Automated fact-checking aims to assess the truthfulness of textual claims based on relevant evidence. However, verifying complex claims that require multi-hop reasoning remains a significant challenge. We propose GraphCheck, a novel framework that transforms claims into entity-relationship graphs for structured and systematic fact-checking. By explicitly modeling both explicit and latent entities and exploring multiple reasoning paths, GraphCheck enhances verification robustness. While GraphCheck excels in complex scenarios, it may be unnecessarily elaborate for simpler claims. To address this, we introduce DP-GraphCheck, a variant that employs a lightweight strategy selector to choose between direct prompting and GraphCheck adaptively. This selective mechanism improves both accuracy and efficiency by applying the appropriate level of reasoning to each claim. Experiments on the HOVER and EX-FEVER datasets demonstrate that our approach outperforms existing methods in verification accuracy, while achieving strong computational efficiency despite its multipath exploration. Moreover, the strategy selection mechanism in DP-GraphCheck generalizes well to other fact-checking pipelines, highlighting the broad applicability of our framework.
论文检索
输入标题、作者或关键词，从 7,591 篇学术成果中精准定位
ToolHaystack: Stress-Testing Tool-Augmented Language Models in Realistic Long-Term Interactions
PDF ↗Large language models (LLMs) have demonstrated strong capabilities in using external tools to address user inquiries. However, most existing evaluations assume tool use in short contexts, offering limited insight into model behavior during realistic long-term interactions. To fill this gap, we introduce ToolHaystack, a benchmark for testing the tool use capabilities in long-term interactions. Each test instance in ToolHaystack includes multiple tasks execution contexts and realistic noise within a continuous conversation, enabling assessment of how well models maintain context and handle various disruptions. By applying this benchmark to 14 state-of-the-art LLMs, we find that while current models perform well in standard multi-turn settings, they often significantly struggle in ToolHaystack, highlighting critical gaps in their long-term robustness not revealed by previous tool benchmarks.
Research on bias in Text-to-Image (T2I) models has primarily focused on demographic representation and stereotypical attributes, overlooking a fundamental question: how does grammatical gender influence visual representation across languages? We introduce a cross-linguistic benchmark examining words where grammatical gender contradicts stereotypical gender associations (e.g., “une sentinelle” - grammatically feminine in French but referring to the stereotypically masculine concept “guard”). Our dataset spans five gendered languages (French, Spanish, German, Italian, Russian) and two gender-neutral control languages (English, Chinese), comprising 800 unique prompts that generated 28,800 images across three state-of-the-art T2I models. Our analysis reveals that grammatical gender dramatically influences image generation: masculine grammatical markers increase male representation to 73% on average (compared to 22% with gender-neutral English), while feminine grammatical markers increase female representation to 38% (compared to 28% in English). These effects vary systematically by language resource availability and model architecture, with high-resource languages showing stronger effects. Our findings establish that language structure itself, not just content, shapes AI-generated visual outputs, introducing a new dimension for understanding bias and fairness in multilingual, multimodal systems.
LLM-as-Judge frameworks are increasingly popular for AI evaluation, yet research findings on the relationship between models’ generation and judgment abilities remain inconsistent. We investigate this relationship through systematic dataset- and instance-level analyses across 11 models and 21 diverse tasks. Despite both capabilities relying on the same underlying knowledge, our analyses reveal they are only weakly correlated, primarily due to LLMs’ sensitivity to the responses being judged. To address this, we propose a self-reference-guided evaluation strategy that leverages a model’s own answers as references. This approach significantly strengthens the correlation between generation and judgment abilities, offering a practical path to align these skills and providing a reliable proxy for model selection in evaluation tasks.
Large Language Models (LLMs) often exhibit deficiencies with complex reasoning tasks, such as maths, which we attribute to the discrepancy between human reasoning patterns and those presented in the LLMs’ training data. When dealing with complex problems, humans tend to think carefully before expressing solutions. However, they often do not articulate their inner thoughts, including their intentions and chosen methodologies. Consequently, critical insights essential for bridging reasoning steps may be absent in training data collected from human sources. To bridge this gap, we proposes inserting insights between consecutive reasoning steps, which review the status and initiate the next reasoning steps. Unlike prior prompting strategies that rely on a single or a workflow of static prompts to facilitate reasoning, insights are proactively generated to guide reasoning processes. We implement our idea as a reasoning framework, named Thinking Before You Speak (TBYS), and design a pipeline for automatically collecting and filtering in-context examples for the generation of insights, which alleviates human labeling efforts and fine-tuning overheads. Experiments on challenging mathematical datasets verify the effectiveness of TBYS. Project website: https://gitee.com/jswrt/TBYS
QEVA: A Reference-Free Evaluation Metric for Narrative Video Summarization with Multimodal Question Answering
PDF ↗Video-to-text summarization remains underexplored in terms of comprehensive evaluation methods. Traditional n-gram overlap-based metrics and recent large language model (LLM)-based approaches depend heavily on human-written reference summaries, limiting their practicality and sensitivity to nuanced semantic aspects. In this paper, we propose QEVA, a reference-free metric evaluating candidate summaries directly against source videos through multimodal question answering. QEVA assesses summaries along three clear dimensions: Coverage, Factuality, and Temporal Coherence. We also introduce MLVU(VS)-Eval, a new annotated benchmark derived from the MLVU dataset, comprising 800 summaries generated from 200 videos using state-of-the-art video-language multimodal models. This dataset establishes a transparent and consistent framework for evaluation. Experimental results demonstrate that QEVA shows higher correlation with human judgments compared to existing approaches, as measured by Kendall’s \tau_b, \tau_c, and Spearman’s \rho. We hope that our benchmark and metric will facilitate meaningful progress in video-to-text summarization research and provide valuable insights for the development of future evaluation methods.
While large language models (LLMs) demonstrate impressive capabilities, their reliance on parametric knowledge often leads to factual inaccuracies. Retrieval-Augmented Generation (RAG) mitigates this by leveraging external documents, yet existing approaches treat retrieved passages as isolated chunks, ignoring valuable structure that is crucial for document organization. Motivated by this gap, we propose Retrieve-DocumentRoute-Read (RDR2), a novel framework that explicitly incorporates structural information throughout the RAG process. RDR2 employs an LLM-based router to dynamically navigate document structure trees, jointly evaluating content relevance and hierarchical relationships to assemble optimal evidence. Our key innovation lies in formulating document routing as a trainable task, with automatic action curation and structure-aware passage selection inspired by human reading strategies. Through comprehensive evaluation on five challenging datasets, RDR2 achieves state-of-the-art performance, demonstrating that explicit structural awareness significantly enhances RAG systems’ ability to acquire and utilize knowledge, particularly in complex scenarios requiring multi-document synthesis.
Forget the Unneeded: Backdooring Large Language Models via Contrastive-enhanced Machine Unlearning
PDF ↗Prompt tuning for Large Language Models (LLMs) is vulnerable to backdoor attacks. Existing methods find backdoor attacks to be a significant threat in data-rich scenarios. However, in data-limited scenarios, these methods have difficulty capturing precise backdoor patterns, leading to weakened backdoor attack capabilities and significant side effects for the LLMs, which limits their practical relevance. To explore this problem, we propose a backdoor attacks through contrastive-enhanced machine unlearning in data-limited scenarios, called BCU. Specifically, BCU introduces a multi-objective machine unlearning method to capture precise backdoor patterns by forgetting the association between non-trigger data and the backdoor patterns, reducing side effects. Moreover, we design a contrastive learning strategy to enhance the association between triggers and backdoor patterns, improving the capability of backdoor attacks. Experimental results on 6 NLP datasets and 4 LLMs show that BCU exhibits strong backdoor attack capabilities and slight side effects, whether the training data is rich or limited. Our findings highlight practical security risks of backdoor attacks against LLMs, necessitating further research for security purposes. Our code is available at https://github.com/AHU-YangSJ/BCU.
Word Meaning Negotiations (WMN) are sequences in conversation where speakers collectively discuss and shape word meaning. These exchanges can provide insight into conversational dynamics and word-related misunderstandings, but they are hard to find in corpora. In order to facilitate data collection and speed up the WMN annotation process, we introduce the task of detecting WMN indicators – utterances where a speaker signals the need to clarify or challenge word meaning. We train a wide range of models and reveal the difficulty of the task. Our models have better precision than previous regular-expression based approaches and show some generalization abilities, but have moderate recall. However, this constitutes a promising first step toward an iterative process for obtaining more data.
Entity resolution is a fundamental problem in data management that aims to identify all duplicate entries within collections of multi-attribute tuples. Most existing works focus on supervised learning, relying on large amounts of high-quality labeled data, including both positive and negative tuple pairs that are meticulously prepared. However, in reality, the manual annotation process is labor-intensive; in particular, selecting high-quality negative data for labeling is both important and challenging. In this paper, we propose an end-to-end ER solution, PUER, to address low-resource entity resolution (ER) by leveraging Large Language Models (LLMs) in a Positive-Unlabeled (PU) learning setting, where only a small number of positively labeled examples, e.g., 50, and unlabeled data are provided. Unlike directly fine-tuning LLMs in a supervised manner, we solve the entity matching task using reinforcement learning and propose a self-adaptive reward function in the process of RL. To enhance performance, we design an iterative workflow based on the co-training mechanism that fully utilizes entity blocking component to assist the entity matching. This workflow aims to improve the robustness and quality of pseudo-labels so that the performance of entity matching improves. Comprehensive experimental results on various benchmark datasets demonstrate the superiority of PUER. Full version and code are available.
Large language models (LLMs) exhibit remarkable versatility in adopting diverse personas. In this study, we examine how assigning a persona influences a model’s reasoning on an objective task. Using activation patching, we take a first step toward understanding how key components of the model encode persona-specific information. Our findings reveal that the early Multi-Layer Perceptron (MLP) layers attend not only to the syntactic structure of the input but also process its semantic content. These layers transform persona tokens into richer representations, which are then used by the middle Multi-Head Attention (MHA) layers to shape the model’s output. Additionally, we identify specific attention heads that disproportionately attend to racial and color-based identities.
Large language models (LLMs) face challenges with internal knowledge inaccuracies and outdated information. Knowledge editing has emerged as a pivotal approach to mitigate these issues. Although current knowledge editing techniques exhibit promising performance in single-hop reasoning tasks, they show limitations when applied to multi-hop reasoning. Drawing on cognitive neuroscience and the operational mechanisms of LLMs, we hypothesize that the residual single-hop knowledge after editing causes edited models to revert to their original answers when processing multihop questions, thereby undermining their performance in multi-hop reasoning tasks. To validate this hypothesis, we conduct a series of experiments that empirically confirm our assumptions. Building on the validated hypothesis, we propose a novel knowledge editing method that incorporates a Knowledge Erasure mechanism for Large language model Editing (KELE). Specifically, we design an erasure function for residual knowledge and an injection function for new knowledge. Through joint optimization, we derive the optimal recall vector, which is subsequently utilized within a rank-one editing framework to update the parameters of targeted model layers. Extensive experiments on GPT-J (6B) and LLaMA-2 (7B) demonstrate that KELE substantially enhances the multi-hop reasoning capability of edited LLMs.
Dense retrieval models usually adopt vectors from the last hidden layer of the document encoder to represent a document, which is in contrast to the fact that representations in different layers of a pre-trained language model usually contain different kinds of linguistic knowledge, and behave differently during fine-tuning. Therefore, we propose to investigate utilizing representations from multiple encoder layers to make up the representation of a document, which we denote Multi-layer Representations (MLR). We first investigate how representations in different layers affect MLR’s performance under the multi-vector retrieval setting, and then propose to leverage pooling strategies to reduce multi-vector models to single-vector ones to improve retrieval efficiency. Experiments demonstrate the effectiveness of MLR over dual encoder, ME-BERT and ColBERT in the single-vector retrieval setting, as well as demonstrate that it works well with other advanced training techniques such as retrieval-oriented pre-training and hard negative mining.
Most recommender systems focus on short-term objectives such as click-through rate, often at the expense of long-term user satisfaction. This can lead to echo chambers, where users are repeatedly exposed to redundant content. While recent efforts integrate Large Language Models (LLMs) into recommendation, they typically inherit this short-sighted focus. In this work, we highlight unpaired feedback—implicit signals such as continued engagement (positive) or silent disengagement (negative) that lack explicit contrastive labels—as a key challenge for long-term recommendation. Effectively learning from such feedback is crucial for improving LLM-based recommenders in dynamic user environments. To this end, we propose ULRec (Unpaired Feedback for Long-Term LLM-based Recommendation Tuning), a simple framework that fine-tunes LLMs using both positive and negative unpaired feedback. ULRec leverages the KTO algorithm to incorporate these signals without requiring paired supervision. Despite its simplicity, ULRec consistently improves long-term recommendation performance, demonstrating the value of modeling unpaired user feedback.
Decision-making tasks have highlighted fundamental challenges in grounding decisions within real-world contexts. Traditional decision knowledge utilization methods often struggle to effectively integrate structured decision constraints, limiting their ability to decompose high-level tasks, maintain logical consistency, and adapt to dynamic environments. To bridge this gap, we introduce StructuThink, a knowledge-structured reasoning framework that enhances LLM-based agents with explicit decision constraints. Specifically, we propose the Task Transition Knowledge Graph (TTKG) that learning decision knowledge in embodied scenarios. Leveraging this knowledge, we propose the StructuThink framework, comprising a subtask chain constructor for grounding natural language instructions and a constraint-based executor for adaptive and consistent decision-making. We validate StructuThink across multiple benchmarks, including ALFWorld and WebShop, where it achieves higher task success rates (improving by up to 7%) and more efficient action sequences (requiring up to 15% fewer steps) than baseline methods. Our approach enables LLMs to more effectively ground decision-making in domain-specific scenarios, enhancing both interpretability and reliability, thus paving the way for more reliable and adaptable decision-making systems.
Guaranteed Guess: A Language Modeling Approach for CISC-to-RISC Transpilation with Testing Guarantees
PDF ↗The hardware ecosystem is rapidly evolving, with increasing interest in translating low-level programs across different *instruction set architectures* (ISAs) in a quick, flexible, and correct way to enhance the portability and longevity of existing code. A particularly challenging class of this transpilation problem is translating between complex- (CISC) and reduced- (RISC) hardware architectures, due to fundamental differences in instruction complexity, memory models, and execution paradigms. In this work, we introduce GG (**G**uaranteed **G**uess), an ISA-centric transpilation pipeline that combines the translation power of pre-trained large language models (LLMs) with the rigor of established software testing constructs. Our method generates candidate translations using an LLM from one ISA to another, and embeds such translations within a software-testing framework to build quantifiable confidence in the translation. We evaluate our GG approach over two diverse datasets, enforce high code coverage (>98%) across unit tests, and achieve functional/semantic correctness of 99% on HumanEval programs and 49% on BringupBench programs, respectively. Further, we compare our approach to the state-of-the-art Rosetta 2 framework on Apple Silicon, showcasing 1.73\times faster runtime performance, 1.47\times better energy efficiency, and 2.41\times better memory usage for our transpiled code, demonstrating the effectiveness of GG for real-world CISC-to-RISC translation tasks. We will open-source our codes, data, models, and benchmarks to establish a common foundation for ISA-level code translation research.
Large language models (LLMs) develop the in-context learning capability through pretraining and instruction tuning, enabling task adaptation without parameter updates. Self-refinement is a manifestation of this capability, which allows LLMs to iteratively refine the output using self-generated feedback. However, empirical observations reveal Inference-Free Self-Refinement (IFSR) in preference alignment: LLMs generate preference-improved output via fixed instructions, requiring no specific feedback, even no initial responses. There are two key components of the IFSR in preference alignment. The refining instruction is a fixed instruction that constrains the output distribution from a preference-semantic perspective. During training, it facilitates joint learning of preference-related semantic representations and data distribution alignment. The pseudo reference response is constructed from paired preference data and serves as a demonstration to guide the output distribution. It mitigates off-policy distributional bias while enhancing token-level preference learning in training. Experiments across multiple datasets demonstrate that incorporating IFSR into preference alignment yields performance improvement over 10%. Further ablation studies reveal additional characteristics and potential principles of IFSR.
AgentDrug: Utilizing Large Language Models in an Agentic Workflow for Zero-Shot Molecular Editing
PDF ↗Molecular editing—modifying a given molecule to improve desired properties—is a fundamental task in drug discovery. While LLMs hold the potential to solve this task using natural language to drive the editing, straightforward prompting achieves limited accuracy. In this work, we propose AgentDrug, an agentic workflow that leverages LLMs in a structured refinement process to achieve significantly higher accuracy. AgentDrug defines a nested refinement loop: the inner loop uses feedback from cheminformatics toolkits to validate molecular structures, while the outer loop guides the LLM with generic feedback and a gradient-based objective to steer the molecule toward property improvement. We evaluate AgentDrug on benchmarks with both single- and multi-property editing under loose and strict thresholds. Results demonstrate significant performance gains over previous methods. With Qwen-2.5-3B, AgentDrug improves accuracy by 20.7% (loose) and 16.8% (strict) on six single-property tasks, and by 7.0% and 5.3% on eight multi-property tasks. With larger model Qwen-2.5-7B, AgentDrug further improves accuracy on 6 single-property objectives by 28.9% (loose) and 29.0% (strict), and on 8 multi-property objectives by 14.9% (loose) and 13.2% (strict).
The quality of a conversation goes beyond the individual quality of each reply, and instead emerges from how these combine into interactional dynamics that give the conversation its distinctive overall “shape”. However, there is no robust automated method for comparing conversations in terms of their overall dynamics. Such methods could enhance the analysis of conversational data and help evaluate conversational agents more holistically.In this work, we introduce a similarity measure for comparing conversations with respect to their dynamics. We design a validation procedure for testing the robustness of the metric in capturing differences in conversation dynamics and for assessing its sensitivity to the topic of the conversations. To illustrate the measure’s utility, we use it to analyze conversational dynamics in a large online community, bringing new insights into the role of situational power in conversations.
Large Reasoning Models (LRMs) demonstrate strong performance on complex tasks through chain-of-thought (CoT) reasoning. However, they suffer from high inference latency due to lengthy reasoning chains. In this paper, we propose SpecCoT, a collaborative framework that combines large and small models for effective yet efficient reasoning. Unlike traditional speculative decoding, which operates at the token level, SpecCoT adopts a step-level verification strategy: the large model first establishes the reasoning direction, and for each intermediate step, the small model generates multiple candidate drafts in parallel. The large model then verifies these drafts, either selecting the most suitable one or rejecting them all and generating its own. SpecCoT approach balances reasoning quality with inference efficiency through fine-grained model cooperation. Experiments across diverse tasks show SpecCoT reduces inference latency by 1.7-4.1\times while maintaining comparable accuracy to standard large model inference.