论文检索

Cong Tan, Yongqi Shao, Hong Huo, Tao Fang

Dense retrieval models commonly use flat indexes to achieve high-precision retrieval by computing exact distances between embedding vectors. However, flat indexes are memory-intensive and inefficient, limiting their scalability in large-scale retrieval tasks. In contrast, quantized indexes enable faster retrieval with significantly lower memory usage, but their accuracy tends to decrease. Therefore, we propose a scalable and efficient training method for the dual-encoder models to improves the retrieval accuracy on quantized indexes. Our approach combines the direct gradient update to the cached target embeddings with large scale negative sampling based on similarity, significantly reducing computational overhead and GPU memory usage. Target embeddings are initialized with a pre-trained encoder and stored in a memory buffer, which is directly updated via backpropagation, thus avoiding the repeated re-encoding of the full corpus. To build a rich set of negatives, we retrieve the top-k most similar targets for each query from cached embeddings using the quantized index, including both query-specific and cross-batch top-k results. This design effectively approximates the truncated softmax distribution. The experiments show that our method achieves performs exceptionally well on quantized indexes, providing a practical and scalable solution for real-world retrieval systems.

Zetian Sun, Dongfang Li, Baotian Hu, Min Zhang

Large language models (LLMs) have achieved remarkable success in a wide range of tasks. However, their reasoning capabilities, particularly in complex domains like mathematics, remain a significant challenge. Value-based process verifiers, which estimate the probability of a partial reasoning chain leading to a correct solution, are a promising approach for improving reasoning. Nevertheless, their effectiveness is often hindered by estimation error in their training annotations, a consequence of the limited number of Monte Carlo (MC) samples feasible due to the high cost of LLM inference. In this paper, we identify that the estimation error primarily arises from high variance rather than bias, and the MC estimator is a Minimum Variance Unbiased Estimator (MVUE). To address the problem, we propose the Compound Monte Carlo Sampling (ComMCS) method, which constructs an unbiased estimator by linearly combining the MC estimators from the current and subsequent steps. Theoretically, we show that our method leads to a predictable reduction in variance, while maintaining an unbiased estimation without additional LLM inference cost. We also perform empirical experiments on the MATH-500 and GSM8K benchmarks to demonstrate the effectiveness of our method. Notably, ComMCS outperforms regression-based optimization method by 2.8 points, the non-variance-reduced baseline by 2.2 points on MATH-500 on Best-of-32 sampling experiment.

Zechen Sun, Zecheng Tang, Juntao Li, Wenpeng Hu, Wenliang Chen, Zhunchen Luo, Qiaoming Zhu

High-quality multi-hop instruction data is critical for enhancing the reasoning capabilities of large language models (LLMs) in complex long-context scenarios, e.g., long-form reasoning. Nevertheless, there is currently a notable scarcity of such datasets within the community, and existing data synthesis approaches typically fail to provide explicit modeling of intermediate reasoning steps, resulting in unverifiable and potentially erroneous samples. To mitigate above issue, we design the Concept-Graph based Multi-hop Instructions Synthesis (CGMIS) framework, which constructs long-form reasoning paths via concept graph traversal and automatically generates verifiable multi-hop data. The CGMIS framework not only guarantees the accuracy and verifiability of the synthesized data but also enables the construction of high-quality multi-hop instruction datasets from arbitrary corpora. Experiments show that fine-tuning with CGMIS-generated data achieves state-of-the-art performance across 13 long-context reasoning tasks on various models, using only 10% of the data volume required by existing methods.

Yiliu Sun, Zicheng Zhao, Yang Wei, Yanfang Zhang, Chen Gong

Reinforcement Learning with Verifiable Rewards (RLVR) significantly enhances the reasoning capability of Large Language Models (LLMs). Current RLVR approaches typically conduct training across all generated tokens, but neglect to explore which tokens (e.g., prefix tokens) actually contribute to reasoning. This uniform training strategy spends substantial effort on optimizing low-return tokens, which in turn impedes the potential improvement from high-return tokens and reduces overall training effectiveness. To address this issue, we propose a novel RLVR approach called Progressive Prefix-token Policy Optimization (PPPO), which highlights the significance of the prefix segment of generated outputs. Specifically, inspired by the well-established human thinking theory of Path Dependence, where early-stage thoughts substantially constrain subsequent thinking trajectory, we identify an analogous phenomenon in LLM reasoning termed Beginning Lock-in Effect (BLE). PPPO leverages this finding by focusing its optimization objective on the prefix reasoning process of LLMs. This targeted optimization strategy can positively influence subsequent reasoning processes, and ultimately improve final results. To improve the learning effectiveness of LLMs on how to start reasoning with high quality, PPPO introduces two training strategies: (a) Progressive Prefix Retention, which shapes a progressive learning process by increasing the proportion of retained prefix tokens during training; (b) Continuation Accumulated Reward, which mitigates reward bias by sampling multiple continuations for one prefix token sequence, and accumulating their scores as the reward signal. Extensive experimental results on various reasoning tasks (e.g., math, physics, chemistry, and biology) demonstrate that our proposed PPPO outperforms representative RLVR methods, with the accuracy improvements of 18.02% on only 26.17% training tokens.

Yanming Sun, Runzhe Zhan, Chi Seng Cheang, Han Wu, Xuebo Liu, Yuyao Niu, Fengying Ye, Kaixin Lan, Lidia S. Chao, Derek F. Wong

REtrieval-Augmented LLM-based Machine Translation (REAL-MT) shows promise for knowledge-intensive tasks like idiomatic translation, but its reliability under noisy retrieval, a common challenge in real-world deployment, remains poorly understood. To address this gap, we propose a noise synthesis framework and new metrics to systematically evaluate REAL-MT’s reliability across high-, medium-, and low-resource language pairs. Using both open- and closed-sourced models, including standard LLMs and large reasoning models (LRMs), we find that models heavily rely on retrieved context, and this dependence is significantly more detrimental in low-resource language pairs, producing nonsensical translations. Although LRMs possess enhanced reasoning capabilities, they show no improvement in error correction and are even more susceptible to noise, tending to rationalize incorrect contexts. Attention analysis reveals a shift from the source idiom to noisy content, while confidence increases despite declining accuracy, indicating poor self-monitoring. To mitigate these issues, we investigate training-free and fine-tuning strategies, which improve robustness at the cost of performance in clean contexts, revealing a fundamental trade-off. Our findings highlight the limitations of current approaches, underscoring the need for self-verifying integration mechanisms.

Weiwei Sun, Shengyu Feng, Shanda Li, Yiming Yang

Although LLM-based agents have attracted significant attention in domains such as software engineering and machine learning research, their role in advancing combinatorial optimization (CO) remains relatively underexplored. This gap underscores the need for a deeper understanding of their potential in tackling structured, constraint-intensive problems---a pursuit currently limited by the absence of comprehensive benchmarks for systematic investigation. To address this, we introduce CO-Bench, a benchmark suite featuring 36 real-world CO problems drawn from a broad range of domains and complexity levels. CO-Bench includes structured problem formulations and curated data to support rigorous investigation of LLM agents. We evaluate multiple agentic frameworks against established human-designed algorithms, revealing the strengths and limitations of existing LLM agents and identifying promising directions for future research.

Qiyao Sun, Xingming Li, Xixiang He, Ao Cheng, Xuanyu Ji, Hailun Lu, Runke Huang, Qingyong Hu

Large language models (LLMs) have achieved remarkable success in various natural language processing tasks, yet they remain prone to generating factually incorrect outputs—known as "hallucinations". While recent approaches have shown promise for hallucination detection by repeatedly sampling from LLMs and quantifying the semantic inconsistency among the generated responses, they rely on fixed sampling budgets that fail to adapt to query complexity, resulting in computational inefficiency. We propose an Adaptive Bayesian Estimation framework for Semantic Entropy with Guided Semantic Exploration, which dynamically adjusts sampling requirements based on observed uncertainty. Our approach employs a hierarchical Bayesian framework to model the semantic distribution, enabling dynamic control of sampling iterations through variance-based thresholds that terminate generation once sufficient certainty is achieved. We also develop a perturbation-based importance sampling strategy to systematically explore the semantic space. Extensive experiments on four QA datasets demonstrate that our method achieves superior hallucination detection performance with significant efficiency gains. In low-budget scenarios, our approach requires about 50% fewer samples to achieve comparable detection performance to existing methods, while delivers an average AUROC improvement of 12.6% under the same sampling budget.

Kai Sun, Wenqiang Li, Bo Dong, Yuxin Lin, Jingyao Zhang, Bin Shi

Large language models (LLMs) are seeing growing adoption in multi-agent systems. In these systems, efficient failure attribution is critical for ensuring robustness and interpretability. Current LLM-based attribution methods often face challenges with lengthy logs and lacking expert knowledge. Drawing inspiration from human debugging strategies, we propose an automated failure attribution framework, Scope Delineation Before Localization, which operates in two key stages: (1) identifying the failure scope and (2) pinpointing the failure step. By decoupling failure attribution into the two stages, our approach alleviates the reasoning workload of LLMs, enabling more precise failure attribution. To support scope delineation, we further introduce two strategies: Stepwise Scope Delineation and Expertise-Assisted Scope Delineation. Experiments on the Who&When dataset validate the efficacy of our two-stage framework, demonstrating substantial improvements over prior methods (up to 24.27% on step-level accuracy).

Kai Sun, Yuxin Lin, Bo Dong, Jingyao Zhang, Bin Shi

The rapid development of large language models (LLMs) has relied on access to high-quality, large-scale datasets, yet growing concerns around data privacy and security have spurred substantial research into pre-training data detection. While state-of-the-art (SOTA) methods such as RECALL and CON-RECALL leverage auxiliary prefixes to enhance detection performance, their dependence on individual prefixes introduces notable instability across varying prefix conditions. To address this, we first conduct a theoretical analysis to assess the impact of prefixes on existing prefix-based methods. Building on the analysis, we propose a novel prefix selection method to identify optimal prefixes. Specifically, our method derives two key criteria Discriminability and Symmetry. These criteria serve to quantify the effectiveness of prefixes in detecting pre-training data, enabling precise selection of high-performing candidate prefixes. Experiments on the WikiMIA dataset demonstrate that our method consistently improves the performance of RECALL and CON-RECALL, achieving gains of up to 21.1% in AUC scores while significantly enhancing robustness.

Chenghao Sun, Zhen Huang, Yonggang Zhang, Xinmei Tian, Xu Shen, Jieping Ye

Large language models (LLMs) present a paradox: they can correctly answer a multi-hop factual query in a high-resource language like English, yet fail on the identical query in another language. This raises a fundamental question about the nature of multilingual knowledge: are facts missing, or merely inaccessible? The underlying mechanisms for this knowledge gap have remained largely unexplored. In this work, we resolve this question by introducing a mechanistic interpretability framework that traces the causal pathways of multi-hop knowledge reasoning. Our analysis reveals a core, non-obvious finding: cross-lingual inconsistencies do not stem from a knowledge deficit. Instead, factual knowledge is robustly stored in a set of **shared, language-agnostic semantic neurons**. The failure originates from **misaligned attention pathways**, where a common set of critical attention heads fails to correctly route information along the reasoning chain to the appropriate knowledge neurons in lower-resource languages. This mechanistic diagnosis motivates a targeted alignment strategy: a surgical fine-tuning of only these critical heads. Experiments demonstrate that our method achieves significant improvements in multilingual multi-hop factuality—with positive cross-lingual transfer—while uniquely preserving general model capabilities, offering a scalable and mechanistically-grounded approach to building more reliable multilingual models.

Yize Sui, Yan Xu, Kun Hu, Jing Ren, Wenjing Yang

Retrieval-Augmented Generation (RAG) enhances the quality of question answering by integrating external knowledge with internal knowledge. A robust RAG system needs to precisely regulate the dependence of the response on the two types of knowledge. The recently proposed context-aware contrastive decoding (CCD) method attempts to achieve this goal by adjusting the knowledge reference weights by comparing the output distribution differences of LLMS when they rely on different knowledge sources. However, these methods are based on probabilistic knowledge reference adjustment strategies (such as the highest probability or entropy), only focus on the relative confidence of the output responses at each decoding step, without considering the absolute confidence of the responses, which may lead to misjudgment of the external knowledge and internal knowledge reference degree in the decoding process. To this end, we propose a novel decoding method, Evidence-guided Contrastive Decoding (ECD), which conducts evidence modeling by constructing the Dirichlet distribution and regards logits as evidence vectors, so as to regulate the reference degree of internal and external knowledge more accurately, and finally improve the quality of generated responses. Extensive evaluations across four public benchmark datasets on three mainstream LLMs have demonstrated the effectiveness and advantages of ECD.

Puzhen Su, Haoran Yin, Miao Yongzhu, Jintao Tang, Shasha Li, Ting Wang

Detecting Alzheimer’s disease (AD) from narrative transcripts challenges large language models (LLMs): pre-training rarely covers this out-of-distribution task, and all transcript demos describe the same scene, producing highly homogeneous contexts. These factors cripple both the model’s built-in task knowledge (task cognition) and its ability to surface subtle, class-discriminative cues (contextual perception). Because cognition is fixed after pre-training, improving in-context learning (ICL) for AD detection hinges on enriching perception through better demonstration (demo) sets. We demonstrate that standard ICL quickly saturates, its demos lack diversity (context width) and fail to convey fine-grained signals (context depth), and that recent task vector (TV) approaches improve broad task adaptation by injecting TV into the LLMs' hidden states (HSs), they are ill-suited for AD detection due to the mismatch of injection granularity, strength and position. To address these bottlenecks, we introduce DA4ICL, a demo-centric anchoring framework that jointly expands context width via Diverse and Contrastive Retrieval (DCR) and deepens each demo's signal via Projected Vector Anchoring (PVA) at every Transformer layer. Across three AD benchmarks, DA4ICL achieves large, stable gains over both ICL and TV baselines, charting a new paradigm for fine-grained, OOD and low-resource LLM adaptation.

Jiayuan Su, Fulin Lin, Zhaopeng Feng, Han Zheng, Teng Wang, Zhenyu Xiao, Xinlong Zhao, Zuozhu Liu, Lu Cheng, Hongwei Wang

Recent advances in large reasoning models (LRMs) have significantly enhanced long-chain reasoning capabilities over standard large language models (LLMs). However, LRMs often produce unnecessarily lengthy outputs even for simple queries, leading to inefficiencies or even accuracy degradation compared to LLMs. To address this, we propose CP-Router, a training-free, model-agnostic routing framework that dynamically selects between an LLM and an LRM, demonstrated with multiple-choice question answering (MCQA) prompts. The routing decision is guided by the prediction uncertainty estimates derived via Conformal Prediction (CP), which provides rigorous coverage guarantees. To improve uncertainty differentiation across inputs, we introduce Full and Binary Entropy (FBE), a novel entropy-based criterion that adaptively selects the appropriate CP threshold. Experiments across MCQA and QA benchmarks—including mathematics, logical reasoning, and Chinese chemistry—demonstrate that CP-Router efficiently reduces token usage while maintaining or even improving accuracy compared to using LRM alone. We further demonstrate the generality and robustness of CP-Router by extending it to diverse model pairings beyond the LLM–LRM setting.

Dongyu Su, Yimin Xiao, Tongguan Wang, Feiyue Xue, Junkai Li, Hui Liu, Ying Sha

The Two-Part Allegorical Saying (TPAS) is a Chinese linguistic phenomenon with a riddle-explanation structure, and an important component of Chinese metaphors. Existing research has primarily used TPAS to assist other semantic tasks, but lacks in-depth exploration of its intrinsic mechanisms: semantic rhetoric, logical reasoning, and metaphorical expression. To address this gap, we construct the first Chinese TPAS Reading Comprehension dataset (CTRC), which contains 18,103 TPASs and 75,296 passages. We frame it as a cloze test where the model selects the most suitable TPAS from candidates to fill passage blanks. To tackle the challenges of this CTRC task, we propose a Multi-view TPAS Contrastive Learning Network (MTCLN). Firstly, the joint vector cross-projection module extracts the rhetorical features of TPAS, such as homophonic puns, through vector space mapping to mitigate the semantic deviations caused by rhetoric. Then, the softened contrastive learning module strengthens the modeling of TPAS logical reasoning through feature association. Finally, the multi-view feature fusion module integrates contextual semantics with diverse TPAS features to facilitate the understanding of metaphorical expressions. Experiments on the CTRC dataset demonstrate that MTCLN achieves an average accuracy of 67.47%, outperforming large language models by 25.48%.

Zichen Song, Qixin Zhang, Ming Li, Yao Shu

The proliferation of Large Language Models (LLMs) has raised concerns over training data privacy. Membership Inference Attacks (MIA), aiming to identify whether specific data was used for training, pose significant privacy risks. However, existing MIA methods struggle to address the scale and complexity of modern LLMs. This paper introduces OR-MIA, a novel MIA framework inspired by model optimization and input robustness. First, training data points are expected to exhibit smaller gradient norms due to optimization dynamics. Second, member samples show greater stability, with gradient norms being less sensitive to controlled input perturbations. OR-MIA leverages these principles by perturbing inputs, computing gradient norms, and using them as features for a robust classifier to distinguish members from non-members. Evaluations on LLMs (70M to 6B parameters) and various datasets demonstrate that OR-MIA outperforms existing methods, achieving over 90% accuracy. Our findings highlight a critical vulnerability in LLMs and underscore the need for improved privacy-preserving training paradigms.

Yuerong Song, Xiaoran Liu, Ruixiao Li, Zhigeng Liu, Zengfeng Huang, Qipeng Guo, Ziwei He, Xipeng Qiu

Diffusion Large Language Models (dLLMs) enable breakthroughs in reasoning and parallel decoding but suffer from prohibitive quadratic computational complexity and memory overhead during inference. Current caching techniques accelerate decoding by storing full-layer states, yet impose substantial memory usage that limit long-context applications. Our analysis of attention patterns in dLLMs reveals persistent cross-layer sparsity, with pivotal tokens remaining salient across decoding steps and low-relevance tokens staying unimportant, motivating selective cache eviction. We propose Sparse-dLLM, the first training-free framework integrating dynamic cache eviction with sparse attention via delayed bidirectional sparse caching. By leveraging the stability of token saliency over steps, it retains critical tokens and dynamically evicts unimportant prefix/suffix entries using an attention-guided strategy. Extensive experiments on LLaDA and Dream series demonstrate Sparse-dLLM achieves up to 10 times higher throughput than vanilla dLLMs, with comparable performance and similar peak memory costs, outperforming previous methods in efficiency and effectiveness.

Seokwon Song, Minsu Park, Gunhee Kim

Source attribution aims to enhance the reliability of AI-generated answers by including references for each statement, helping users validate the provided answers. However, existing work has primarily focused on text-only scenario and largely overlooked the role of multimodality. We introduce MAVIS, the first benchmark designed to evaluate multimodal source attribution systems that understand user intent behind visual questions, retrieve multimodal evidence, and generate long-form answers with citations. Our dataset comprises 157K visual QA instances, where each answer is annotated with fact-level citations referring to multimodal documents. We develop fine-grained automatic metrics along three dimensions of informativeness, groundedness, and fluency, and demonstrate their strong correlation with human judgments. Our key findings are threefold: (1) LVLMs with multimodal RAG generate more informative and fluent answers than unimodal RAG, but they exhibit weaker groundedness for image documents than for text documents, a gap amplified in multimodal settings. (2) Given the same multimodal documents, there is a trade-off between informativeness and groundedness across different prompting methods. (3) Our proposed method highlights mitigating contextual bias in interpreting image documents as a crucial direction for future research.

Ruike Song, Zeen Song, Huijie Guo, Wenwen Qiang

External reasoning systems combine language models with process reward models (PRMs) to select high-quality reasoning paths for complex tasks such as mathematical problem solving. However, these systems are prone to reward hacking, where high-scoring but logically incorrect paths are assigned high scores by the PRMs, leading to incorrect answers. From a causal inference perspective, we attribute this phenomenon primarily to the presence of confounding semantic features. To address it, we propose Causal Reward Adjustment (CRA), a method that mitigates reward hacking by estimating the true reward of a reasoning path. CRA trains sparse autoencoders on the PRM’s internal activations to recover interpretable features, then corrects confounding by using backdoor adjustment. Experiments on math solving datasets demonstrate that CRA mitigates reward hacking and improves final accuracy, without modifying the policy model or retraining PRM.

Aarush Sinha, Pavan Kumar S, Roshan Balaji, Nirav Pravinbhai Bhatt

Hard negatives are essential for training effective retrieval models. Hard-negative mining typically relies on ranking documents using cross-encoders or static embedding models based on similarity metrics such as cosine distance. Hard negative mining becomes challenging for biomedical and scientific domains due to the difficulty in distinguishing between source and hard negative documents. However, referenced documents naturally share contextual relevance with the source document but are not duplicates, making them well-suited as hard negatives. In this work, we propose BiCA: Biomedical Dense Retrieval with Citation-Aware Hard Negatives, an approach for hard-negative mining by utilizing citation links in 20,000 PubMed articles for improving a domain-specific small dense retriever. We fine-tune the GTE_small and GTE_Base models using these citation-informed negatives and observe consistent improvements in zero-shot dense retrieval using nDCG@10 for both in-domain and out-of-domain tasks on BEIR and outperform baselines on long-tailed topics in LoTTE using Success@5. Our findings highlight the potential of leveraging document link structure to generate highly informative negatives, enabling state-of-the-art performance with minimal fine-tuning and demonstrating a path towards highly data-efficient domain adaptation.

Shuzheng Si, Haozhe Zhao, Cheng Gao, Yuzhuo Bai, Zhitong Wang, Bofei Gao, Kangyang Luo, Wenhao Li, Yufei Huang, Gang Chen 等

Teaching large language models (LLMs) to be faithful in the provided context is crucial for building reliable information-seeking systems. Therefore, we propose a systematic framework, CANOE, to reduce faithfulness hallucinations of LLMs across different downstream tasks without human annotations. Specifically, we first synthesize short-form question-answering (QA) data with four diverse tasks to construct high-quality and easily verifiable training data without human annotation. Also, we propose Dual-GRPO, a rule-based reinforcement learning method that includes three tailored rule-based rewards derived from synthesized short-form QA data, while simultaneously optimizing both short-form and long-form response generation. Notably, Dual-GRPO eliminates the need to manually label preference data to train reward models and avoids over-optimizing short-form generation when relying only on the synthesized short-form QA data. Experimental results show that CANOE greatly improves the faithfulness of LLMs across 11 different tasks, even outperforming the most advanced LLMs, e.g., GPT-4o and OpenAI o1.