论文检索

Dongsheng Chen, Yingqi Zhu, Xingyue Zhang, Wenqing Zhou, Lei Li

Despite the remarkable generation capabilities demonstrated by large language models (LLMs), the issue of hallucination remains a critical challenge. This is largely attributed to the models’ tendency to fit spurious dependencies in pre-training data rather than underlying causal logic. To address this, from an information-theoretic perspective, this paper proposes a unified contrastive decoding framework based on dynamic pointwise mutual information (Dynamic PMI). Under this framework, we design three fine-grained input transformation strategies targeting context, syntax, and semantics to construct dynamic background distributions. These strategies systematically disentangle and suppress spurious dependencies induced by context priors, lexical co-occurrences, and syntactic structures, thereby guiding the model to prioritize underlying causal logic. Experiments on extensive discriminative and generative benchmarks demonstrate that our method significantly improves the model’s factuality and reasoning robustness. Notably, despite employing a single-model architecture, our framework surpasses state-of-the-art dual-model strategies while maintaining high computational efficiency. Furthermore, the framework exhibits strong cross-model generalizability and effectively alleviates the over-refusal tendency in open-ended generation.

Rose Sathyanathan, Kinshuk Vasisht, Danish Pruthi

Millions of users turn to AI models for their information needs. It is conceivable that a large number of user queries contain assumptions that may be factually inaccurate. Prior work notes that large language models (LLMs) often fail to challenge such erroneous assumptions, and can reinforce users’ misinformed opinions. However, given the recent advances, especially in model’s reasoning capabilities, we revisit whether large reasoning models (LRMs) can reason about the underlying assumptions and respond to user queries appropriately. We construct queries with varying degrees of presuppositions spanning health, science, and general knowledge, and use it to evaluate several widely-deployed models When compared to non-reasoning models, we find that reasoning models achieve a slightly higher accuracy (2-11%), but they still fail to challenge a large fraction (26-42%) of false presuppositions. Further, reasoning models remain susceptible to how strongly the presupposition is expressed.

Chau Minh Pham, Zichao Wang, Puneet Mathur, Alexa Siu, Akriti Jain, Aparna Garimella, Ananya B. Sai, Nedim Lipka, Mohit Iyyer, Varun Manjunatha

Large language models are increasingly used to draft long-form multimodal documents, but their end-to-end performance on professional report generation remains systematically understudied. We introduce AnalystBench, a continually extensible benchmark of 20 real-world report generation tasks grounded in multimodal document collections, where models must process millions of input tokens to produce long-form professional reports. Using expert-validated quality checklists and groundedness evaluation, we evaluate LLMs and coding agents and find that the best model, GPT-5.1, scores highly on executive summarization tasks (exceeding 90% on quality checklists) but degrades substantially on tasks requiring long-horizon synthesis over large inputs (dropping to 25-40%). Agent-based generation substantially benefits strong closed-source models such as GPT-5.1, with checklist scores improving by 20.24 percentage points and visual coverage by 37.41 points over vanilla generation, but offers little or negative gains for open-source models like DeepSeek-R1 (-3.02 points). Expert reviewers note that while generated reports are grounded and clearly separate factual description from interpretation, they often fall short in actionability, clarity, and quantitative precision, which highlights the gap between system performance and real-world professional needs.

Xiaoyong Mei, Tingting Zuo, Da Chen, Guangyu Hu, Xiangyu Wen, Chao Duan, Mingyan Zhang, Fudan Zheng

Multi-role dialogue summarization requires modeling complex interactions among multiple speakers while preserving role-specific information and factual consistency. However, most existing methods optimize for automatic metrics such as ROUGE and BERTScore, which favor surface-level imitation of references rather than genuine gains in faithfulness or alignment with human preferences. We propose a novel framework that couples explicit cognitive-style reasoning with reward-based optimization for multi-role dialogue summarization. Our method first distills structured reasoning traces (e.g., step-by-step inferences and intermediate reflections) from a large teacher model and uses them as auxiliary supervision to initialize a reasoning-aware summarizer via staged supervised fine-tuning. It then applies GRPO with a dual-principle reward that blends metric-based signals with human-aligned criteria targeting key information coverage, implicit inference, factual faithfulness, and conciseness. Experiments on multilingual multi-role dialogue benchmarks show that our method matches strong baselines on ROUGE and BERTScore. Specifically, results on CSDS confirm the framework’s stability in semantic consistency, while in-depth analysis on SAMSum demonstrates clear gains in factual faithfulness and model-based preference alignment. These findings underscore the value of reasoning-aware and preference-aware training for reliable dialogue summarization. Code will be made accessible upon acceptance, checkpoints and datasets are now available at https://huggingface.co/NebulaPixel.

Yule Xie, Jiaxin Ding, Cheng Deng, Shiqing Gao, Junran Zhang, Sibo Zhang, Zeyuan Wang, Ke Wu, Xin Ding, Luoyi Fu 等

Reinforcement learning (RL) has recently shown remarkable ability to enhance reasoning in large language models (LLMs), yet its potential in scientific domains beyond mathematics remains largely unexplored. Geoscience questions couple broad factual knowledge with multi-step inference and often rely on visual evidence such as maps, cross-sections, and diagrams, making them a challenging but verifiable testbed for RL-based reasoning. To enable this study, we introduce GeoMC-10K, a dataset of 10,000 geoscience multiple-choice questions spanning physical to human geography and high-school to professional levels; over 30% of the questions are image dependent. To support text-only RL on these multimodal questions, we design GeoM2T, a multi-agent framework that converts multimodal questions into descriptive text while preserving answerability and difficulty. Fine-tuning LLaMA-3.1-8B and Qwen-3-8B with Group Relative Policy Optimization (GRPO), incorporating a factual reward mechanism, yields GR1, which achieves absolute accuracy improvements of 5.9% and 13.3%, respectively, and it generalizes to out-of-distribution geoscience benchmarks. Together, GeoMC-10K, GeoM2T, and GR1 establish a scalable benchmark and baseline for RL-enhanced geoscience reasoning.

Yuchen Wu, Liang Ding, Li Shen, Dacheng Tao

Aligning Large Language Models (LLMs) to be faithful to new knowledge in complex, multi-hop reasoning tasks is a critical, yet unsolved, challenge. We find that SFT-based methods, e.g., Reason-KE, while state-of-the-art, suffer from a "faithfulness gap": they optimize for format mimicry rather than sound reasoning. This gap enables the LLM’s powerful parametric priors to override new contextual facts, resulting in critical factual hallucinations (e.g., incorrectly reasoning "Houston" from "NASA" despite an explicit edit). To solve this core LLM alignment problem, we propose **Reason-KE++**, an SFT+RL framework that instills process-level faithfulness. Its core is a Stage-aware Reward mechanism that provides dense supervision for intermediate reasoning steps (e.g., Decomposition, Sub-answer Correctness). Crucially, we identify that naive outcome-only RL is a deceptive trap for LLM alignment: it collapses reasoning integrity (e.g., 19.00% Hop acc) while superficially boosting final accuracy. Our process-aware framework sets **a new SOTA of 95.48%** on MQUAKE-CF-3k (+5.28%), demonstrating that for complex tasks, aligning the reasoning process is essential for building trustworthy LLMs.

Lvhua Wu, Xuefeng Jiang, Sheng Sun, Yan Lei, Tian Wen, Yuwei Wang, Min Liu

The rapid spread of fake news threatens social stability and public trust, highlighting the urgent need for its effective detection.Although large language models (LLMs) show potential in fake news detection, they are limited by knowledge cutoff and easily generate factual hallucinations when handling time-sensitive news.Furthermore, the thinking of a single LLM easily falls into early stance locking and confirmation bias, making it hard to handle both content reasoning and fact checking simultaneously.To address these challenges, we propose ZoFia, a two-stage zero-shot fake news detection framework.In the first retrieval stage, we propose novel Hierarchical Salience and Salience-Calibrated Minimum Marginal Relevance (SC-MMR) algorithm to extract core entities accurately, which drive dual-source retrieval to overcome knowledge and evidence gaps.In the subsequent stage, a multi-agent system conducts multi-perspective reasoning and verification in parallel and achieves an explainable and robust result via adversarial debate.Comprehensive experiments on two public datasets show that ZoFia outperforms existing zero-shot baselines and even most few-shot methods.Our code has been open-sourced to facilitate the research community at https://github.com/SakiRinn/ZoFia.

Ke Yang, Shuguang Yuan, Jing Yu, Chi Chen

Retrieval-Augmented Generation (RAG) enhances the factual accuracy of Large Language Model (LLM) outputs based on external knowledge bases. These knowledge bases often carry significant intellectual property (IP) value, raising the urgent need for robust watermarking techniques to protect IP. However, existing RAG watermarking methods remain in their infancy, facing challenges such as limited encoding capacity and potential degradation of RAG performance or knowledge quality. In this paper, we propose knowledge-infused and multi-bit watermarking (KMW) for RAG knowledge bases. It generates watermark text to infuse the knowledge base by benign knowledge completion and a tailored generative watermarking algorithm. Each generated text can carry a multi-bit watermark segment. For effective detection, we design a Watermark Text Indexer that optimizes queries for steady retrieval of watermarked texts. Experiments on multiple datasets and LLMs show KMW reliably extracts watermarks from adversarial RAGs. It is robust against knowledge selection, alteration, expansion, and RAG setting restrictions, while remaining stealthy and secure. This highlights that KMW ensures effective IP protection for RAG systems. Our code is available here https://github.com/iieSKLCSDsg/KMW.

Yurun Song, Xiangqing Shen, Jianfei Yu, Rui Xia

While large language models (LLMs) have achieved remarkable success, their reliability in knowledge-intensive tasks is often compromised by factual hallucinations. Integrating Knowledge Graphs (KGs) addresses this issue; however, existing approaches typically rely on simple graph traversal.This paradigm decouples topological navigation from logical operations (e.g., temporal filtering, aggregation), leading to imprecise retrieval and heavy post-processing burdens.Although semantic parsing offers a solution by grounding reasoning in logical forms, it traditionally suffers from a dependency on scarce supervised annotations.To bridge this gap, we propose Interactive Semantic Parsing, a framework that formulates reasoning as the sequential generation of executable logical clauses. This design allows logical constraints to be dynamically interleaved with graph search, while optimizing via reinforcement learning with only final answer feedback eliminates the need for gold program annotations.To tackle the sparse reward challenge in the vast symbolic space, we introduce a distance-aware process reward to evaluate intermediate steps based on their topological proximity to the answer.Experimental results on WebQSP and CWQ demonstrate that our method achieves state-of-the-art performance, particularly on complex queries, validating the effectiveness of our dense reward signal in enabling robust reasoning without supervision.Our code is available at https://github.com/NUSTM/ISP-KGR.

Hongtao Duan, Lu Jiang, Minying Zhang, Xiaobing Zhu, Tianpeng Bu, Hao Jiang, Xinyu Wei, Lulu hu

Tool calling requires Large Language Models (LLMs) to generate structured decisions including tool names and schema-constrained arguments, where small decoding mistakes can cause hard failures. Existing methods either rely on costly tool-use training data or only constrain syntax, leaving tool selection and argument value errors largely unsolved. We analyze tool calling failures through a Where–When lens: (Where) failures correlate with persistent uncertainty in late transformer layers, (When) uncertainty concentrates on content-bearing tokens (tool names and argument values) rather than schema tokens. Based on this, and motivated by evidence that transformer Feed Forward Networks (FFNs) act as key–value style memories that store and retrieve factual or associative mappings, we propose Memory Space Tool Retracing (MemTR), a weight-free decoding-time method that retrieves relevant tool evidence from the tool library and mixes it into the FFN-output at the uncertain layer, treating FFNs as key–value memories. Through extensive experiments on various model families (Qwen, Llama, and xLAM) and benchmarks (BFCL, ACEBench, APIBank), MemTR reduces tool calling failures by 2%–9% with only 1%–2% runtime overhead, without any fine-tuning or additional tool-use training data.

Jin Wang, Kaiwen Luo, Liang Lin, Weiliu Wang, Yitian Chen, Moayad Aloqaily, Xuehai Tang, Zhenhong Zhou, Kun Wang, Li Sun 等

While Audio Large Language Models (ALLMs) have achieved remarkable progress in understanding and generation, their potential privacy implications remain largely unexplored. This paper takes the first step to investigate whether ALLMs inadvertently leak user privacy solely through acoustic voiceprints and introduces \textit{HearSay}, a comprehensive benchmark constructed from over 22,000 real-world audio clips. To ensure data quality, the benchmark is meticulously curated through a rigorous pipeline involving automated profiling and human verification, guaranteeing that all privacy labels are grounded in factual records. Extensive experiments on \textit{HearSay} yield three critical findings:\textbf{Significant Privacy Leakage}: ALLMs inherently extract private attributes from voiceprints, reaching 92.89% accuracy on gender and effectively profiling social attributes.\textbf{Insufficient Safety Mechanisms}: Alarmingly, existing safeguards are severely inadequate; most models fail to refuse privacy-intruding requests, exhibiting near-zero refusal rates for physiological traits.\textbf{Reasoning Amplifies Risk}: Chain-of-Thought (CoT) reasoning exacerbates privacy risks in capable models by uncovering deeper acoustic correlations.These findings expose critical vulnerabilities in ALLMs, underscoring the urgent need for targeted privacy alignment.The codes and dataset are available at https://github.com/JinWang79/HearSay_Benchmark

Ziming Li, Yu Tian, Tian Lan, Jiang Li, Zehua Duo, Guanglai Gao, Xiangdong Su

While Large Language Models (LLMs) excel at capturing communicative intent, this capability introduces a side effect: Pragmatic Hallucination, where models over-interpret literal contexts to generate non-factual inferences. To quantify this, we introduce the PaCE (Pragmatics-as-Context Evaluation) benchmark, comprising over 3,000 manually verified "context-flip" samples. Evaluations across nine mainstream models reveal a significant Context Sensitivity Gap (CSG), with literal accuracy consistently lagging behind pragmatic reasoning. Attribution analysis indicates that Reinforcement Learning from Human Feedback (RLHF) exacerbates this bias, and neither parameter scaling nor Chain-of-Thought (CoT) fully mitigates it. Crucially, "Strict Prompting" effectively reverses the CSG, demonstrating that the phenomenon stems from behavioral lock-in during training rather than inherent capability deficiencies. Furthermore, error patterns exhibit high systematic correlation across diverse architectures. This study highlights that current alignment paradigms lack precise control over pragmatic boundaries, underscoring the necessity for a "Literal Grounding" mechanism in future safety frameworks.

Langming Liu, Kangtao Lv, Haibin Chen, Weidong Zhang, Yejing Wang, Shilei Liu, Xin Tong, Yujin Yuan, Yongwei Wang, Wenbo Su 等

Large language models (LLMs), despite their powerful capabilities, suffer from factual hallucinations where they generate verifiable falsehoods. We identify a root of this issue: the imbalanced data distribution in the pretraining corpus, which leads to a state of "low-probability truth" and "high-probability falsehood". Recent approaches, such as teaching models to say "I don’t know" or post-hoc knowledge editing, either evade the problem or face catastrophic forgetting. To address this issue from its root, we propose PretrainRL, a novel framework that integrates reinforcement learning into the pretraining phase to consolidate factual knowledge. The core principle of PretrainRL is "debiasing then learning." It actively reshapes the model’s probability distribution by down-weighting high-probability falsehoods, thereby making "room" for low-probability truths to be learned effectively. To enable this, we design an efficient negative sampling strategy to discover these high-probability falsehoods and introduce novel metrics to evaluate the model’s probabilistic state concerning factual knowledge. Extensive experiments on three public benchmarks demonstrate that PretrainRL significantly alleviates factual hallucinations and outperforms state-of-the-art methods.

Jiaang Li, Zhendong Mao, Quan Wang, Yuning Wan, Yongdong Zhang

Retrieval-Augmented Generation (RAG) enhances the factuality of Large Language Models (LLMs) by incorporating retrieved documents and/or generated context. However, LLMs often exhibit a stylistic bias when presented with mixed contexts, favoring fluent but hallucinated generated content over factually grounded yet disorganized retrieved evidence. This phenomenon reveals that the utility of retrieved information is bottlenecked by its presentation. To bridge this gap, we propose QREAM, a style-controlled rewriter that aligns retrieved documents with a question-oriented style while preserving facts, better for LLM readers to utilize. Our framework consists of two stages: (1) QREAM-ICL, which uses stylistic seeds to guide iterative rewriting exploration; and (2) QREAM-FT, a lightweight student model distilled from denoised ICL outputs. QREAM-FT employs dual-criteria rejection sampling, filtering based on answer correctness and factual consistency to ensure high-quality supervision. QREAM seamlessly integrates into existing RAG pipelines as a plug-and-play module. Experiments demonstrate that QREAM consistently enhances advanced RAG pipelines, yielding up to 8% relative improvement with negligible latency overhead, effectively balancing question relevance with factual grounding.

Xiaoliang Xu, Huang Yuan, Junmei Wang, Can Xu

Automatic instruction generation offers a low-cost, high-efficiency pathway for fine-tuning large language models (LLMs). However, existing methods struggle in knowledge-intensive domains and complex reasoning tasks due to their dependence on high-quality seed data, limited coverage of single-document knowledge, and repetitive content. To overcome these limitations, this paper presents GCIG, a GraphRAG-based Cross-document Instruction Generation framework. We begin by constructing an enhanced knowledge graph to provide a structural representation of the raw corpus, followed by LLM-driven selection of reliable subgraph-text pairs based on factuality and logical complementarity. Subsequently, we adaptively generate diverse questions through task-aware prompts and context-sensitive retrieval. Finally, we employ Chain-of-Thought reasoning to anchor entity paths and integrate scattered evidence, thereby closing logical gaps and improving answer coherence. Experiments on knowledge-intensive and multi-hop question-answering tasks demonstrate that GCIG outperforms existing methods, producing instruction data with stronger logical consistency and broader knowledge coverage for effective LLM fine-tuning. The code and data are publicly available at https://github.com/WhitEiller/GCIG.

Haohan Yuan, Haopeng Zhang

Reasoning has substantially improved Large Language Models (LLMs) on analytical tasks such as mathematics and code generation, but its value for abstractive summarization remains unclear. To address this gap, we adapt general reasoning strategies to the summarization setting and conduct a large-scale comparative study of 8 reasoning strategies and 3 Large Reasoning Models (LRMs) across 8 diverse datasets, evaluating both summary quality and factual faithfulness. Our results show that reasoning is not a universal solution and its effectiveness depends strongly on the strategy and the summarization setting. In particular, we find a trade-off between summary quality and factual faithfulness. Explicit reasoning strategies often improve reference-based quality, but may weaken factual grounding, whereas implicit reasoning in LRMs shows the opposite tendency. We further find that increasing an LRM’s internal reasoning budget does not reliably improve summarization and can even reduce factual consistency. These findings suggest that, for summarization, more reasoning is not always better. Effective reasoning should preserve faithful compression rather than induce over-elaboration.

Yingtao Ren, Xiao Luo, Yu-Cheng Chang, Chin-teng Lin

Multi-step retrieval-augmented generation has attracted increasing attention due to its capacity to improve the factuality of large language models with iterative retrieved knowledge. However, the performance of multi-step RAG systems is susceptible to potential retrieval noise and fabricated documents in real-world scenarios. Current approaches usually utilize supervised fine-tuning on predetermined noisy contexts to enhance the robustness. However, their performance remains inadequate when it comes to more complicated long-context scenarios due to the lack of adaptability. Towards this end, we propose a novel framework named Context-attended Adversarial Reinforcement Learning (CARE) for multi-step RAG systems against attacks. The core of our CARE is to conduct reinforcement learning on adversarial samples which are alternatingly enhanced with text gradients. In particular, our CARE includes a reward model to identify the accuracy of responses, which is minimized for the generation of adversarial samples with text gradients. These context-attended noisy samples are then utilized for reinforcement learning to maximize the rewards. The whole framework is conducted alternatingly from easy to hard samples to ensure the smoothness of the optimization. Extensive experiments on multi-step RAG benchmark datasets are conducted to validate the superiority of our proposed CARE in multiple noisy scenarios. Our code is available at https://github.com/yingtaoren/CARE.

Yezi Liu, Hanning Chen, Wenjun Huang, Yang Ni, Mohsen Imani

Large Language Models (LLMs) encode vast factual knowledge, yet their inability to selectively forget specific information hinders privacy protection, bias mitigation, and post-deployment correction. We present LoRA-based Unlearning with Negative Examples (LUNE), a lightweight framework that performs negative-only unlearning by updating only low-rank adapters while freezing the backbone, thereby localizing edits and avoiding disruptive global changes. Leveraging Low-Rank Adaptation (LoRA), LUNE targets intermediate representations to suppress (or replace) requested knowledge with an order-of-magnitude lower compute and memory than full fine-tuning or direct weight editing. Extensive experiments on multiple factual unlearning tasks show that LUNE: (I) achieves effectiveness comparable to full fine-tuning and memory-editing methods; and (II) reduces computational cost by about an order of magnitude.

Guanyu Zheng, Wang Zhenyu, He Tingting, Xv Wang, Haochang Wang, Yaokai Huang, Tiejun Zhao

Lifelong knowledge editing aims to inject a stream of factual updates into large language models (LLMs) without retraining, yet existing memory-based editors often suffer from catastrophic forgetting as edits accumulate. We argue that a key factor is the coupled knowledge memory mechanism, where addressing (routing) and storage (writing via memory-module updates) are entangled. This entanglement makes it difficult to confine the effects of each edit to its intended scope, particularly in multi-domain and associated-fact editing streams, where updates either span diverse semantic domains or repeatedly modify related attributes of the same subject. Consequently, updating memory for one edit inadvertently alters the routing and stored representations of previously injected edits, leading to catastrophic forgetting as edits accumulate. We propose **DKME**, which decouples addressing from storage via two stages: decoupled semantic addressing learns a fact-aware manifold for scope-aware routing, and partitioned memory storage localizes edits to memory partitions identified by unsupervised clustering in the embedding space. Experiments on three benchmarks, including HalluEditBench, CKnowEdit, and WikiData_\text{counterfact}, demonstrate that DKME consistently achieves a more favorable trade-off between editing success and locality compared to baselines, while maintaining more stable performance as the edit scale increases.

Song-Li Wu, Zhaocheng Du, Xianquan Wang, Jingyi Wang

Lifelong knowledge editing (LKE) aims to incrementally correct factual inaccuracies in large language models (LLMs), but sequential edits can lead to substantial degradation of capabilities. Existing approaches primarily rely on static parameter regularization, which restricts knowledge integration and fails to prevent cumulative capability degradation. We argue that an important source of this degradation lies in the temporal mismatch between locally editable factual knowledge and procedural knowledge, which is gradually acquired, guides task execution, and cannot be reliably updated by rapid edits. To this end, we formulate LKE as a dual-timescale process, explicitly decoupling fast-updating factual knowledge from slow-evolving procedural knowledge. Based on this formulation, we propose CaPEdit, a framework that preserves model capabilities under LKE. It first synthesizes procedural knowledge across successive edits, and subsequently performs parameter updates guided jointly by factual supervision and the synthesized procedural signal. To ensure stability under long edit sequences, CaPEdit is trained via a hybrid optimization scheme, combining step-wise updates for rapid factual correction with trajectory-level optimization to facilitate gradual procedural adaptation. Experiments demonstrate that CaPEdit improves capability preservation across all fundamental capabilities by 49.78%, achieves superior editing performance, and requires only 18.07% of the editing time of most existing methods.