论文检索

Xi Chen, Aske Plaat, Niki Van Stein

Chain‑of‑thought (CoT) prompting boosts Large Language Models accuracy on multi‑step tasks, yet whether the generated ``thoughts'' reflect the true internal reasoning process is unresolved. We present the first feature‑level causal study of CoT faithfulness. Combining sparse autoencoders with activation patching, we extract monosemantic features from Pythia‑70M and Pythia‑2.8B while they tackle GSM8K math problems under CoT and plain (noCoT) prompting. Swapping a small set of CoT‑reasoning features into a noCoT run raises answer log‑probabilities significantly in the 2.8B model, but has no reliable effect in 70M, revealing a clear contrast for these two scales. CoT also leads to significantly higher activation sparsity and feature interpretability scores in the larger model, signalling more modular internal computation. For example, the model's confidence in generating correct answers improves from 1.2 to 4.3. We introduce patch‑curves and random‑feature patching baselines, showing that useful CoT information is not only present in the top-K patches but widely distributed. Overall, our results indicate that CoT can induce more interpretable internal structures in high-capacity LLMs, validating its role as a structured prompting method.

Tianxiang Chen, Zhentao Tan, Xiaofan Bo, Yue Wu, Tao Gong, Qi Chu, Jieping Ye

Effectively handling long contexts is challenging for Large Language Models (LLMs) due to the rarity of long texts, high computational demands, and substantial forgetting of short-context abilities. Recent approaches have attempted to construct long contexts for instruction tuning, but these methods often require LLMs or human interventions, which are both costly and limited in length and diversity. Also, the drop in short-context performances of present long-context LLMs remains significant. In this paper, we introduce Flora, an effortless (human/LLM-free) long-context construction strategy. Flora can markedly enhance the long-context performance of LLMs by arbitrarily assembling short instructions based on categories and instructing LLMs to generate responses based on long-context meta-instructions. This enables Flora to produce contexts of arbitrary length and scale with rich diversity, while only slightly compromising short-context performance. Experiments on Llama3-8B-Instruct and QwQ-32B show that LLMs enhanced by Flora excel in three long-context benchmarks while maintaining strong performances in short-context tasks.

Sirui Chen, Mengshi Zhao, Lei Xu, Yuying Zhao, Beier Zhu, Hanwang Zhang, Shengjie Zhao, Chaochao Lu

Recent advances in large language models (LLMs) have greatly improved their reasoning and decision-making abilities when deployed as agents. Richer reasoning, however, often comes at the cost of longer chain of thought (CoT), hampering interaction efficiency in real-world scenarios. Nevertheless, there still lacks systematic definition of LLM‑Agent efficiency, hindering targeted improvements. To this end, we introduce dual‑efficiency, comprising (i) step-level efficiency, which minimizes tokens per step, and (ii) trajectory-level efficiency, which minimizes the number of steps to complete a task. Building on this definition, we propose DEPO, a dual-efficiency preference‑based optimization method that jointly rewards succinct responses and fewer action steps. Experiments on WebShop and BabyAI show that DEPO cuts token usage by up to 60.9% and steps by up to 26.9%, while achieving up to a 29.3% improvement in task performance. DEPO also generalizes to three out-of-domain math benchmarks and retains its efficiency gains when trained on only 25% of the data.

Shengyuan Chen, Chuang Zhou, Zheng Yuan, Qinggang Zhang, Zeyang Cui, Hao Chen, Yilin Xiao, Jiannong Cao, Xiao Huang

Large language models (LLMs) often suffer from hallucination, generating factually incorrect statements when handling questions beyond their knowledge and perception. Retrieval-augmented generation (RAG) addresses this by retrieving query-relevant contexts from knowledge bases to support LLM reasoning. Recent advances leverage pre-constructed graphs to capture the relational connections among distributed documents, showing remarkable performance in complex tasks. However, existing Graph-based RAG (GraphRAG) methods rely on a costly process to transform the corpus into a graph, introducing overwhelming token cost and update latency. Moreover, real-world queries vary in type and complexity, requiring different logic structures for accurate reasoning. The pre-built graph may not align with these required structures, resulting in ineffective knowledge retrieval. To this end, we propose a Logic-aware Retrieval Augmented Generation framework (LogicRAG) that dynamically extracts reasoning structures at inference time to guide adaptive retrieval without any pre-built graph. LogicRAG begins by decomposing the input query into a set of subproblems and constructing a directed acyclic graph (DAG) to model the logical dependencies among them. To support coherent multi-step reasoning, LogicRAG then linearizes the graph using topological sort, so that subproblems can be addressed in a logically consistent order. Besides, LogicRAG applies graph pruning to reduce redundant retrieval and uses context pruning to filter irrelevant context, significantly reducing the overall token cost. Extensive experiments demonstrate that LogicRAG achieves both superior performance and efficiency compared to state-of-the-art baselines.

Qiguang Chen, Dengyun Peng, Jinhao Liu, Huikang Su, Jiannan Guan, Libo Qin, Wanxiang Che

Recent advancements in large language models (LLMs) have greatly improved their ability to perform complex reasoning tasks through Long Chain-of-Thought (CoT). However, this approach often results in substantial redundancy, impairing computational efficiency and causing significant delays in real-time applications. To improve efficiency, current methods often rely on human-defined difficulty priors, which do not align with the LLM's self-awared difficulty, leading to inefficiencies. In this paper, we introduce the Dynamic Reasoning-Boundary Self-Awareness Framework (DR. SAF), which enables LLMs to dynamically assess and adjust their reasoning depth in response to problem complexity. DR. SAF integrates three key components: Boundary Self-Awareness Alignment, Adaptive Reward Management, and a Boundary Preservation Mechanism. These components allow models to optimize their reasoning processes, balancing efficiency and accuracy without compromising performance. Our experimental results demonstrate that DR. SAF achieves a 49.27% reduction in total response tokens with minimal loss in accuracy. The framework also delivers a 6.59x gain in token efficiency and a 5x reduction in training time, making it well-suited to resource-limited settings. During extreme training, DR. SAF can even surpass traditional instruction-based models in token efficiency with more than 16% accuracy improvement.

Ling-Chun Chen, Hsi-Wen Chen, Ming-Syan Chen

Large Language Models (LLMs) have achieved remarkable success across reasoning and knowledge-intensive tasks, yet their static pretraining leaves them unable to handle rapidly evolving or domain-specific knowledge. Retrieval-Augmented Generation (RAG) addresses this by grounding LLM outputs in dynamically retrieved evidence, improving factual accuracy and reducing hallucinations. However, standard RAG pipelines struggle with temporally sensitive queries, especially when documents contain fuzzy or indirect time expressions (e.g., “a few years later”). This leads to Temporal Misalignment, where topically relevant but temporally incorrect results are retrieved. To overcome this, we propose DeFuzzRAG, a lightweight framework that enhances temporal robustness in RAG. DeFuzzRAG employs a small local language model to infer concrete time scopes from vague expressions and applies metadata-based filtering to realign retrieval with the query’s temporal intent. Experiments on a benchmark of fuzzified queries demonstrate that DeFuzzRAG substantially improves retrieval accuracy, raising Hit Rate by 15.7% while maintaining efficiency and model-agnostic integration. Our findings highlight the importance of temporal reasoning in RAG and establish DeFuzzRAG as a practical, plug-and-play solution for deploying temporally robust LLM systems in real-world settings.

Lei Chen, Tengteng Cheng, BoYu Gao, Zitao Liu, Weiqi Luo

Automated grading of student responses still faces numerous challenges, particularly when dealing with complex and ambiguous answers. In particular, large models are prone to scoring bias when handling uncertain responses, and few-shot reasoning methods often lack stability, which limits their applicability in real educational scenarios. To tackle these challenges, we propose the Contrastive Error Mining and FineTuning (CEM-FT) framework, which automatically identifies high-value hard samples by analyzing scoring disagreements between a full fine-tuned model and a few-shot model. A lightweight LoRA adapter is then trained on these samples to refine model performance with minimal computational overhead. Experiments on the SciEntsbank, Beetle, and Mohler datasets show that CEM-FT can improve QWK by up to 3.9% compared to the fine-tuned Qwen model on SciEntsbank datasets, which is a significant improvement over the few-shot baseline. The proposed framework substantially enhances both scoring accuracy and consistency, providing a practical, robust solution for reliable automated assessment with large language models.

Junjie Chen, Weihang Su, Zhumin Chu, Haitao Li, Yujia Zhou, Dingbo Yuan, Xudong Wang, Jun Zhou, Yiqun Liu, Min Zhang 等

The rapid development of large language models (LLMs) has highlighted the need for efficient and reliable methods to evaluate their performance. Traditional evaluation methods often face challenges like high costs, limited task formats, dependence on human references, and systematic biases. To address these limitations, we propose Auto-PRE, an automatic LLM evaluation framework inspired by the peer review process. Unlike previous approaches that rely on human annotations, Auto-PRE automatically selects evaluator LLMs based on three core traits: consistency, pertinence, and self-confidence, which correspond to the instruction, content, and response stages, respectively, and collectively cover the entire evaluation process. Experiments on three representative tasks, including summarization, non-factoid QA, and dialogue generation, demonstrate that Auto-PRE achieves state-of-the-art performance while significantly reducing evaluation costs. Furthermore, the structured and scalable design of our automatic qualification exam framework provides valuable insights into automating the evaluation of LLMs-as-judges, paving the way for more advanced LLM-based evaluation frameworks.

Haoyang Chen, Richong Zhang, Junfan Chen

Large language models (LLMs) perform in-context learning (ICL) with minimal supervised examples, which benefits various natural language processing (NLP) tasks. One of the critical research focus is the selection of prompt demonstrations. Current approaches typically employ retrieval models to select the top-K most semantically similar examples as demonstrations. However, we argue that existing methods are limited since the label consistency is not guaranteed during demonstration selection. Our cognition derives from the Bayesian view of ICL and our rethinking of ICL from the transductive label propagation perspective. We treat ICL as a transductive learning method and incorporate latent concepts from Bayesian view and deduce that similar demonstrations guide the concepts of query, with consistent labels serving as estimates. Based on this understanding, we establish a label propagation framework to link label consistency with propagation error bounds. To model label consistency, we propose a data synthesis method, leveraging both semantic and label information, and use TopK sampling with Synthetic Data (TopK-SD) to acquire demonstrations with consistent labels. TopK-SD outperforms original TopK sampling on multiple benchmarks. Our work provides a new perspective for understanding the working mechanisms within ICL.

Haokun Chen, Jianing Li, Yao Zhang, Jinhe Bi, Yan Xia, Jindong Gu, Volker Tresp

Multimodal Large Language Models (MLLMs) achieve impressive performance once optimized on massive datasets. Such datasets often contain sensitive or copyrighted content, raising significant data privacy concerns. Regulatory frameworks mandating the 'right to be forgotten' drive the need for machine unlearning. This technique allows for the removal of target data without resource-consuming retraining. However, while well-studied for text, visual concept unlearning in MLLMs remains underexplored. A primary challenge is precisely removing a target visual concept without disrupting model performance on related entities. To address this, we introduce AUVIC, a novel visual concept unlearning framework for MLLMs. AUVIC applies adversarial perturbations to enable precise forgetting. This approach effectively isolates the target concept while avoiding unintended effects on similar entities. To evaluate our method, we construct VCUBench. It is the first benchmark designed to assess visual concept unlearning in group contexts. Experimental results demonstrate that AUVIC achieves state-of-the-art target forgetting rates while incurs minimal performance degradation on non-target concepts.

Hang Chen, Xinyu Yang, Jiaying Zhu, Wenya Wang

Circuit graph discovery has emerged as a fundamental approach to elucidating the skill mechanistic of language models. Despite the output faithfulness of circuit graphs, they suffer from atomic ablation, which causes the loss of causal dependencies between connected components. In addition, their discovery process, designed to preserve output faithfulness, inadvertently captures extraneous effects other than an isolated target skill. To alleviate these challenges, we introduce skill paths, which offer a more refined and compact representation by isolating individual skills within a linear chain of components. To enable skill path extracting from circuit graphs, we propose a three-step framework, consisting of decomposition, pruning, and post-hoc causal mediation. In particular, we offer a complete linear decomposition of the transformer model which leads to a disentangled computation graph. After pruning, we further adopt causal analysis techniques, including counterfactuals and interventions, to extract the final skill paths from the circuit graph. To underscore the significance of skill paths, we investigate three generic language skills—Previous Token Skill, Induction Skill, and In-Context Learning Skill—using our framework. Experiments support two crucial properties of these skills, namely stratification and inclusiveness.

Guo Chen, Junjie Huang, Huaijin Xie, Fei Sun, Tao Jia

Retrieval-Augmented Generation (RAG) effectively enhances Large Language Models (LLMs) by incorporating retrieved external knowledge into the generation process. Reasoning models improve LLM performance in multi-hop QA tasks, which require integrating and reasoning over multiple pieces of evidence across different documents to answer a complex question. However, they often introduce substantial computational costs, including increased token consumption and inference latency. To better understand and mitigate this trade-off, we conduct a comprehensive study of reasoning strategies for reasoning models in RAG multi-hop QA tasks. Our findings reveal that reasoning models adopt structured strategies to integrate retrieved and internal knowledge, primarily following two modes: Context-Grounded Reasoning, which relies directly on retrieved content, and Knowledge-Reconciled Reasoning, which resolves conflicts or gaps using internal knowledge. To this end, we propose a novel Lightweight Rerank Reasoning Strategy Framework for RAG (LiR³AG) to enable non-reasoning models to transfer reasoning strategies by restructuring retrieved evidence into coherent reasoning chains. LiR³AG significantly reduce the average 98% output tokens overhead and 58.6% inferencing time while improving 8B non-reasoning model's F1 performance ranging from 6.2% to 22.5% to surpass the performance of 32B reasoning model in RAG, offering a practical and efficient path forward for RAG systems.

Canyu Chen, Baixiang Huang, Zekun Li, Zhaorun Chen, Shiyang Lai, Xiongxiao Xu, Jia-Chen Gu, Jindong Gu, Huaxiu Yao, Chaowei Xiao 等

Large Language Models (LLMs) have emerged as a new information channel. Meanwhile, one critical but under-explored question is: Is it possible to bypass the safety alignment and inject harmful information into LLMs stealthily? In this paper, we propose to reformulate knowledge editing as a new type of safety threat for LLMs, namely Editing Attack, and conduct a systematic investigation with a newly constructed dataset EditAttack. Specifically, we focus on two typical safety risks of Editing Attack including Misinformation Injection and Bias Injection. For the first risk, we find that editing attacks can inject both commonsense and long-tail misinformation into LLMs, and the effectiveness for the former one is particularly high. For the second risk, we discover that not only can biased sentences be injected into LLMs with high effectiveness, but also one single biased sentence injection can degrade the overall fairness. Then, we further illustrate the high stealthiness of editing attacks. Our discoveries demonstrate the emerging misuse risks of knowledge editing techniques on compromising the safety alignment of LLMs and the feasibility of disseminating misinformation or bias with LLMs as new channels.

Boyu Chen, Zirui Guo, Zidan Yang, Yuluo Chen, Junze Chen, Zhenghao Liu, Chuan Shi, Cheng Yang

Retrieval-augmented generation (RAG) improves the response quality of large language models (LLMs) by retrieving knowledge from external databases. Typical RAG approaches split the text database into chunks, organizing them in a flat structure for efficient searches. To better capture the inherent dependencies and structured relationships across the text database, researchers propose to organize textual information into an indexing graph, known as graph-based RAG. However, we argue that the limitation of current graph-based RAG methods lies in the redundancy of the retrieved information, rather than its insufficiency. Moreover, previous methods use a flat structure to organize retrieved information within the prompts, leading to suboptimal performance. To overcome these limitations, we propose PathRAG, which retrieves key relational paths from the indexing graph, and converts these paths into textual form for prompting LLMs. Specifically, PathRAG effectively reduces redundant information with flow-based pruning, while guiding LLMs to generate more logical and coherent responses with path-based prompting. Experimental results show that PathRAG consistently outperforms state-of-the-art baselines across six datasets and five evaluation dimensions.

Da Chang, Peng Xue, Yu Li, Yongxiang Liu, Pengxiang Xu, Shixun Zhang

Parameter-Efficient Fine-Tuning (PEFT) methods are crucial for adapting large pre-trained models. Among these, LoRA is considered a foundational approach. Building on this, the influential DoRA method enhances performance by decomposing weight updates into magnitude and direction. However, its underlying mechanism remains unclear, and it introduces significant computational overhead. In this work, we first identify that DoRA's success stems from its capacity to increase the singular value entropy of the weight update matrix, which promotes a more uniform update distribution akin to full fine-tuning. We then reformulate DoRA into a mathematically equivalent and more efficient matrix form, revealing it as a learnable weight conditioning method. Based on this insight, we propose a unified framework for designing advanced PEFT methods by exploring two orthogonal dimensions: the architectural placement and the transformation type of the conditioning matrix. Within this framework, we introduce two novel methods: (1) Pre-Diag, which applies a diagonal conditioning matrix before the LoRA update to efficiently calibrate the pre-trained weights, thereby enhancing performance while reducing training time; and (2) Skewed Orthogonal Rotation Adaptation (SORA), which employs a parameter-efficient orthogonal rotation to perform a more powerful, norm-preserving transformation of the feature space. Extensive experiments on natural language understanding and generation tasks demonstrate that our proposed methods achieve superior performance and efficiency compared to both LoRA and DoRA.

Ziyi Cao, Qingyi Si, Jingbin Zhang, Bingquan Liu

Large language models face significant cost challenges in long-sequence inference. To address this, reusing historical Key-Value (KV) Cache for improved inference efficiency has become a mainstream approach. Recent advances further enhance throughput by sparse attention mechanisms to select the most relevant KV Cache, thereby reducing sequence length. However, such techniques are limited to single-context scenarios, where historical KV Cache is computed sequentially with causal-attention dependencies. In retrieval-augmented generation (RAG) scenarios, where retrieved documents as context are unknown beforehand, each document’s KV Cache is computed and stored independently (termed multiple-context KV Cache), lacking cross-attention between contexts. This renders existing methods ineffective. Although prior work partially recomputes multiple-context KV Cache to mitigate accuracy loss from missing cross-attention, it requires retaining all KV Cache throughout, failing to reduce memory overhead. This paper presents SamKV, the first exploration of attention sparsification for multiple-context KV Cache. Specifically, SamKV takes into account the complementary information of other contexts when sparsifying one context, and then locally recomputes the sparsified information. Experiments demonstrate that our method compresses sequence length to 15% without accuracy degradation compared with full-recomputation baselines, significantly boosting throughput in multi-context RAG scenarios.

Yukun Cao, Zirui Xu, Dongyang Li, Zhihao Guo, Luobin Huang, Lisheng Wang

Knowledge Graph (KG)-based Retrieval-Augmented Generation (RAG) shifts the contents of retrieval from narrative text to a relational knowledge network, empowering large language models (LLMs) to harness structured relationships between entities. However, conventional KG-RAG approaches are resource-intensive, requiring either query decomposition with multiple LLM rounds or parameterized static knowledge injection to update the model. Although subgraph reasoning aims to address these issues, most current methods are based on heuristic shortest path and multi-hop graph traversal algorithms. The retrieved subgraphs suffer from incompleteness and semantic drift, and neglect the interaction between subgraph and LLMs in terms of fine-grained structural semantics. We propose a dual-constraint subgraph optimization for KG-RAG (DCTR). It improves subgraph retrieval and generates high-quality subgraphs with structural integrity and information salience for LLMs. Specifically, it formulates subgraph generation as a two-stage graph-theoretic constrained optimization problem to create compact and complete pseudo-labels. Since these pseudo-labels are discrete, a smooth approximation is employed to convert them into a differentiable representation, thereby optimizing the retriever to highlight key information while extracting subgraphs. On two benchmark datasets, DCTR significantly enhances subgraph quality, achieving state-of-the-art performance in LLM reasoning.

Yi Cao, Weijie Shi, Wei-Jie Xu, Yucheng Shen, Yue Cui, Hanghui Guo, Shimin Di, Ziyi Liu, Jiaming Li, Alexander Zhou 等

Large Reasoning Models (LRMs) have recently demonstrated impressive performance across a range of reasoning tasks by generating intermediate thoughts. However, these models can suffer from overthinking—generating excessive tokens that contribute little to final accuracy while increasing inference cost. To mitigate this, we propose TIV (Thought Injection via Vectors), an innovative framework that compresses token-level reasoning into compact vectors without sacrificing performance. Rather than generating explicit thoughts, TIV injects learnable vectors into the post-attention hidden states of the final token across Transformer layers, enabling implicit and lightweight reasoning. We further introduce a two-stage reinforcement learning strategy: the first stage calibrates the model's reasoning distribution, and the second distills it into a vector-based policy optimized for both accuracy and brevity. Experiments on three reasoning benchmarks show that TIV preserves over 99% of the original accuracy while reducing output length by more than 65% on average, reaching up to 80% in some cases. Moreover, TIV consistently achieves superior trade-offs between accuracy and efficiency compared to existing methods, distinguishing itself as a state-of-the-art (SOTA) approach for efficient reasoning in LRMs.

Ding Cao, Yuchen Cai, Yuqing Huang, Xuesong He, Rongxi Guo, Guiquan Liu, Guangzhong Sun

Sequential knowledge editing techniques aim to continuously update knowledge in large language models at low cost, preventing models from generating outdated or incorrect information. However, existing sequential editing methods suffer from a significant decline in editing success rates after long-term editing. Through theoretical analysis and experiments, our findings reveal that as the number of edits increases, the model's output increasingly deviates from the desired target, leading to a drop in editing success rates. We refer to this issue as the superimposed noise accumulation problem. Our further analysis demonstrates that the problem is related to the erroneous activation of irrelevant knowledge and conflicts between activated knowledge. Based on this analysis, a method named DeltaEdit is proposed that reduces conflicts between knowledge through dynamic orthogonal constraint strategies. Experiments show that DeltaEdit significantly reduces superimposed noise, achieving a 16.8% improvement in editing performance over the strongest baseline.

Boxi Cao, Ruotong Pan, Hongyu Lin, Xianpei Han, Le Sun

Multiple-choices question answering (MCQA) has emerged as one of the most popular task formats for large language models (LLMs) evaluation. Unfortunately, there exist substantial evidence that the evaluation of current MCQA benchmarks suffers from significant answer bias, which severely undermines the reliability of the evaluation conclusions. Specifically, many LLMs achieve performance significantly higher than random selection even when the questions are omitted from input information. To this end, we conduct a systematic investigation of the attribution of answer bias, and demonstrate a strong correlation between the degree of data contamination and the severity of answer bias, while the position of options and the popularity of answers have relatively minor effects. Building on these insights, we further propose OPD, a straightforward yet effective tool for contamination detection and dataset debiasing without requiring access to the model’s internal training data. Our findings and algorithms provide valuable insights for the design of future trustworthy LLM evaluation protocols.