ACL 2025 论文列表

The evaluation of text generated by LLMs remains a challenge for question answering, retrieval augmented generation (RAG), summarization, and many other natural language processing tasks. Evaluating the factuality of LLM generated responses is particularly important in medical question answering, where the stakes are high. One method of evaluating the factuality of text is through the use of information nuggets (answer keys). Nuggets are text representing atomic facts that may be used by an assessor to make a binary decision as to whether the fact represented by said nugget is contained in an answer. Although manual nugget extraction is expensive and time-consuming, recent RAG shared task evaluations have explored automating the nuggetization of text with LLMs. In this work, we explore several approaches to nugget generation for medical question answering and evaluate their alignment with expert human nugget generation. We find providing an example and extracting nuggets from an answer to be the best approach to nuggetization. While, overall, we found the capabilities of LLMs to distill atomic facts limited, Llama 3.3 performed the best out of the models we tested.

Conceptual abstraction and analogy-making are crucial for human learning, reasoning, and adapting to unfamiliar domains. Recently, large language models (LLMs) have made the synthesis of analogical data possible, which, however, still heavily relies on extensive human efforts to be annotated. This paper empirically examines the LLMs’ capability to annotate story-level analogical data. Specifically, we propose a novel multi-stage progressive reasoning prompt framework A3E (Automated Analogy Annotation Expert), which is based on the structure mapping theory from cognitive psychology and efficiently annotates candidate story pairs across six fine-grained categories. We use A3E to evaluate how well the state-of-the-art LLMs can serve as analogy annotators. Experimental results demonstrate that our proposed A3E achieves an average performance gain of + 73% across a range of prompting baselines and base LLMs. The code and data is available at https://github.com/zhangxjohn/A3E.

Recently, a number of repository-level code generation benchmarks–such as CoderEval, DevEval, RepoEval, RepoBench, and LongCode-Arena–have emerged to evaluate the capabilities of large language models (LLMs) beyond standalone benchmarks like HumanEval and MBPP. Thus, a natural question is, would LLMs have similar performance in real world coding tasks as their performance in these benchmarks? Unfortunately, one cannot answer this question, since these benchmarks consist of short completions, synthetic examples, or focus on limited scale repositories, failing to represent real-world coding tasks.To address these challenges, we create RepoCod, a Python code-generation benchmark containing complex tasks with realistic dependencies in real-world large projects and appropriate metrics for evaluating source code. It includes 980 whole-function generation tasks from 11 popular projects, 50.8% of which require repository-level context. RepoCod includes 314 developer-written test cases per instance for better evaluation. We evaluate ten LLMs on RepoCod and find that none achieves more than 30% pass@1 on RepoCod, indicating the necessity of building stronger LLMs that can help developers in real-world software development. In addition, we found that retrieval-augmented generation achieves better results than using target function dependencies as context.

Recent calls for pluralistic alignment emphasize that AI systems should address the diverse needs of all people. Yet, efforts in this space often require sorting people into fixed buckets of pre-specified diversity-defining dimensions (e.g., demographics), risking smoothing out individualistic variations or even stereotyping. To achieve an authentic representation of diversity that respects individuality, we propose individualistic alignment. While individualistic alignment can take various forms, in this paper, we introduce IndieValueCatalog, a dataset transformed from the influential World Values Survey (WVS), to study language models (LMs) on the specific challenge of individualistic value reasoning. Given a sample of an individual’s value-expressing statements, models are tasked with predicting their value judgments in novel cases. With IndieValueCatalog, we reveal critical limitations in frontier LMs’ abilities to predict individualistic values with accuracies only ranging between 55% to 65%. Moreover, our results highlight that a precise description of individualistic values cannot be approximated only via demographic information. Finally, we train a series of IndieValueReasoners to reveal new patterns and dynamics into global human values.

With the popularity of large language models and their high-quality text generation capabilities, researchers are using them as auxiliary tools for text summary writing. Although summaries generated by these large language models are smooth and capture key information sufficiently, the quality of their output depends on the prompt, and the generated text is somewhat procedural to a certain extent. We construct LecSumm to verify whether language models truly capture human writing preferences, in which we recruit 200 college students to write summaries for lecture notes on ten different machine-learning topics and analyze writing preferences in real-world human summaries through the dimensions of length, content depth, tone & style, and summary format. We define the method of capturing human writing preferences by language models as finetuning pre-trained models with data and designing prompts to optimize the output of large language models. The results of translating the analyzed human writing preferences into prompts and conducting experiments show that both models still fail to capture human writing preferences effectively. Our LecSumm dataset brings new challenges to finetuned and prompt-based large language models on the task of human-centered text summarization.

Unvoiced electromyography (EMG) is an effective communication tool for individuals unable to produce vocal speech. However, most prior methods rely on paired voiced and unvoiced EMG signals, along with speech data, for unvoiced EMG-to-text conversion, which is not practical for these individuals. Given the rise of large language models (LLMs) in speech recognition, we explore their potential to understand unvoiced speech. To this end, we address the challenge of learning from unvoiced EMG alone and propose a novel EMG adaptor module that maps EMG features to an LLM’s input space, achieving an average word error rate of 0.49 on a closed-vocabulary unvoiced EMG-to-text task. Even with a conservative data availability of just six minutes, our approach improves performance over specialized models by nearly 20%. While LLMs have been shown to be extendable to new language modalities—such as audio—understanding articulatory biosignals, like unvoiced EMG, is more challenging. This work takes a crucial first step toward enabling LLMs to comprehend unvoiced speech using surface EMG.

This study evaluates Large Language Models’ (LLMs) ability to simulate non-native-like English use observed in human second language (L2) learners interfered with by their native first language (L1). In dialogue-based interviews, we prompt LLMs to mimic L2 English learners with specific L1s (e.g., Japanese, Thai, Urdu) across seven languages, comparing their outputs to real L2 learner data. Our analysis examines L1-driven linguistic biases, such as reference word usage and avoidance behaviors, using information-theoretic and distributional density measures. Results show that modern LLMs (e.g., Qwen2.5, LLAMA3, DeepseekV3, GPT 4o) replicate L1-dependent patterns observed in human L2 data, with distinct influences from various languages (e.g., Japanese, Korean, and Mandarin significantly affect tense agreement, and Urdu influences noun-verb collocations). Our results reveal LLMs’ potential for L2 dialogue generation and evaluation for future educational applications.

Large Language Models (LLMs) are increasingly being used to automate programming tasks. However, the capabilities of LLMs in reasoning about program semantics are still inadequately studied, leaving substantial potential for further exploration. This paper introduces FormalBench, a comprehensive benchmark designed to evaluate the reasoning abilities of Large Language Models (LLMs) on program semantics. Specifically, it utilizes the task of synthesizing formal program specifications as a proxy measure for assessing the semantic reasoning of LLMs. This task requires both comprehensive reasoning over all possible program executions and the generation of precise, syntactically correct expressions that adhere to formal syntax and semantics. Using this benchmark, we evaluated the ability of LLMs to synthesize consistent and complete specifications. Our findings show that LLMs perform well with simple control flows but struggle with more complex structures, especially loops, even with advanced prompting. Additionally, LLMs exhibit limited robustness against semantic-preserving transformations. We also highlight common failure patterns and design self-repair prompts, improving success rates by 25%. FormalBench is packaged as an executable library and has been released at https://github.com/thanhlecongg/FormalBench/.

Peer review is fundamental to scientific research, but the growing volume of publications has intensified the challenges of this expertise-intensive process. While LLMs show promise in various scientific tasks, their potential to assist with peer review, particularly in identifying paper limitations, remains understudied. We first present a comprehensive taxonomy of limitation types in scientific research, with a focus on AI. Guided by this taxonomy, for studying limitations, we present LimitGen, the first comprehensive benchmark for evaluating LLMs’ capability to support early-stage feedback and complement human peer review. Our benchmark consists of two subsets: LimitGen-Syn, a synthetic dataset carefully created through controlled perturbations of high-quality papers, and LimitGen-Human, a collection of real human-written limitations. To improve the ability of LLM systems to identify limitations, we augment them with literature retrieval, which is essential for grounding identifying limitations in prior scientific findings. Our approach enhances the capabilities of LLM systems to generate limitations in research papers, enabling them to provide more concrete and constructive feedback.

The surge of LLM studies makes synthesizing their findings challenging. Analysis of experimental results from literature can uncover important trends across studies, but the time-consuming nature of manual data extraction limits its use.Our study presents a semi-automated approach for literature analysis that accelerates data extraction using LLMs.It automatically identifies relevant arXiv papers, extracts experimental results and related attributes, and organizes them into a structured dataset, LLMEvalDB.We then conduct an automated literature analysis of frontier LLMs, reducing the effort of paper surveying and data extraction by more than 93% compared to manual approaches.We validate LLMEvalDB by showing that it reproduces key findings from a recent manual analysis of Chain-of-Thought (CoT) reasoning and also uncovers new insights that go beyond it, showing, for example, that in-context examples benefit coding & multimodal tasks but offer limited gains in math reasoning tasks compared to zero-shot CoT.Our automatically updatable dataset enables continuous tracking of target models by extracting evaluation studies as new data becomes available. Through LLMEvalDB and empirical analysis, we provide insights into LLMs while facilitating ongoing literature analyses of their behavior.

Communication among humans relies on conversational grounding, allowing interlocutors to reach mutual understanding even when they do not have perfect knowledge and must resolve discrepancies in each other’s beliefs. This paper investigates how large language models (LLMs) manage common ground in cases where they (don’t) possess knowledge, focusing on facts in the political domain where the risk of misinformation and grounding failure is high. We examine LLMs’ ability to answer direct knowledge questions and loaded questions that presuppose misinformation.We evaluate whether loaded questions lead LLMs to engage in active grounding and correct false user beliefs, in connection to their level of knowledge and their political bias.Our findings highlight significant challenges in LLMs’ ability to engage in grounding and reject false user beliefs, raising concerns about their role in mitigating misinformation in political discourse.

We introduce TestCase-Eval, a new benchmark for systematic evaluation of LLMs in test-case generation. TestCase-Eval includes 500 algorithm problems and 100,000 human-crafted solutions from the Codeforces platform. It focuses on two pivotal tasks: (1) Fault Coverage, which measures how well LLM-generated test sets probe diverse input scenarios and cover a wide range of potential failure modes. (2) Fault Exposure, which evaluates whether LLMs can craft a tailored test input that reveals a specific incorrect code implementation. We provide a comprehensive assessment of 19 state-of-the-art open-source and proprietary LLMs on TestCase-Eval, offering insights into their strengths and limitations in generating effective test cases for algorithm problems.

Attributed Question Answering (AQA) has attracted wide attention, but there are still several limitations in evaluating the attributions, including lacking fine-grained attribution categories, relying on manual annotations, and failing to compare attributions with only subtle differences. To bridge these gaps, we introduce Complex Attributed Question Answering (CAQA), a large-scale benchmark containing comprehensive attribution categories, automatically generated using Knowledge Graphs (KGs), and complex attribution scenarios. We have conducted extensive experiments to verify the effectiveness of CAQA, including the benchmarking of 25 automatic evaluators, their comparison with human evaluators, the testing of LLM evaluators fine-tuned by CAQA and so on. These experiments also lead to a series of important findings that can benefit the future research of AQA.

While pre-trained multimodal representations (e.g., CLIP) have shown impressive capabilities, they exhibit significant compositional vulnerabilities leading to counterintuitive judgments. We introduce Multimodal Adversarial Compositionality (MAC), a benchmark that leverages large language models (LLMs) to generate deceptive text samples to exploit these vulnerabilities across different modalities and evaluates them through both sample-wise attack success rate and group-wise entropy-based diversity. To improve zero-shot methods, we propose a self-training approach that leverages rejection-sampling fine-tuning with diversity-promoting filtering, which enhances both attack success rate and sample diversity. Using smaller language models like Llama-3.1-8B, our approach demonstrates superior performance in revealing compositional vulnerabilities across various multimodal representations, including images, videos, and audios.

The radioactive nature of Large Language Model (LLM) watermarking enables the detection of watermarks inherited by student models when trained on the outputs of watermarked teacher models, making it a promising tool for preventing unauthorized knowledge distillation. However, the robustness of watermark radioactivity against adversarial actors remains largely unexplored. In this paper, we investigate whether student models can acquire the capabilities of teacher models through knowledge distillation while avoiding watermark inheritance. We propose two categories of watermark removal approaches: pre-distillation removal through untargeted and targeted training data paraphrasing (UP and TP), and post-distillation removal through inference-time watermark neutralization (WN). Extensive experiments across multiple model pairs, watermarking schemes and hyper-parameter settings demonstrate that both TP and WN thoroughly eliminate inherited watermarks, with WN achieving this while maintaining knowledge transfer efficiency and low computational overhead. Given the ongoing deployment of watermarking techniques in production LLMs, these findings emphasize the urgent need for more robust defense strategies.

Recent works integrating Knowledge Graphs (KGs) have shown promising improvements in enhancing the reasoning capabilities of Large Language Models (LLMs). However, existing benchmarks primarily focus on closed-ended tasks, leaving a gap in evaluating performance on more complex, real-world scenarios. This limitation also hinders a thorough assessment of KGs’ potential to reduce hallucinations in LLMs. To address this, we introduce OKGQA, a new benchmark specifically designed to evaluate LLMs augmented with KGs in open-ended, real-world question answering settings. OKGQA reflects practical complexities through diverse question types and incorporates metrics to quantify both hallucination rates and reasoning improvements in LLM+KG models. To consider the scenarios in which KGs may contain varying levels of errors, we propose a benchmark variant, OKGQA-P, to assess model performance when the semantics and structure of KGs are deliberately perturbed and contaminated. In this paper, we aims to (1) explore whether KGs can make LLMs more trustworthy in an open-ended setting, and (2) conduct a comparative analysis to shed light on method design. We believe this study can facilitate a more complete performance comparison and encourages continuous improvement in integrating KGs with LLMs to mitigate hallucination, and make LLMs more trustworthy.

Interpreting the internal process of neural models has long been a challenge. This challenge remains relevant in the era of large language models (LLMs) and in-context learning (ICL); for example, ICL poses a new issue of interpreting which example in the few-shot examples contributed to identifying/solving the task. To this end, in this paper, we design synthetic diagnostic tasks of inductive reasoning, inspired by the generalization tests in linguistics; here, most in-context examples are ambiguous w.r.t. their underlying rule, and one critical example disambiguates the task demonstrated. The question is whether conventional input attribution (IA) methods can track such a reasoning process, i.e., identify the influential example, in ICL. Our experiments provide several practical findings; for example, a certain simple IA method works the best, and the larger the model, the generally harder it is to interpret the ICL with gradient-based IA methods.

Prompt injection attacks manipulate large language models (LLMs) by misleading them to deviate from the original input instructions and execute maliciously injected instructions, because of their instruction-following capabilities and inability to distinguish between the original input instructions and maliciously injected instructions. To defend against such attacks, recent studies have developed various detection mechanisms. If we restrict ourselves specifically to works which perform detection rather than direct defense, most of them focus on direct prompt injection attacks, while there are few works for the indirect scenario, where injected instructions are indirectly from external tools, such as a search engine. Moreover, current works mainly investigate injection detection methods and pay less attention to the post-processing method that aims to mitigate the injection after detection.In this paper, we investigate the feasibility of detecting and removing indirect prompt injection attacks, and we construct a benchmark dataset for evaluation. For detection, we assess the performance of existing LLMs and open-source detection models, and we further train detection models using our crafted training datasets. For removal, we evaluate two intuitive methods: (1) the *segmentation removal method*, which segments the injected document and removes parts containing injected instructions, and (2) the *extraction removal method*, which trains an extraction model to identify and remove injected instructions.

Large Vision-Language Models often generate hallucinated content that is not grounded in its visual inputs. While prior work focuses on mitigating hallucinations, we instead explore leveraging hallucination correction as a training objective to improve video-language alignment. We introduce HACA, a self-training framework learning to correct hallucinations in descriptions that do not align with the video content. By identifying and correcting inconsistencies, HACA enhances the model’s ability to align video and textual representations for spatio-temporal reasoning. Our experimental results show consistent gains in video-caption binding and text-to-video retrieval tasks, demonstrating that hallucination correction-inspired tasks serve as an effective strategy for improving vision and language alignment.