论文检索

Ziniu Liu, Shuheng Zhou, Mingqing Liu, Hao Deng, Huijia Zhu

Visual token pruning has emerged as a pivotal strategy to alleviate the computational bottleneck in Multimodal Large Language Models (MLLMs), yet it frequently compromises the integrity of visual understanding in pursuit of efficiency. Existing methods face a fundamental tension: vision-centric approaches are susceptible to the attention sink phenomenon and operate in a query-agnostic manner, whereas text-guided methods often create an overly narrow focus, discarding essential background context and failing on ambiguous queries. In this paper, we propose CrisPrune, a training-free and model-agnostic method that reconciles efficiency with understanding by integrating visual saliency and text relevance. Specifically, we introduce intrinsic visual saliency with robust normalization to identify information-rich regions characterized by significant visual features. Simultaneously, we design dual-source text relevance to synergize explicit instruction alignment with implicit scene priors. Finally, we reformulate the selection process using a Determinantal Point Process (DPP) to balance token quality and spatial diversity. Extensive experiments demonstrate that CrisPrune significantly outperforms state-of-the-art methods. On LLaVA-NeXT, it achieves a 13 \times decrease in FLOPs while maintaining 97% of the original performance with 94.4% of visual tokens pruned, effectively bridging the gap between efficiency and holistic understanding.

Wenqian Yu, Shuo Chen, Zhijiang Li, Zhipeng Wang, Jindong Gu

Large language models (LLMs) face a critical alignment challenge: balancing safety with helpfulness. Excessive safety can lead to over-refusal, where models reject harmful-looking yet benign queries, severely limiting utility.Existing training-free interventions offer an efficient way to mitigate over-refusal without re-training, but suffer from high inference overhead and architecture dependency. Our work explores a complementary direction: rather than applying post-hoc corrections to model outputs, our goal is to intrinsically reshape the distributions of harmful and benign samples within the model’s decision space. In this paper, we argue that a lightweight training-based approach can more effectively distinguish between harmful and benign samples. We propose Single Token Alignment (STA), which optimizes only a single-token prefix (e.g., 4,096 parameters) while keeping the base model frozen. To address the inherent challenge of achieving robust refinement through such a minimal parameter interface, STA employs a mixed weighting mechanism integrated with its optimization objective. This mechanism incorporates hard weighting via stringent data filtering to provide clear, unbiased learning signals, and soft weighting through a focal mechanism to prioritize challenging cases.Extensive experiments across 9 models and 10 datasets demonstrate that STA achieves a superior safety-helpfulness balance for LLMs, MLLMs, and reasoning models, offering a highly efficient and generalizable solution for refining safety alignment.

Dong She, Xianrong Yao, Liqun Chen, Jinghe Yu, Yang Gao, Zhanpeng Jin

Vision-Language Models (VLMs) have demonstrated strong capabilities in perception, yet holistic Affective Image Content Analysis (AICA)—which integrates perception, reasoning, and generation into a unified framework—remains underexplored. To address this, we introduce AICA-Bench, a comprehensive benchmark comprising three core tasks: Emotion Understanding (EU), Reasoning (ER), and Generation (EGCG). We evaluate 23 VLMs, revealing critical gaps: models struggle with intensity calibration and suffer from descriptive shallowness in open-ended tasks. To bridge these gaps, we propose Grounded Affective Tree (GAT) Prompting, a training-free framework that integrates visual scaffolding with hierarchical reasoning. Experiments show that GAT effectively corrects intensity errors and significantly enhances descriptive depth, establishing a robust baseline for future affective multimodal research.

Zijian Zheng, Yonghe Lu

Short text clustering has gained significant prominence due to its ubiquity in real-world applications. Despite the recent success of contrastive clustering, existing paradigms still suffer from two critical bottlenecks: (1) conventional data augmentation provides limited semantic granularity and may introduce unintended noise; and (2) the absence of global optimization for cluster assignments often precipitates the accumulation of pseudo-label noise, thereby compromising semantic consistency. To bridge these gaps, we propose MAST, a Multi-view Alignment Strategy with Transport-based clustering. MAST constructs complementary structural views to capture multi-granularity semantic features and introduces a multi-view contrastive objective that jointly aligns original, augmented, and structure-enhanced embeddings. To mitigate representation over-smoothing, we incorporate structure-aware negative reweighting and intermediate-layer negative sampling. Furthermore, MAST employs high-confidence guided refinement and an optimal transport-based pseudo-label alignment mechanism to enforce global semantic consistency across multiple views. Extensive experiments on several benchmark datasets demonstrate that MAST consistently outperforms state-of-the-art methods, establishing a new competitive baseline for short text clustering.

Yiju Guo, Tianyi Hu, Zexu Sun, Yankai Lin

Reinforcement Learning with Verifiable Rewards (RLVR) has advanced LLM reasoning, but remains constrained by inefficient exploration under limited rollout budgets, leading to low sampling success and unstable training in complex tasks. We find that many exploration failures arise not from problem difficulty, but from a small number of prompt tokens that introduce interference. Building on this insight, we propose the Less Noise Sampling Framework (LENS), which first purifies prompts by identifying and removing interference tokens. then transfers successful rollouts from the purified setting to supervise policy optimization on the original noisy prompts, enabling the model to learn to ignore interference in the real-world, noisy prompting settings. Experimental results show that LENS significantly outperforms GRPO, delivering higher performance and faster convergence, with a 3.88% average gain and over 1.6 \times speedup. Our work highlights the critical role of pruning interference tokens in improving rollout efficiency, offering a new perspective for RLVR research.

Yueqin Yin, Yaxi Li, Xin Liu, Xun Wang, Kaiqiang Song, Simin Ma, Shujian Liu, Sathish Reddy Indurthi, Haoyun Deng, Pengcheng He 等

Large language models often hallucinate, producing content that is factually incorrect or not grounded in the sources. Reliable faithfulness verification is critical for trustworthy deployment. In the provided-source (closed-world) setting, existing verifiers either classify whole passages in one step or check sentences independently, overlooking cross-sentence context. We present ContextCheck, a framework for sentence-level faithfulness verification with context-aware disambiguation. Each sentence is verified against the grounding document while conditioning on preceding sentences, enabling pronouns and references to be resolved directly in context. This design avoids the separate decontextualization step of rewriting claims into self-contained forms, casting verification as a context-conditioned task. Fine-tuned from Llama-3.1-8B-Instruct, ContextCheck sets a new state of the art on three context-dependent datasets; it improves Macro F1 by over 10 points compared to the strongest baselines, and matches or slightly surpasses the strongest baselines on 14 standard single-sentence datasets compared to prior 8B-scale verifiers (average Macro F1 73.5 vs. 72.8). These results show that ContextCheck offers a practical and effective approach for sentence-level hallucination detection.

Yoshinari Fujinuma

Large Language Models (LLMs) are commonly used as evaluators in various applications, but the reliability of the outcomes remains a challenge. One such challenge is using LLMs-as-judges for direct assessment, i.e., assigning scores from a specified range without any references. Using summarization as our primary testbed, we first show that this challenge stems from LLM judge outputs being associated with score range bias, i.e., LLM judge outputs are highly sensitive to pre-defined score ranges. We also show that similar biases exist among models from the same family. We then mitigate this bias through contrastive decoding, achieving up to 11.7% relative improvement in Spearman correlation with human judgments, averaged across score ranges.

Shangqing Tu, Yaxuan Li, Yushi Bai, Lei Hou, Juanzi Li

Parallel scaling has emerged as a powerful paradigm to enhance reasoning capabilities in large language models (LLMs) by generating multiple Chain-of-Thought (CoT) traces simultaneously. However, this approach introduces significant computational inefficiency due to *inter-trace redundancy*—our analysis reveals that over 80% of parallel reasoning traces yield identical final answers, representing substantial wasted computation. To address this critical efficiency bottleneck, we propose **DeepPrune**, a novel framework that enables efficient parallel scaling through dynamic pruning. Our method features a specialized judge model trained with oversampling techniques to accurately predict answer equivalence from partial reasoning traces, achieving 0.7072 AUROC on equivalence prediction across unseen reasoning models. This is combined with an online greedy clustering algorithm that dynamically prunes redundant paths while preserving answer diversity. Comprehensive evaluations across three challenging benchmarks (AIME 2024, AIME 2025, and GPQA) and multiple reasoning models demonstrate that DeepPrune achieves remarkable token reduction ranging from 65.73% to 88.50% compared to conventional consensus sampling, while maintaining competitive accuracy within 3.4 percentage points. Our work establishes a new standard for efficient parallel reasoning, making high-performance reasoning more efficient. Our code and data are here: https://github.com/THU-KEG/DeepPrune/

Cheng Liu, Xiaolei Liu, Xingyu Li, Bangzhou Xin, Kangyi Ding

Existing jailbreak defense paradigms primarily rely on static detection of prompts, outputs, or internal states, often neglecting the dynamic evolution of risk during decoding. This oversight leaves risk signals embedded in decoding trajectories underutilized, constituting a critical blind spot in current defense systems. In this work, we empirically demonstrate that hidden states in critical layers during the decoding phase carry stronger and more stable risk signals than input jailbreak prompts. Specifically, the hidden representations of tokens generated during jailbreak attempts progressively approach high-risk regions in the latent space. Based on this observation, we propose TrajGuard, a training-free, decoding-time defense framework. TrajGuard aggregates hidden-state trajectories via a sliding window to quantify risk in real time, triggering a lightweight semantic adjudication only when risk within a local window persistently exceeds a threshold. This mechanism enables the immediate interruption or constraint of subsequent decoding. Extensive experiments across 12 jailbreak attacks and various open-source LLMs show that TrajGuard achieves an average defense rate of 95%. Furthermore, it reduces detection latency to 5.2 ms/token while maintaining a false positive rate below 1.5%. These results confirm that hidden-state trajectories during decoding can effectively support real-time jailbreak detection, highlighting a promising direction for defenses without model modification.

Xidong Yang, Wenhao Li, Junjie Sheng, Yun Hua, Haosheng Chen, Chuyun Shen, Xiangfeng Wang

Reinforcement learning (RL) remains fundamentally limited by poor data efficiency and weak generalization. Prior episodic RL methods attempt to alleviate this via external memory modules, yet they suffer from two key limitations: a representation bottleneck caused by shallow encoders, and a retrieval dilemma where episodic memory is accessed indiscriminately.To address these challenges, we propose Agentic Episodic Control (AEC), a novel architecture that integrates large language models (LLMs) into episodic RL.AEC uses an LLM-based semantic augmenter to generate semantic representations from raw observations, and a critical state recognizer to selectively retrieve valuable experiences.This transforms memory usage from passive similarity matching into strategic, context-aware recall.Across five BabyAI-Text environments, AEC achieves 2–6× higher data efficiency than baselines and is the only method to solve complex tasks like UnlockLocal with over 90% success.It further demonstrates strong cross-task and cross-environment generalization, maintaining performance even under distribution shifts.AEC shows that combining LLM-derived priors with reinforcement learning yields more sample-efficient and adaptable agents. Code is available at https://github.com/Xidong-Yang/Agentic_Episodic_Control.

Rei Minamoto, Yusuke Oda, Daisuke Kawahara

Sensitive personal information can appear in large-scale pre-training corpora for large language models (LLMs). Detecting and filtering such information is therefore essential to ensure compliance with privacy regulations and prevent unintended information leakage. However, in contrast to English and other languages, research into sensitive personal information has been limited in the Japanese language. In this study, we focus on sensitive personal data defined as special care-required personal information (SCPI) under Japan’s Act on the Protection of Personal Information (APPI). We construct an SCPI dataset using LLM-based annotation and train machine learning models to rapidly detect SCPI in text. As a result, our SCPI classifier can effectively identify information related to SCPI. This study is the first to explore SCPI detection in Japanese text corpora, highlighting the challenges of accurate detection.

Yihao Ding, Siwen Luo, Yue Dai, Yanbei Jiang, Zechuan Li, Qiang Sun, Geoffrey Martin, Wei Liu, Yifan Peng

Visually Rich Document Understanding (VRDU) has become a pivotal area of research, driven by the need to automatically interpret documents that contain intricate visual, textual, and structural elements. Recently, Multimodal Large Language Models (MLLMs) have demonstrated significant promise in this domain, including both OCR-based and OCR-free approaches for information extraction from document images. This survey reviews recent advances in MLLM-based VRDU, highlighting emerging trends and promising research directions with a focus on two key aspects: (1) techniques for representing and integrating textual, visual, and layout features; (2) training paradigms, including pretraining, instruction tuning, and training strategies. Moreover, we address challenges such as data scarcity, handling multi-page and multilingual documents, and integrating emerging trends such as Retrieval-Augmented Generation and agentic frameworks. Our analysis offers a roadmap for advancing MLLM-based VRDU toward more scalable, reliable, and adaptable systems.

Jiakun Li, Xingwei He, Kefan Li, Hongzheng Chai, Hongyue Yu, Yuan Yuan

Test-time scaling improves the reasoning performance of large language models but often results in token-inefficient overthinking, where models continue reasoning beyond what is necessary for a correct answer. Existing dynamic early-exit methods typically rely on single-step confidence signals, which are often unreliable for detecting reasoning convergence in multi-step settings. To mitigate this limitation, we propose TRACE, a training-free framework for efficient test-time scaling that determines when to terminate reasoning based on temporal aggregation of multi-step evidence rather than instantaneous signals. TRACE detects reasoning convergence over time by aggregating two complementary signals across recent reasoning steps: answer consistency, capturing the persistence of predicted answers, and confidence trajectory, modeling the temporal evolution of model confidence. Benefiting from these two factors, TRACE can accurately determine whether the reasoning process has converged, thereby promptly halting inference and effectively avoiding redundant reasoning steps. Extensive experiments on multiple challenging benchmarks show that TRACE reduces reasoning token usage by 25–30% on average while maintaining accuracy within 1–2% of full-length reasoning, consistently outperforming existing dynamic reasoning methods.

Wenhao Zhang, Lin Mu, Li Ni, Peiquan Jin, Yiwen Zhang

Low-rank adaptation (LoRA) is a widely used strategy for efficient fine-tuning of large language models (LLMs), but its strictly linear structure fundamentally limits expressive capacity. The bilinear formulation of weight updates captures only first-order dependencies between low-rank factors, restricting the modeling of nonlinear and higher-order parameter interactions.In this paper, we propose Polynomial Expansion Rank Adaptation (PERA), a novel method that introduces structured polynomial expansion directly into the low-rank factor space.By expanding each low-rank factor to synthesize high-order interaction terms before composition, PERA transforms the adaptation space into a polynomial manifold capable of modeling richer nonlinear coupling without increasing rank or inference cost.We provide theoretical analysis demonstrating that PERA offers enhanced expressive capacity and more effective feature utilization compare to existing linear adaptation approaches.Empirically, PERA consistently outperforms state-of-the-art methods across diverse benchmarks. Notably, our experiments show that incorporating high-order nonlinear components—particularly square terms—is crucial for enhancing expressive capacity and maintaining strong and robust performance under various rank settings.

Yuan Sui, Yufei He, Tri Cao, Sophia Simeng Han, Yulin Chen, Bryan Hooi

Large Language Models (LLMs) often struggle with computational efficiency and error propagation in multi-step reasoning tasks. While recent advancements on prompting and post-training have enabled LLMs to perform step-wise reasoning, they still tend to explore unproductive solution paths without effective backtracking or strategy adjustment. In this paper, we propose Meta-Reasoner, a new framework that empowers LLMs to “think about how to think”. It optimizes the inference process by dynamically adapting reasoning strategies in real-time. Our approach employs contextual multi-armed bandits (CMABs) to learn an adaptive policy. It learns to evaluate the current state of LLM’s reasoning and determine optimal strategy that is most likely to lead to a successful outcome during inference, like whether to backtrack, switch to a new approach, or restart the problem-solving process. This meta-guidance helps avoid unproductive paths exploration during inference and hence improves computational efficiency. We evaluate Meta-Reasoner on math problems (e.g., Game-of-24, TheoremQA) and scientific tasks (e.g., SciBench). Results show that our method outperform previous SOTA methods by 9-12% in accuracy, while reducing inference time by 28-35% under the same compute budget. Additional experiments on creative writing demonstrate the generalizability of our approach to diverse reasoning-intensive tasks.

Haiyan Wu, Chenchen Wang, Chaoqun Sun, Chengxiong Lu, Yan-Hong Chen, Zhiqiang Zhang, Xiaoqing Feng

Retrieval-Augmented Generation (RAG) has emerged as a widely adopted paradigm for grounding Large Language Models (LLMs) in external knowledge. Recent agentic RAG systems introduce multi-turn reasoning, but they often lack the capacity to evaluate the utility of retrieved information, leading to brittle reasoning and suboptimal decision-making. We propose Reflective RAG, an agentic framework that incorporates self-evaluation to dynamically optimize retrieval and generation strategy. At its core, Reflective RAG employs a reflection tagging mechanism that allows the model to critique the relevance of retrieved content, thereby explicitly guiding its subsequent policy. To ensure robust learning, we introduce a two-stage training procedure that partially decouples evaluation semantics from strategy optimization. First, during supervised fine-tuning (SFT), the model learns to generate accurate reflection signals by self-correcting labels based on internal uncertainty. Second, a reinforcement learning (RL) stage optimizes the agent’s strategy using these reflections, stabilized by dynamic KL regularization. Evaluations across five knowledge-intensive QA benchmarks demonstrate that Reflective RAG consistently outperforms strong agentic baselines. Further analysis demonstrates its improved training stability and stronger generalization to complex multi-hop reasoning tasks.

Junxi Wang, Te Sun, Jiayi Zhu, Junxian Li, Haowen Xu, Zichen Wen, Xuming Hu, Zhiyu li, Linfeng Zhang

Vision agent memory has shown remarkable effectiveness in long-video understanding; however, storing such memory for videos incurs substantial overhead, leading to high costs in both storage and computation. To address this issue, we propose StreamMeCo, an efficient Stream Agent Memory Compression framework. Specifically, based on the connectivity of the memory graph, StreamMeCo introduces edge-free minmax sampling for isolated nodes and edge-aware weight pruning for connected nodes, evicting redundant memory nodes while maintaining accuracy. In addition, we introduce a time-decay memory retrieval mechanism to mitigate the performance degradation caused by memory compression. Extensive experiments on three challenging benchmark datasets (M3-Bench-robot, M3-Bench-web, and Video-MME-Long) demonstrate that under 70% memory graph compression, StreamMeCo achieves a 1.87× speedup in memory retrieval while delivering an average accuracy improvement of 1.0%. Our code is available at https://github.com/Celina-love-sweet/StreamMeCo.

Yebo Wu, Han Jin, Zhijiang Guo, Li Li

Multimodal Large Language Models (MLLMs) have demonstrated remarkable reasoning capabilities yet continue to suffer from hallucination, where generated text contradicts visual content. In this paper, we introduce Dual-Anchor Introspective Decoding (DaID), a novel contrastive decoding framework that dynamically calibrates each token generation by mining the model’s internal perceptual discrepancies. Specifically, DaID identifies a Spotlight layer to amplify visual factual signals and a Shadow layer to suppress textual inertia. By leveraging visual attention distributions to guide this dual-anchor selection process, our method ensures precise, token-specific adaptation. Experimental results across multiple benchmarks and MLLMs demonstrate that DaID significantly mitigates hallucination while enhancing general reasoning capabilities.

Lin Mu, Guoji Wang, Li Ni, Lei Sang, Zhize Wu, Peiquan Jin, Yiwen Zhang

Large Language Models (LLMs) have shown strong potential for recommendation (LLMRec) due to their powerful reasoning and generalization abilities. However, effectively aligning the textual semantics modeled by LLMs with the collaborative signals remains a key challenge. Existing methods either translate collaborative information into textual prompts or inject pre-trained embeddings into the LLM, both of which treat structural information as static input and fail to capture high-order relational dependencies.To bridge this gap, we propose GraphLoRA, a novel framework that generalizes low-rank adaptation from independent to structure-aware propagation. GraphLoRA embeds a trainable graph message-passing network within the low-rank adaptation pathway, enabling structural signals to propagate through the parameter space.This design allows collaborative topology to explicitly guide parameter updates, fostering deep integration between graph-structured and textual semantic information. Extensive experiments on multiple benchmarks demonstrate that GraphLoRA not only outperforms state-of-the-art LLM-based recommendation methods but also achieves superior generalization, effectively balancing structural reasoning capability with computational efficiency.

Yujie Liu, Zonglin Yang, Tong Xie, Jinjie Ni, Ben Gao, Yuqiang Li, Shixiang Tang, Wanli Ouyang, Erik Cambria, Dongzhan Zhou

Large language models (LLMs) have shown potential in assisting scientific research, yet their ability to discover high-quality research hypotheses remains unexamined due to the lack of a dedicated benchmark. To address this gap, we introduce the first large-scale benchmark for evaluating LLMs on a sufficient set of scientific discovery sub-tasks—inspiration retrieval, hypothesis composition, and hypothesis ranking—where sufficient means that perfectly solving these sub-tasks perfectly solves the overall discovery task. We develop an automated LLM-based framework that extracts critical components—research questions, background surveys, inspirations, and hypotheses—from papers across 12 disciplines, with expert validation confirming its accuracy. To prevent data contamination, we focus exclusively on publications from 2024 onward, ensuring minimal overlap with LLM pretraining data; our automated framework further enables automatic extraction of even more recent papers as LLM pretraining cutoffs advance, supporting scalable and contamination-free automatic renewal of this discovery benchmark. Our evaluation shows that, across disciplines, LLMs excel at inspiration retrieval—an out-of-distribution task—suggesting their ability to surface novel knowledge associations.