ACL 2025 Papers

Recently, large language model (LLM)-based agents have achieved significant success in interactive environments, attracting significant academic and industrial attention. Despite these advancements, current research predominantly focuses on English scenarios. In reality, there are over 7,000 languages worldwide, all of which demand access to comparable agentic services. Nevertheless, the development of language agents remains inadequate for meeting the diverse requirements of multilingual agentic applications. To fill this gap, we introduce X-WebAgentBench, a novel multilingual agent benchmark in an interactive web environment, which evaluates the planning and interaction performance of language agents across multiple languages, thereby contributing to the advancement of global agent intelligence. Additionally, we assess the performance of various LLMs and cross-lingual alignment methods, examining their effectiveness in enhancing agents. Our findings reveal that even advanced models like GPT-4o, when combined with cross-lingual techniques, fail to achieve satisfactory results. We hope that X-WebAgentBench can serve as a valuable benchmark for multilingual agent scenario in real-world applications.

The Turing test examines whether AIs exhibit human-like behaviour in natural language conversations. The traditional setting limits each participant to one message at a time and requires constant human participation. This fails to reflect a natural conversational style and hinders the evaluation of dialogue agents based on Large Language Models (LLMs) in complex and prolonged interactions. This paper proposes X-Turing, which enhances the original test with a burst dialogue pattern, allowing more dynamic exchanges using consecutive messages. It further reduces human workload by iteratively generating dialogues that simulate the long-term interaction between the agent and a human to compose the majority of the test process. With the pseudo-dialogue history, the agent then engages in a shorter dialogue with a real human, which is paired with a human-human conversation on the same topic to be judged using questionnaires. We introduce the X-Turn Pass-Rate metric to assess the human likeness of LLMs across varying durations. While LLMs like GPT-4 initially perform well, achieving pass rates of 51.9% and 38.9% during 3 turns and 10 turns of dialogues respectively, their performance drops as the dialogue progresses, which underscores the difficulty in maintaining consistency in the long term.

We introduce the Exemplar-Based Expository Text Generation task, aiming to generate an expository text on a new topic using an exemplar on a similar topic. Current methods fall short due to their reliance on extensive exemplar data, difficulty in adapting topic-specific content, and issues with long-text coherence. To address these challenges, we propose the concept of Adaptive Imitation and present a novel Recurrent Plan-then-Adapt (RePA) framework. RePA leverages large language models (LLMs) for effective adaptive imitation through a fine-grained plan-then-adapt process. RePA also enables recurrent segment-by-segment imitation, supported by two memory structures that enhance input clarity and output coherence. We also develop task-specific evaluation metrics–imitativeness, adaptiveness, and adaptive-imitativeness–using LLMs as evaluators. Experimental results across our collected three diverse datasets demonstrate that RePA surpasses existing baselines in producing factual, consistent, and relevant texts for this task.

Large Multimodal Models (LMMs) have demonstrated impressive performance on existing medical Visual Question Answering (Med-VQA) benchmarks. However, high reported accuracy does not necessarily reflect their true diagnostic reliability in clinical settings. This study reveals that state-of-the-art models perform worse than random guessing on medical diagnosis questions when subjected to simple Probing Evaluation for Medical Diagnosis (ProbMed). ProbMed challenges models through probing evaluation and procedural diagnosis. Particularly, probing evaluation features pairing ground-truth questions with adversarial counterparts that feature negated and hallucinated attributes, while procedural diagnosis requires reasoning across various dimensions for each image, including modality recognition, organ identification, clinical findings, abnormalities, and positional grounding. Our evaluation reveals that even top-performing models like GPT-4o, GPT-4V, and Gemini Pro perform worse than random guessing on specialized diagnostic questions, indicating significant limitations in handling fine-grained medical inquiries. Furthermore, our ablation study on open-source models (e.g., LLaVA, LLaVA-Med, and Med-Flamingo) identifies poor visual understanding as a primary bottleneck—a limitation that can be partially mitigated by incorporating visual descriptions generated by GPT-4o, resulting in an average performance improvement of 9.44%. These findings underscore the urgent need for more robust evaluation methods and domain-specific expertise to ensure the reliability of LMMs in high-stakes medical applications.

Recent advances in large vision-language models (LVLMs) have shown promise for embodied task planning, yet they struggle with fundamental challenges like dependency constraints and efficiency. Existing approaches either solely optimize action selection or directly leverage pre-trained models as world models during inference, overlooking the benefits of learning to model the world as a way to enhance planning capabilities. We propose Dual Preference Optimization (D2PO), a new learning framework that jointly optimizes state prediction and action selection through preference learning, enabling LVLMs to understand environment dynamics for better planning. To automatically collect trajectories and stepwise preference data without human annotation, we introduce a tree search mechanism for extensive exploration via trial-and-error. Extensive experiments on VoTa-Bench demonstrate that our D2PO-based method significantly outperforms existing methods and GPT-4o when applied to Qwen2-VL (7B), LLaVA-1.6 (7B), and LLaMA-3.2 (11B), achieving superior task success rates with more efficient execution paths.

Annotating event descriptions with their aspectual features is often seen as a pre-requisite to temporal reasoning. However, a recent study by Pruś et al. (2024) has shown that non-experts’ annotations of the aspectual class of English verb phrases can disagree with both expert linguistic annotations and each another. They hypothesised that people use their world knowledge to tacitly conjure their own contexts, leading to disagreement between them. In this paper, we test that hypothesis by adding context to Pruś et al.’s examples and mirroring their experiment. Our results show that whilst their hypothesis explains some of the disagreement, some examples continue to yield divided responses even with the additional context. Finally, we show that outputs from GPT-4, despite to some degree capturing the aspectual class division, are not an accurate predictor of human answers.

Social psychologists have shown that Warmth (W) and Competence (C) are the primary dimensions along which we assess other people and groups. These dimensions impact various aspects of our lives from social competence and emotion regulation to success in the work place and how we view the world. More recent work has started to explore how these dimensions develop, why they have developed, and what they constitute. Of particular note, is the finding that warmth has two distinct components: Trust (T) and Sociability (S). In this work, we introduce Words of Warmth, the first large-scale repository of manually derived word–warmth (as well as word–trust and word–sociability) associations for over 26k English words. We show that the associations are highly reliable. We use the lexicons to study the rate at which children acquire WCTS words with age. Finally, we show that the lexicon enables a wide variety of bias and stereotype research through case studies on various target entities. Words of Warmth is freely available at: http://saifmohammad.com/warmth.html

Retrieval-augmented generation (RAG) enhances the quality of LLM generation by providing relevant chunks, but retrieving accurately from external knowledge remains challenging due to missing contextually important words in query. We present Word2Passage, a novel approach that improves retrieval accuracy by optimizing word importance in query expansion. Our method generates references at word, sentence, and passage levels for query expansion, then determines word importance by considering both their reference level origin and characteristics derived from query types and corpus analysis. Specifically, our method assigns distinct importance scores to words based on whether they originate from word, sentence, or passage-level references. Extensive experiments demonstrate that Word2Passage outperforms existing methods across various datasets and LLM configurations, effectively enhancing both retrieval accuracy and generation quality. The code is publicly available at https://github.com/DISL-Lab/Word2Passage

The exponential growth of offensive language on social media tends to fuel online harassment and challenges detection mechanisms. Hate speech detection is commonly treated as a monolingual or multilingual sentence-level classification task. However, profane language tends to contain code-mixing, a combination of more than one language, which requires a more nuanced detection approach than binary classification. A general lack of available code-mixed datasets aggravates the problem. To address this issue, we propose five word-level annotated hate speech datasets, EN and DE from social networks, one subset of the DE-EN Offensive Content Detection Code-Switched Dataset, one DE-EN code-mixed German rap lyrics held-out test set, and a cross-domain held-out test set. We investigate the capacity of fine-tuned German-only, German-English bilingual, and German-English code-mixed token classification XLM-R models to generalize to code-mixed hate speech in German rap lyrics in zero-shot domain transfer as well as across different domains. The results show that bilingual fine-tuning facilitates not only the detection of code-mixed hate speech, but also neologisms, addressing the inherent dynamics of profane language use.

Human readers can efficiently comprehend scrambled words, a phenomenon known as Typoglycemia, primarily by relying on word form; if word form alone is insufficient, they further utilize contextual cues for interpretation. While advanced large language models (LLMs) exhibit similar abilities, the underlying mechanisms remain unclear. To investigate this, we conduct controlled experiments to analyze the roles of word form and contextual information in semantic reconstruction and examine LLM attention patterns. Specifically, we first propose SemRecScore, a reliable metric to quantify the degree of semantic reconstruction, and validate its effectiveness. Using this metric, we study how word form and contextual information influence LLMs’ semantic reconstruction ability, identifying word form as the core factor in this process. Furthermore, we analyze how LLMs utilize word form and find that they rely on specialized attention heads to extract and process word form information, with this mechanism remaining stable across varying levels of word scrambling. This distinction between LLMs’ fixed attention patterns primarily focused on word form and human readers’ adaptive strategy in balancing word form and contextual information provides insights into enhancing LLM performance by incorporating human-like, context-aware mechanisms. Code is available on: https://github.com/Aurora-cx/TypoLLM.

The goal of conversational product search (CPS) is to develop an intelligent, chat-based shopping assistant that can directly interact with customers to understand shopping intents, ask clarification questions, and find relevant products. However, training such assistants is hindered mainly due to the lack of reliable and large-scale datasets. Prior human-annotated CPS datasets are extremely small in size and lack integration with real-world product search systems. We propose a novel approach, TRACER, which leverages large language models (LLMs) to generate realistic and natural conversations for different shopping domains. TRACER’s novelty lies in grounding the generation to dialogue plans, which are product search trajectories predicted from a decision tree model, that guarantees relevant product discovery in the shortest number of search conditions. We also release the first target-oriented CPS dataset Wizard of Shopping (WoS), containing highly natural and coherent conversations (3.6k) from three shopping domains. Finally, we demonstrate the quality and effectiveness of WoS via human evaluations and downstream tasks.

Large Language Models (LLMs) have achieved impressive results across a broad array of tasks, yet their capacity for complex, domain-specific mathematical reasoning—particularly in wireless communications—remains underexplored. In this work, we introduce WirelessMathBench, a novel benchmark specifically designed to evaluate LLMs on mathematical modeling challenges to wireless communications engineering. Our benchmark consists of 587 meticulously curated questions sourced from 40 state-of-the-art research papers, encompassing a diverse spectrum of tasks ranging from basic multiple-choice questions to complex equation completion tasks, including both partial and full completions, all of which rigorously adhere to physical and dimensional constraints. Through extensive experimentation with leading LLMs, we observe that while many models excel in basic recall tasks, their performance degrades significantly when reconstructing partially or fully obscured equations, exposing fundamental limitations in current LLMs. Even DeepSeek-R1, the best performer on our benchmark, achieves an average accuracy of only 38.05%, with a mere 7.83% success rate in full equation completion. By publicly releasing WirelessMathBench along with the evaluation toolkit, we aim to advance the development of more robust, domain-aware LLMs for wireless system analysis and broader engineering applications.

Graphical User Interface (GUI) automation relies on accurate GUI grounding. However, obtaining large-scale, high-quality labeled data remains a key challenge, particularly in desktop environments like Windows Operating System (OS). Existing datasets primarily focus on structured web-based elements, leaving a gap in real-world GUI interaction data for non-web applications. To address this, we introduce a new framework that leverages LLMs to generate large-scale GUI grounding data, enabling automated and scalable labeling across diverse interfaces. To ensure high accuracy and reliability, we manually validated and refined 5,000 GUI coordinate-instruction pairs, creating WinSpot—the first benchmark specifically designed for GUI grounding tasks in Windows environments. WinSpot provides a high-quality dataset for training and evaluating visual GUI agents, establishing a foundation for future research in GUI automation across diverse and unstructured desktop environments.

Documents are fundamental to preserving and disseminating information, often incorporating complex layouts, tables, and charts that pose significant challenges for automatic document understanding (DU). While vision-language large models (VLLMs) have demonstrated improvements across various tasks, their effectiveness in processing long-context vision inputs remains unclear. This paper introduces WikiMixQA, a benchmark comprising 1,000 multiple-choice questions (MCQs) designed to evaluate cross-modal reasoning over tables and charts extracted from 4,000 Wikipedia pages spanning seven distinct topics. Unlike existing benchmarks, WikiMixQA emphasizes complex reasoning by requiring models to synthesize information from multiple modalities. We evaluate 12 state-of-the-art vision-language models, revealing that while proprietary models achieve ~70% accuracy when provided with direct context, their performance deteriorates significantly when retrieval from long documents is required. Among these, GPT-4-o is the only model exceeding 50% accuracy in this setting, whereas open-source models perform considerably worse, with a maximum accuracy of 27%. These findings underscore the challenges of long-context, multi-modal reasoning and establish WikiMixQA as a crucial benchmark for advancing document understanding research.

We introduce WiCkeD, a simple method to increase the complexity of existing multiple-choice benchmarks by randomly replacing a choice with “None of the above”, a method often used in educational tests. We show that WiCkeD can be automatically applied to any existing benchmark, making it more challenging. We apply WiCkeD to 6 popular benchmarks and use it to evaluate 18 open-weight LLMs. The performance of the models drops12.1 points on average with respect to the original versions of the datasets. When using chainof-thought on 3 MMLU datasets, the performance drop for the WiCkeD variant is similar to the one observed when using the LLMs directly, showing that WiCkeD is also challenging for models with enhanced reasoning abilities. WiCkeD also uncovers that some models are more sensitive to the extra reasoning required, providing additional information with respect to the original benchmarks.We relase our code and data at github.com/anonymized.

Vision Language Models (VLMs) have achieved remarkable progress in multimodal tasks, yet they often struggle with visual arithmetic, seemingly simple capabilities like object counting or length comparison, which are essential for relevant complex tasks like chart understanding and geometric reasoning. In this work, we first investigate the root causes of this deficiency through a suite of probing tasks focusing on basic visual arithmetic. Our analysis reveals that while pre-trained vision encoders typically capture sufficient information, the text decoder often fails to decode it correctly for arithmetic reasoning. To address this, we propose CogAlign, a novel post-training strategy inspired by Piaget’s theory of cognitive development. CogAlign trains VLMs to recognize invariant properties under visual transformations. We demonstrate that this approach significantly improves the performance of three diverse VLMs on our proposed probing tasks. Furthermore, CogAlign enhances performance by an average of 4.6% on CHOCOLATE and 2.9% on MATH-VISION, outperforming or matching supervised fine-tuning methods while requiring only 60% less training data. These results highlight the effectiveness and generalizability of CogAlign in improving fundamental visual arithmetic capabilities and their transfer to downstream tasks.

Large Language Models (LLMs) are valued for their strong performance across various tasks, but they also produce inaccurate or misleading outputs. Uncertainty Estimation (UE) quantifies the model’s confidence and helps users assess response reliability. However, existing UE methods have not been thoroughly examined in scenarios like Retrieval-Augmented Generation (RAG), where the input prompt includes non-parametric knowledge. This paper shows that current UE methods cannot reliably estimate the correctness of LLM responses in the RAG setting. We propose an axiomatic framework to identify deficiencies in existing UE methods. Our framework introduces five constraints that an effective UE method should meet after incorporating retrieved documents into the LLM’s prompt. Experimental results reveal that no existing UE method fully satisfies all the axioms, explaining their suboptimal performance in RAG. We further introduce a simple yet effective calibration function based on our framework, which not only satisfies more axioms than baseline methods but also improves the correlation between uncertainty estimates and correctness.

The safety alignment of large language models (LLMs) remains vulnerable, as their initial behavior can be easily jailbroken by even relatively simple attacks. Since infilling a fixed template between the input instruction and initial model output is a common practice for existing LLMs, we hypothesize that this template is a key factor behind their vulnerabilities: LLMs’ safety-related decision-making overly relies on the aggregated information from the template region, which largely influences these models’ safety behavior. We refer to this issue as template-anchored safety alignment. In this paper, we conduct extensive experiments and verify that template-anchored safety alignment is widespread across various aligned LLMs. Our mechanistic analyses demonstrate how it leads to models’ susceptibility when encountering inference-time jailbreak attacks. Furthermore, we show that detaching safety mechanisms from the template region is promising in mitigating vulnerabilities to jailbreak attacks. We encourage future research to develop more robust safety alignment techniques that reduce reliance on the template region.

Despite the remarkable successes of Large Language Models (LLMs), the underlying Transformer architecture has inherent limitations in handling complex reasoning tasks. Chain-of-Thought (CoT) prompting has emerged as a practical workaround, but most CoT-based methods rely on a single generic prompt like “think step by step,” with no task-specific adaptation. These approaches expect the model to discover an effective reasoning path on its own, forcing it to search through a vast prompt space. In contrast, many work has explored task-specific prompt designs to boost performance. However, these designs are typically developed through trial and error, lacking a theoretical ground. As a result, prompt engineering remains largely ad hoc and unguided.In this paper, we provide a theoretical framework that explains why some prompts succeed while others fail. We show that prompts function as selectors, extracting specific task-relevant information from the model’s full hidden state during CoT reasoning. Each prompt defines a unique trajectory through the answer space, and the choice of this trajectory is crucial for task performance and future navigation in the answer space.We analyze the complexity of finding optimal prompts and the size of the prompt space for a given task. Our theory reveals principles behind effective prompt design and shows that naive CoT—using model-self-guided prompt like “think step by step” —can severely hinder performance. Showing that optimal prompt search can lead to over a 50% improvement on reasoning tasks through experiments, our work provide a theoretical foundation for prompt engineering.