ACL 2025 论文列表

Multimodal Large Language Models (MLLMs) have expanded the capabilities of traditional language models by enabling interaction through both text and images. However, ensuring the safety of these models remains a significant challenge, particularly in accurately identifying whether multimodal content is safe or unsafe—a capability we term safety awareness. In this paper, we introduce MMSafeAware, the first comprehensive multimodal safety awareness benchmark designed to evaluate MLLMs across 29 safety scenarios with 1,500 carefully curated image-prompt pairs. MMSafeAware includes both unsafe and over-safety subsets to assess models’ abilities to correctly identify unsafe content and avoid over-sensitivity that can hinder helpfulness. Evaluating nine widely used MLLMs using MMSafeAware reveals that current models are not sufficiently safe and often overly sensitive; for example, GPT-4V misclassifies 36.1% of unsafe inputs as safe and 59.9% of benign inputs as unsafe. We further explore three methods to improve safety awareness—prompting-based approaches, visual contrastive decoding, and vision-centric reasoning fine-tuning—but find that none achieve satisfactory performance. Our findings highlight the profound challenges in developing MLLMs with robust safety awareness, underscoring the need for further research in this area. All the code and data will be publicly available to facilitate future research.

Real-world open-domain questions can be complex, especially when answering them requires integrating information from multiple sources. Effectively identifying the necessary information involves *aligning* it with the available data and its organization. However, existing RAG solutions address the alignment problem in a limited manner. Using off-the-shelf LLMs for question decomposition lacks awareness of the available data and its structure, often resulting in suboptimal retrieval performance. Alternatively, iteratively generating follow-up queries and interacting with the data collection, as explored in agentic RAG approaches, shows potential but is often *inefficient* since each successive query depends on previous results rather than being guided by the overall organization of the available data. To address the *alignment* problem, we introduce an LLM-based retrieval method — ARM, designed to better align questions with the organization of the data collection. Instead of solely matching query utterance, ARM explores *relationships among data objects*, enabling a retrieve-all-at-once solution for complex queries. Experimental results demonstrate that ARM significantly outperforms existing RAG methods on various complex open-domain QA tasks across multiple modalities, achieving superior retrieval performance and downstream accuracy while significantly lowering monetary costs.

Large Language Models (LLMs) with API-calling capabilities enabled building effective Language Agents (LA), while also revolutionizing the conventional task-oriented dialogue (TOD) paradigm. However, current approaches face a critical dilemma: TOD systems are often trained on a limited set of target APIs, requiring new data to maintain their quality when interfacing with new services, while LAs are not trained to maintain user intent over multi-turn conversations. Because both robust multi-turn management and advanced function calling are crucial for effective conversational agents, we evaluate these skills on three popular benchmarks: MultiWOZ 2.4 (TOD), BFCL V3 (LA), and API-Bank (LA)—and our analyses reveal that specialized approaches excel in one domain but underperform in the other. To bridge this chasm, we introduce **CoALM** (**C**onversational **A**gentic **L**anguage **M**odel), a unified approach that integrates both conversational and agentic capabilities. We created **CoALM-IT**, a carefully constructed multi-task dataset that interleave multi-turn ReAct reasoning with complex API usage. Using CoALM-IT, we train three models **CoALM 8B**, **CoALM 70B**, and **CoALM 405B**, which outperform top domain-specific models, including GPT-4o, across all three benchmarks. This demonstrates the feasibility of a single model approach for both TOD and LA, setting a new standard for conversational agents.

Understanding clients’ thoughts and beliefs is fundamental in counseling, yet current evaluations of LLM therapists often fail to assess this ability. Existing evaluation methods rely on client simulators that clearly disclose internal states to the therapist, making it difficult to determine whether an LLM therapist can uncover unexpressed perspectives. To address this limitation, we introduce MindVoyager, a novel evaluation framework featuring a controllable and realistic client simulator which dynamically adapts itself based on the ongoing counseling session, offering a more realistic and challenging evaluation environment. We further introduce evaluation metrics that assess the exploration ability of LLM therapists by measuring their thorough understanding of client’s beliefs and thoughts.

Code security and usability are both essential for various coding assistant applications driven by large language models (LLMs). Current code security benchmarks focus solely on single evaluation task and paradigm, such as code completion and generation, lacking comprehensive assessment across dimensions like secure code generation, vulnerability repair and discrimination. In this paper, we first propose CoV-Eval, a multi-task benchmark covering various tasks such as code completion, vulnerability repair, vulnerability detection and classification, for comprehensive evaluation of LLM code security. Besides, we developed VC-Judge, an improved judgment model that aligns closely with human experts and can review LLM-generated programs for vulnerabilities in a more efficient and reliable way. We conduct a comprehensive evaluation of 20 proprietary and open-source LLMs. Overall, while most LLMs identify vulnerable codes well, they still tend to generate insecure codes and struggle with recognizing specific vulnerability types and performing repairs. Extensive experiments and qualitative analyses reveal key challenges and optimization directions, offering insights for future research in LLM code security.

Hallucination has emerged as a critical challenge for large language models (LLMs) and large vision-language models (LVLMs), particularly in high-stakes medical applications. Despite its significance, dedicated research on medical hallucination remains unexplored. In this survey, we first provide a unified perspective on medical hallucination for both LLMs and LVLMs, and delve into its causes. Subsequently, we review recent advancements in detecting, evaluating, and mitigating medical hallucinations, offering a comprehensive overview of evaluation benchmarks, metrics, and strategies developed to tackle this issue. Moreover, we delineate the current challenges and delve into new frontiers, thereby shedding light on future research. We hope this work coupled with open-source resources can provide the community with quick access and spur breakthrough research in medical hallucination.

Large language models excel at problem-solving but often struggle with complex reasoning and factual accuracy. While chain-of-thought and retrieval-augmented generation help break down problems and retrieve knowledge, they still falter on challenging tasks like competitive programming due to frequent reasoning errors and irrelevant retrieval. To address this, we introduce Critic-guided planning with Retrieval-augmentation, CR-Planner, a novel framework that leverages fine-tuned critic models to guide both reasoning and retrieval processes through planning. CR-Planner iteratively selects and executes sub-goals, guided by critic models. A sub-goal critic identifies promising sub-goals from reasoning, query generation, and retrieval, while an execution critic evaluates outputs of sub-goal executions. We employ Monte Carlo Tree Search to collect data for critic training, allowing systematic exploration of action sequences and effective navigation toward the final answer. We evaluate CR-Planner on challenging domain-knowledge-intensive and reasoning-heavy tasks, including competitive programming, theorem-driven math reasoning, and complex domain retrieval problems. It significantly outperforms baselines, demonstrating effectiveness in both reasoning and retrieval.

Recent advancements in Vision-Language Models (VLMs) have opened new possibilities in automatic grading of handwritten student responses, particularly in mathematics. However, a comprehensive study to test the ability of VLMs to evaluate and reason over handwritten content remains absent. To address this gap, we introduce FERMAT, a benchmark designed to assess VLMs’ ability to detect, localize and correct errors in handwritten mathematical content. FERMAT spans four key error dimensions - computational, conceptual, notational, and presentation - and comprises over 2,200 handwritten math solutions derived from 609 manually curated problems from grades 7-12 with intentionally introduced perturbations. Using FERMAT we benchmark nine VLMs across three tasks: error detection, localization, and correction. Our results reveal significant shortcomings in current VLMs in reasoning over handwritten text, with Gemini-1.5-Pro achieving the highest error correction rate (77%). We also observed that some models struggle with processing handwritten content, as their accuracy improves when handwritten inputs are replaced with printed text or images. These findings highlight the limitations of current VLMs and reveal new avenues for improvement. We will release FERMAT and all the associated resources in the open-source to drive further research.

Non-verbal communication (NVC) is an integral part of human language, but it has been overlooked in natural language processing research. Studying NVC in general is challenging because of its high variance in interpretation among individuals and cultures, but mime—the theatrical technique of suggesting intent using only gesture, expression, and movement—is a subset of NVC with much lower human interpretation variance. As a gateway for evaluating vision-language models on their understanding of NVC, we propose Mime Identification-based Multimodal Evaluation (MIME), a gesture recognition task built upon a novel corpus of mimed activity comprising 86 unique gestures with a variety of perturbations applied to the avatar, background, and viewpoint for evaluating recognition robustness. We find that both open-weight and API-based vision-language models perform significantly worse than humans at identifying mimed gestures in MIME, motivating the need for increased research for instilling more robust understanding of human actions for VLMs.

Vision-language models (VLMs) integrate textual and visual information, enabling the model to process visual inputs and leverage visual information to generate predictions. Such models are demanding for tasks such as visual question answering, image captioning, and visual grounding. However, some recent work found that VLMs often rely heavily on textual information, ignoring visual information, but are still able to achieve competitive performance in vision-language (VL) tasks. This survey reviews modality collapse analysis work to provide insights into the reason for this unintended behavior. It also reviews probing studies for fine-grained vision-language understanding, presenting current findings on information encoded in VL representations and highlighting potential directions for future research.

While ChatGPT and GPT-based models are able to effectively perform many tasks without additional fine-tuning, they struggle with tasks related to extremely low-resource languages and indigenous languages. Uniform Meaning Representation (UMR), a semantic representation designed to capture the meaning of texts in many languages, is well-positioned to be leveraged in the development of low-resource language technologies. In this work, we explore the downstream utility of UMR for low-resource languages by incorporating it into GPT-4 prompts. Specifically, we examine the ability of GPT-4 to perform translation from three indigenous languages (Navajo, Arápaho, and Kukama), with and without demonstrations, as well as with and without UMR annotations. Ultimately, we find that in the majority of our test cases, integrating UMR into the prompt results in a statistically significant increase in performance, which is a promising indication of future applications of the UMR formalism.

Natural Language Processing tasks that aim to infer an author’s private states, e.g., emotions and opinions, from their written text, typically rely on datasets annotated by third-party annotators. However, the assumption that third-party annotators can accurately capture authors’ private states remains largely unexamined. In this study, we present human subjects experiments on emotion recognition tasks that directly compare third-party annotations with first-party (author-provided) emotion labels. Our findings reveal significant limitations in third-party annotations—whether provided by human annotators or large language models (LLMs)—in faithfully representing authors’ private states. However, LLMs outperform human annotators nearly across the board. We further explore methods to improve third-party annotation quality. We find that demographic similarity between first-party authors and third-party human annotators enhances annotation performance. While incorporating first-party demographic information into prompts leads to a marginal but statistically significant improvement in LLMs’ performance. We introduce a framework for evaluating the limitations of third-party annotations and call for refined annotation practices to accurately represent and model authors’ private states.

Spatial relation reasoning is a crucial task for multimodal large language models (MLLMs) to understand the objective world. However, current benchmarks have issues like relying on bounding boxes, ignoring perspective substitutions, or allowing questions to be answered using only the model’s prior knowledge without image understanding. To address these issues, we introduce SpatialMQA, a human-annotated spatial relation reasoning benchmark based on COCO2017, which enables MLLMs to focus more on understanding images in the objective world. To ensure data quality, we design a well-tailored annotation procedure, resulting in SpatialMQA consisting of 5,392 samples. Based on this benchmark, a series of closed- and open-source MLLMs are implemented and the results indicate that the current state-of-the-art MLLM achieves only 48.14% accuracy, far below the human-level accuracy of 98.40%. Extensive experimental analyses are also conducted, suggesting the future research directions. The benchmark and codes are available at https://huggingface.co/datasets/liuziyan/SpatialMQA.

This paper introduces MISS-QA, the first benchmark specifically designed to evaluate the ability of models to interpret schematic diagrams within scientific literature. MISS-QA comprises 3,000 expert-annotated examples over 983 scientific papers. In this benchmark, models are tasked with interpreting schematic diagrams that illustrate research overviews and answering corresponding information-seeking questions based on the broader context of the paper. To ensure reliable and consistent evaluation, we propose an automated evaluating protocol powered by open-source LLMs trained on human-scored data. We assess the performance of 18 frontier multimodal foundation models, including o1, Claude-3.5, Llama-3.2-Vision, and Qwen2-VL. We reveal a significant performance gap between these models and human experts on MISS-QA. Our analysis of model performance on unanswerable questions and our detailed error analysis further highlight the strengths and limitations of current models, offering key insights to enhance models in comprehending multimodal scientific literature.

Medical Vision-Language Pre-training (MedVLP) has made significant progress in enabling zero-shot tasks for medical image understanding. However, training MedVLP models typically requires large-scale datasets with paired, high-quality image-text data, which are scarce in the medical domain. Recent advancements in Large Language Models (LLMs) and diffusion models have made it possible to generate large-scale synthetic image-text pairs. This raises the question: Can MedVLP succeed using purely synthetic data? To address this, we use off-the-shelf generative models to create synthetic radiology reports and paired Chest X-ray (CXR) images, and propose an automated pipeline to build a diverse, high-quality synthetic dataset, enabling a rigorous study that isolates model and training settings, focusing entirely from the data perspective.Our results show that MedVLP models trained exclusively on synthetic data outperform those trained on real data by 3.8% in averaged AUC on zero-shot classification. Moreover, using a combination of synthetic and real data leads to a further improvement of 9.07%. Additionally, MedVLP models trained on synthetic or mixed data consistently outperform those trained on real data in zero-shot grounding, as well as in fine-tuned classification and segmentation tasks.Our analysis suggests MedVLP trained on well-designed synthetic data can outperform models trained on real datasets, which may be limited by low-quality samples and long-tailed distributions[^1].[^1]: All data and code will be released upon acceptance.

As the capabilities of Multimodal Large Language Models (MLLMs) improve, the need for higher-order evaluation of them is increasing. However, there is a lack of work evaluating MLLM for higher-order perception and understanding of Chinese visual content. To address this, we introduce the CII-Bench, which aims to assess MLLMs’ such capabilities for Chinese images. To ensure the authenticity of the Chinese context, images in CII-Bench are sourced from the Chinese Internet and manually reviewed, with corresponding answers also manually crafted. Additionally, CII-Bench incorporates images that represent Chinese traditional culture, such as famous Chinese traditional paintings, which can deeply reflect the model’s understanding of Chinese traditional culture. Through experiments on multiple MLLMs using CII-Bench, significant findings emerged. There is a large gap between MLLMs and humans in performance. The highest MLLM accuracy is 64.4%, while the human average is 78.2% and the peak is 81.0%. MLLMs perform poorly on traditional culture images, indicating limitations in understanding high-level semantics and lacking a deep knowledge base of Chinese traditional culture. Moreover, most models have higher accuracy when image emotion hints are added to the prompts. We believe CII-Bench will help MLLMs better understand Chinese semantics and specific images, and move forward the development of expert artificial general intelligence (AGI). Our project is publicly available at https://cii-bench.github.io.

The massive user-generated content (UGC) available in Chinese social media is giving rise to the possibility of studying internet buzzwords. In this paper, we study if large language models (LLMs) can generate accurate definitions for these buzzwords based on UGC as examples. Our work serves a threefold contribution. First, we introduce CHEER, the first dataset of Chinese internet buzzwords, each annotated with a definition and relevant UGC. Second, we propose a novel method, called RESS, to effectively steer the comprehending process of LLMs to produce more accurate buzzword definitions, mirroring the skills of human language learning. Third, with CHEER, we benchmark the strengths and weaknesses of various off-the-shelf definition generation methods and our RESS. Our benchmark demonstrates the effectiveness of RESS while revealing a crucial shared challenge: comprehending unseen buzzwords and leveraging sufficient, high-quality UGC to facilitate this comprehension. In this paper, we believe our work lays the groundwork for future advancements in LLM-based definition generation. Our dataset and code will be openly released.

Argument schemes represent stereotypical patterns of reasoning that occur in everyday arguments. However, despite their usefulness, argument scheme classification, that is classifying natural language arguments according to the schemes they are instances of, is an under-explored task in NLP. In this paper we present a systematic evaluation of large language models (LLMs) for classifying argument schemes based on Walton’s taxonomy. We experiment with seven LLMs in zero-shot, few-shot, and chain-of-thought prompting, and explore two strategies to enhance task instructions: employing formal definitions and LLM-generated descriptions. Our analysis on both manually annotated and automatically generated arguments, including enthymemes, indicates that while larger models exhibit satisfactory performance in identifying argument schemes, challenges remain for smaller models. Our work offers the first comprehensive assessment of LLMs in identifying argument schemes, and provides insights for advancing reasoning capabilities in computational argumentation.

Recently, o1-like models have drawn significant attention, where these models produce the long Chain-of-Thought (CoT) reasoning steps to improve the reasoning abilities of existing Large Language Models (LLMs). In this paper, to understand the qualities of these long CoTs and measure the critique abilities of existing LLMs on these long CoTs, we introduce the DeltaBench including the generated long CoTs from different o1-like models (e.g., QwQ, DeepSeek-R1) for different reasoning tasks (e.g., Math, Code, General Reasoning), to measure the ability to detect errors in long COT reasoning. Based on DeltaBench, we first perform fine-grained analysis of the generated long CoTs to discover the effectiveness and efficiency of different o1-like models. Then, we conduct extensive evaluations of existing process reward models (PRMs) and critic models to detect the errors of each annotated process, which aims to investigate the boundaries and limitations of existing PRMs and critic models. Finally, we hope that DeltaBench could guide developers to better understand the long CoT reasoning abilities of their models.