to adequately assess or mitigate the perpetuation of harmful regional stereotypes in language technologies. This paper presents LACES, a stereotype association dataset, for 15 Latin American countries. This dataset includes 4,789 stereotype associations[The de-identified dataset can be accessed via GitHub], manually created and annotated by 83 participants. The dataset was developed through targeted community partnerships across Latin America. Additionally, in this paper, we propose a novel adaptive data collection methodology that uniquely integrates the sourcing of new stereotype entries and the validation of existing data within a single, unified workflow. This approach results in a resource with more unique stereotypes than previous static collection methods, enabling a more efficient stereotype collection. The paper further supports the quality of LACES by demonstrating reduced efficacy of debiasing methods on this dataset in comparison to existing popular stereotype benchmarks.Content Warning: This research involves the study of social biases. Consequently, the paper contains examples of discriminatory language and stereotypes that may be sensitive or upsetting to readers. These examples are included for the purpose of scientific analysis and do not reflect the views of the authors.

Kaiser Sun, Fan Bai, Mark Dredze

Large language models (LLMs) rely on both contextual knowledge and parametric memory, yet these sources can conflict. Prior analysis largely focused on contextual question answering, suggesting that models tend to favor parametric knowledge under conflict, but this setting assumes that tasks should always rely on the provided passage. It therefore remains unclear how LLMs behave when tasks demand different kinds and degrees of knowledge utilization. We address this gap with a model-agnostic diagnostic framework that holds underlying knowledge constant while injecting controlled conflicts across tasks with varying knowledge requirements. Evaluating representative open-source LLMs, we find that: (1) performance degradation under conflict correlates with a task’s knowledge reliance rather than conflict plausibility alone; (2) strategies such as explanatory rationales or reiteration increase context reliance, helping context-only tasks but harming those that require parametric knowledge; and (3) these behaviors bias model-based evaluation, raising concerns about the reliability of LLMs as judges. Together, our findings show that context–memory conflict is fundamentally task-dependent and motivate task-aware approaches to balancing context and memory in LLM deployment and evaluation.

He Wang, Yu Gu, Fangfang Li, Zhigang Wang, Zhenghao Liu, Ning Wang, Xiaohua Li, Ge Yu

The autoregressive inference in large language models requires repeated computation across transformer layers. While caching intermediate key-value (KV) pairs eliminates redundancy, it introduces severe memory overhead, particularly in long-context settings. Most existing cache compression methods operate solely on either quantization or eviction, based on importance estimation of cached data. However, they are limited by coarse compression choices and inaccurate importance assessment, leading to suboptimal inference quality. To address this, we propose HqeKV, a hybrid compression framework built on both quantization and eviction, offering finer-grained compression options that adapt smoothly to the varying importance of cached KV pairs. An integrated optimizer automatically selects the best compression action for each cached element, maximizing quality while insulating end-users from tedious low-level tuning details. We further design a joint K–V importance metric to provide more accurate importance assessment results so that the optimizer can make smarter decisions. Additionally, HqeKV supports flexible conversion policies across multiple quantization precision levels, to further reduce quality degradation. Extensive experiments show that HqeKV improves output quality under the same memory constraints, outperforming state-of-the-art alternatives. Code is available at https://github.com/skywclouds/HqeKV.

Tongxi Wang

Large language models (LLMs) excel across many tasks, yet inference is still dominated by strictly token-by-token autoregression. Existing acceleration methods largely patch this pipeline and miss core human-reading ingredients: content-adaptive foresight, chunk-structure-aware compute allocation, and train–test consistency for preview/skimming. We propose the Fovea–Block–Skip Transformer (FBS), which injects a causal, trainable loop into Transformers via Parafovea-Attention Window (PAW), Chunk-Head (CH), and Skip-Gate (SG). Across diverse benchmarks, FBS improves the quality-efficiency trade-off without increasing parameters, and ablations show the three modules are complementary.

Yangzhuo Li, Shengpeng Ji, Yifu Chen, Tianle Liang, Haoyu Yang, Junboli, Jun Fang, Lin Li, Qingyang Hong

Integrating explicit Chain-of-Thought (CoT) into end-to-end spoken dialogue models enhances intelligence but incurs prohibitive latency. While the "Thinking-while-Talking" paradigm alleviates this delay, it fundamentally compromises block atomicity, severing the logical connection between interleaved thought and speech. To address this, we present Dual-Reasoner, employing a Streaming Masking Mechanism underpinned by our Dual-Think-30k dataset to guarantee uninterrupted audio streaming. Crucially, to strictly align the fragmented thinking blocks to service speech generation, we introduce the Atomic-Consistency Restoration framework. To secure comprehensive capabilities in high-difficulty reasoning, this mechanism utilizes a quadruple-constraint system to reconstruct logical atomicity, ensuring that "think" chunks act as a rigorous anchor for "talk" outputs. Experimental results demonstrate that Dual-Reasoner achieves comprehensive reasoning enhancements within ultra-low latency constraints: it elevates the VoiceBench score from 67.24 to 73.41 over the baseline, while significantly reducing the Time-to-First-Audio (TTFA) from 20.35s to 3.65s and the Real-Time Factor (RTF) from 7.04 to 1.05.

Jinyang Wu, Chonghua Liao, Mingkuan Feng, Shuai Zhang, Zhengqi Wen, Haoran Luo, Ling Yang, Huazhe Xu, Jianhua Tao

Reinforcement learning (RL) has emerged as an effective paradigm for enhancing model reasoning. However, existing RL methods like GRPO often rely on unstructured self-sampling to fit scalar rewards, often producing inefficient rollouts that fail to capture transferable problem-solving strategies. To address these limitations, we propose **TemplateRL**, a structured template-guided RL framework that augments policy optimization with explicit template guidance. Our approach first constructs a problem-solving template library via MCTS on a small seed set, then seamlessly integrates this high-level structured guidance into RL training. By guiding rollout generation to align with proven template structures, TemplateRL significantly improves high-quality trajectory hit rates while reducing ineffective exploration. This structure-guided design steers the policy toward validated strategic patterns, stabilizing training dynamics, and enhancing RL sampling efficiency. Notably, the explicit template library is interpretable, editable, and supports online updates-enabling continuous updates during both training and inference. Extensive experiments demonstrate that TemplateRL outperforms GRPO by 99% on AIME and 41% on AMC, with superior stability on weak models and remarkable cross-domain generalization, highlighting its potential for broader tasks.