ACL 2025 论文列表

The safe deployment of large language models (LLMs) necessitates comprehensive safety evaluations through red teaming. However, existing methods face challenges in managing semantic intricacies and optimizing the efficiency of the search process. To overcome these limitations, we propose Better Red Teaming (BRT)—an innovative framework that reconceptualizes test case generation as a strategic planning problem, leveraging Monte Carlo Tree Search (MCTS). A notable advancement of our approach is the incorporation of LLMs as world models, enabling the prediction of state transitions and simulation of long-term outcomes throughout the search process. By jointly optimizing objectives related to conditional mutual information and diversity, we improve the world model’s capacity to follow actions while maintaining output diversity. Extensive experiments conducted across a range of LLM architectures demonstrate that BRT achieves state-of-the-art attack success rates without sacrificing computational efficiency.

Process supervision, i.e., evaluating each step, is critical for complex large language model (LLM) reasoning and test-time searching with increased inference compute. Existing approaches, represented by process reward models (PRMs), primarily focus on rewarding signals up to the current step, exhibiting a one-directional nature and lacking a mechanism to model the distance to the final target. To address this problem, we draw inspiration from the A* algorithm, which states that an effective supervisory signal should simultaneously consider the incurred cost and the estimated cost for reaching the target. Building on this key insight, we introduce BiRM, a novel process supervision model that not only evaluates the correctness of previous steps but also models the probability of future success. We conduct extensive experiments on mathematical reasoning tasks and demonstrate that BiRM provides more precise evaluations of LLM reasoning steps, achieving an improvement of 3.1% on Gaokao2023 over PRM under the Best-of-N sampling method. Besides, in search-based strategies, BiRM provides more comprehensive guidance and outperforms ORM by 5.0% and PRM by 3.8% respectively on MATH-500.

Despite their remarkable capabilities, LLMs learn word representations that exhibit the undesirable yet poorly understood feature of anisotropy. In this paper, we argue that the second moment in Adam is a cause of anisotropic embeddings, and suggest a modified optimizer called Coupled Adam to mitigate the problem. Our experiments demonstrate that Coupled Adam significantly improves the quality of embeddings, while also leading to better upstream and downstream performance on large enough datasets.

Significant efforts have been made in the NLP community to facilitate the automatic analysis of climate-related corpora by tasks such as climate-related topic detection, climate risk classification, question answering over climate topics, and many more. In this work, we perform a reproducibility study on 8 tasks and 29 datasets, testing 6 models. We find that many tasks rely heavily on surface-level keyword patterns rather than deeper semantic or contextual understanding. Moreover, we find that 96% of the datasets contain annotation issues, with 16.6% of the sampled wrong predictions of a zero-shot classifier being actually clear annotation mistakes, and 38.8% being ambiguous examples.These results call into question the reliability of current benchmarks to meaningfully compare models and highlight the need for improved annotation practices. We conclude by outlining actionable recommendations to enhance dataset quality and evaluation robustness.

Large Vision-Language Models (LVLMs) have demonstrated remarkable performance across diverse tasks. Despite great success, recent studies show that LVLMs encounter substantial limitations when engaging with visual graphs. To study the reason behind these limitations, we propose VGCure, a comprehensive benchmark covering 22 tasks for examining the fundamental graph understanding and reasoning capacities of LVLMs. Extensive evaluations conducted on 14 LVLMs reveal that LVLMs are weak in basic graph understanding and reasoning tasks, particularly those concerning relational or structurally complex information. Based on this observation, we propose a structure-aware fine-tuning framework to enhance LVLMs with structure learning abilities through three self-supervised learning tasks. Experiments validate the effectiveness of our method in improving LVLMs’ performance on fundamental and downstream graph learning tasks, as well as enhancing their robustness against complex visual graphs.

The growing excitement around the ability of large language models (LLMs) to tackle various tasks has been tempered by their propensity for generating unsubstantiated information (hallucination) and by their inability to effectively handle inconsistent inputs. To detect such issues, we propose the novel task of Query-Conditioned Natural Language Inference (QC-NLI), where the goal is to determine the semantic relationship (e.g. entailment or not entailment) between two documents conditioned on a query; we demonstrate that many common tasks regarding inconsistency detection can be formulated as QC-NLI problems. We focus on three applications in particular: fact verification, intrinsic hallucination detection, and document inconsistency detection. We convert existing datasets for these tasks into the QC-NLI format, and manual annotation confirms their high quality. Finally, we employ zero- and few-shot prompting methods to solve the QC-NLI prediction problem for each task, showing the critical importance of conditioning on the query.

Large Audio-Language Models (LALMs), such as GPT-4o, have recently unlocked audio dialogue capabilities, enabling direct spoken exchanges with humans. The potential of LALMs broadens their applicability across a wide range of practical scenarios supported by audio dialogues. However, given these advancements, a comprehensive benchmark to evaluate the performance of LALMs in the open-ended audio dialogue understanding remains absent currently. To address this gap, we propose an **A**udio **D**ialogue **U**nderstanding **Bench**mark **(ADU-Bench),** which consists of 4 benchmark datasets. They assess the open-ended audio dialogue ability for LALMs in 3 general scenarios, 12 skills, 9 multilingual languages, and 4 categories of ambiguity handling. Notably, *we firstly propose the evaluation of ambiguity handling* in audio dialogues that expresses different intentions beyond the same literal meaning of sentences, *e.g.,* ‘“Really!?”‘ with different intonations. In summary, ADU-Bench includes over 20,000 open-ended audio dialogues for the assessment of LALMs. Through extensive experiments conducted on 16 LALMs, our analysis reveals that existing LALMs struggle with mathematical symbols and formulas, understanding human behavior such as roleplay, comprehending multiple languages, and handling audio dialogue ambiguities from different phonetic elements, such as intonations, pause positions, and homophones. The benchmark is available at https://adu-bench.github.io/.

Do Large Language Models (LLMs) hold positions that conflict with your country’s values? Occasionally they do! However, existing works primarily focus on ethical reviews, failing to capture the diversity of national values, which encompass broader policy, legal, and moral considerations. Furthermore, current benchmarks that rely on spectrum tests using manually designed questionnaires are not easily scalable. To address these limitations, we introduce NaVAB, a comprehensive benchmark to evaluate the alignment of LLMs with the values of five major nations: China, the United States, the United Kingdom, France, and Germany. NaVAB implements a national value extraction pipeline to efficiently construct value assessment datasets. Specifically, we propose a modeling procedure with instruction tagging to process raw data sources, a screening process to filter value-related topics and a generation process with a Conflict Reduction mechanism to filter non-conflicting values. We conduct extensive experiments on various LLMs across countries, and the results provide insights into assisting in the identification of misaligned scenarios. Moreover, we demonstrate that NaVAB can be combined with alignment techniques to effectively reduce value concerns by aligning LLMs’ values with the target country.

Current advanced long-context language models offer great potential for real-world software engineering applications. However, progress in this critical domain remains hampered by a fundamental limitation: the absence of a rigorous evaluation framework for long code understanding. To gap this obstacle, we propose a long code understanding benchmark LongCodeU from four aspects (8 tasks) to evaluate LCLMs’ long code understanding ability required for practical applications, including code unit perception, intra-code unit understanding, inter-code unit relation understanding, and long code documentation understanding. We evaluate 9 popular LCLMs on LongCodeU (i.e., 6 general models and 3 code models). Our experimental results reveal key limitations in current LCLMs’ capabilities for long code understanding. Particularly, the performance of LCLMs drops dramatically when the long code length is greater than 32K, falling far short of their claimed 128K to 1M context windows. In the four aspects, inter-code unit relation understanding is the most challenging for LCLMs. Our study provides valuable insights for optimizing LCLMs and driving advancements in software engineering.

Large Language Models (LLMs) have shown promise in software vulnerability detection, particularly on function-level benchmarks like Devign and BigVul. However, real-world detection requires interprocedural analysis, as vulnerabilities often emerge through multi-hop function calls rather than isolated functions. While repository-level benchmarks like ReposVul and VulEval introduce interprocedural context, they remain computationally expensive, lack pairwise evaluation of vulnerability fixes, and explore limited context retrieval, limiting their practicality.We introduce JITVul, a JIT vulnerability detection benchmark linking each function to its vulnerability-introducing and fixing commits. Built from 879 CVEs spanning 91 vulnerability types, JITVul enables comprehensive evaluation of detection capabilities. Our results show that ReAct Agents, leveraging thought-action-observation and interprocedural context, perform better than LLMs in distinguishing vulnerable from benign code. While prompting strategies like Chain-of-Thought help LLMs, ReAct Agents require further refinement. Both methods show inconsistencies, either misidentifying vulnerabilities or over-analyzing security guards, indicating significant room for improvement.

Large Language Models (LLMs) demonstrate exceptional proficiency in general-purpose tasks but struggle with numerical reasoning, particularly in low-resource languages like Bengali. Despite advancements, limited research has explored their numerical reasoning capabilities in these languages. To address this gap, we present BenNumEval (Bengali Numerical Evaluation), a benchmark designed to assess LLMs on numerical reasoning tasks in Bengali. It comprises six diverse tasks and a total of 3.2k samples curated from real-world problem-solving scenarios. Our extensive evaluations reveal that even with advanced prompting techniques such as Cross-Lingual Prompting (XLP) and Cross-Lingual Chain-of-Thought Prompting (XCoT), LLMs fall notably short of human-level performance, particularly when using Bengali Native Prompting (BNaP). These findings underscore the substantial gap between current LLM capabilities and human expertise in numerical reasoning, highlighting the need for more robust and linguistically inclusive AI models to advance Bengali Language Processing and equitable AI development. The source code for the system and evaluation pipeline is publicly available on GitHub.

In the epoch of multilingual large language models (LLMs), it is still challenging to evaluate the models’ understanding of lower-resourced languages, which motivates further development of expert-crafted natural language understanding benchmarks. We introduce BelarusianGLUE — a natural language understanding benchmark for Belarusian, an East Slavic language, with ≈15K instances in five tasks: sentiment analysis, linguistic acceptability, word in context, Winograd schema challenge, textual entailment. A systematic evaluation of BERT models and LLMs against this novel benchmark reveals that both types of models approach human-level performance on easier tasks, such as sentiment analysis, but there is a significant gap in performance between machine and human on a harder task — Winograd schema challenge. We find the optimal choice of model type to be task-specific: e.g. BERT models underperform on textual entailment task but are competitive for linguistic acceptability. We release the datasets (https://hf.co/datasets/maaxap/BelarusianGLUE) and evaluation code (https://github.com/maaxap/BelarusianGLUE).

The surge in online content has created an urgent demand for robust detection systems, especially in non-English contexts where current tools demonstrate significant limitations. We introduce forePLay, a novel Polish-language dataset for erotic content detection, comprising over 24,000 annotated sentences. The dataset features a multidimensional taxonomy that captures ambiguity, violence, and socially unacceptable behaviors. Our comprehensive evaluation demonstrates that specialized Polish language models achieve superior performance compared to multilingual alternatives, with transformer-based architectures showing particular strength in handling imbalanced categories. The dataset and accompanying analysis establish essential frameworks for developing linguistically-aware content moderation systems, while highlighting critical considerations for extending such capabilities to morphologically complex languages.

A core aspect of compositionality, systematicity is a desirable property in ML models as it enables strong generalization to novel contexts. This has led to numerous studies proposing benchmarks to assess systematic generalization, as well as models and training regimes designed to enhance it. Many of these efforts are framed as addressing the challenge posed by Fodor and Pylyshyn. However, while they argue for systematicity of representations, existing benchmarks and models primarily focus on the systematicity of behaviour. We emphasize the crucial nature of this distinction. Furthermore, building on Hadley’s (1994) taxonomy of systematic generalization, we analyze the extent to which behavioural systematicity is tested by key benchmarks in the literature across language and vision. Finally, we highlight ways of assessing systematicity of representations in ML models as practiced in the field of mechanistic interpretability.

Large Language Models (LLMs) excel at text summarization, a task that requires models to select content based on its importance. However, the exact notion of salience that LLMs have internalized remains unclear. To bridge this gap, we introduce an explainable framework to systematically derive and investigate information salience in LLMs through their summarization behavior. Using length-controlled summarization as a behavioral probe into the content selection process, and tracing the answerability of Questions Under Discussion throughout, we derive a proxy for how models prioritize information. Our experiments on 13 models across four datasets reveal that LLMs have a nuanced, hierarchical notion of salience, generally consistent across model families and sizes. While models show highly consistent behavior and hence salience patterns, this notion of salience cannot be accessed through introspection, and only weakly correlates with human perceptions of information salience.

Language model evaluation is a daunting task: prompts are brittle, corpus-level perplexities are vague, and the choice of benchmarks are endless. Finding examples that show meaningful, generalizable differences between two LMs is crucial to understanding where one model succeeds and another fails. Can this process be done automatically? In this work, we propose methodology for automated comparison of language models that uses performance-aware contextual embeddings to find fine-grained features of text where one LM outperforms another. Our method, which we name BehaviorBox, extracts coherent features that demonstrate differences with respect to the ease of generation between two LMs. Specifically, BehaviorBox finds features that describe groups of words in fine-grained contexts, such as “conditional ‘were’ in the phrase ‘if you were’” and “exclamation marks after emotional statements”, where one model outperforms another within a particular datatset. We apply BehaviorBox to compare models that vary in size, model family, and post-training, and enumerate insights into specific contexts that illustrate meaningful differences in performance which cannot be found by measures such as corpus-level perplexity alone.

Due to the demand for efficient fine-tuning of large language models, Low-Rank Adaptation (LoRA) has been widely adopted as one of the most effective parameter-efficient fine-tuning methods. Nevertheless, while LoRA improves efficiency, there remains room for improvement in accuracy. Herein, we adopt a novel perspective to assess the characteristics of LoRA ranks. The results reveal that different ranks within the LoRA modules not only exhibit varying levels of importance but also evolve dynamically throughout the fine-tuning process, which may limit the performance of LoRA. Based on these findings, we propose BeamLoRA, which conceptualizes each LoRA module as a beam where each rank naturally corresponds to a potential sub-solution, and the fine-tuning process becomes a search for the optimal sub-solution combination. BeamLoRA dynamically eliminates underperforming sub-solutions while expanding the parameter space for promising ones, enhancing performance with a fixed rank. Extensive experiments across three base models and 12 datasets spanning math reasoning, code generation, and commonsense reasoning demonstrate that BeamLoRA consistently enhances the performance of LoRA, surpassing the other baseline methods.

Model merging is a widespread technology in large language models (LLMs) that integrates multiple task-specific LLMs into a unified one, enabling the merged model to inherit the specialized capabilities of these LLMs. Most task-specific LLMs are sourced from open-source communities and have not undergone rigorous auditing, potentially imposing risks in model merging. This paper highlights an overlooked privacy risk: *an unsafe model could compromise the privacy of other LLMs involved in the model merging*. Specifically, we propose *PhiMM*, a privacy attack approach that trains a phishing model capable of stealing privacy using a crafted privacy phishing instruction dataset. Furthermore, we introduce a novel model cloaking method that mimics a specialized capability to conceal attack intent, luring users into merging the phishing model. Once victims merge the phishing model, the attacker can extract personally identifiable information (PII) or infer membership information (MI) by querying the merged model with the phishing instruction. Experimental results show that merging a phishing model increases the risk of privacy breaches. Compared to the results before merging, PII leakage increased by 3.9% and MI leakage increased by 17.4% on average. We release the code of *PhiMM* through an anonymous link.

In the rapidly evolving field of image generation, achieving precise control over generated content and maintaining semantic consistency remain significant limitations, particularly concerning grounding techniques and the necessity for model fine-tuning. To address these challenges, we propose BayesGenie, an off-the-shelf approach that integrates Large Language Models (LLMs) with Bayesian Optimization to facilitate precise and user-friendly image editing. Our method enables users to modify images through natural language descriptions without manual area marking, while preserving the original image’s semantic integrity. Unlike existing techniques that require extensive pre-training or fine-tuning, our approach demonstrates remarkable adaptability across various LLMs through its model-agnostic design. BayesGenie employs an adapted Bayesian optimization strategy to automatically refine the inference process parameters, achieving high-precision image editing with minimal user intervention. Through extensive experiments across diverse scenarios, we demonstrate that our framework outperforms existing methods in both editing accuracy and semantic preservation, as validated using different LLMs including Claude3 and GPT-4.