ACL 2025 论文列表

Current Multimodal Large Language Models (MLLMs) excel in general visual reasoning but remain underexplored in Abstract Visual Reasoning (AVR), which demands higher-order reasoning to identify abstract rules beyond simple perception. Existing AVR benchmarks focus on single-step reasoning, emphasizing the end result but neglecting the multi-stage nature of reasoning process. Past studies found MLLMs struggle with these benchmarks, but it doesn’t explain how they fail. To address this gap, we introduce MultiStAR, a Multi-Stage AVR benchmark, based on RAVEN, designed to assess reasoning across varying levels of complexity. Additionally, existing metrics like accuracy only focus on the final outcomes while do not account for the correctness of intermediate steps. Therefore, we propose a novel metric, MSEval, which considers the correctness of intermediate steps in addition to the final outcomes. We conduct comprehensive experiments on MultiStAR using 17 representative close-source and open-source MLLMs. The results reveal that while existing MLLMs perform adequately on basic perception tasks, they continue to face challenges in more complex rule detection stages. The dataset and code will be available after acceptance.

Large language models (LLMs) have shown impressive few-shot generalization on many tasks via in-context learning (ICL). Despite their success in showing such emergent abilities, the scale and complexity of larger models also lead to unprecedentedly high computational demands and deployment challenges. In reaction, researchers explore transferring the powerful capabilities of larger models to more efficient and compact models by typically aligning the output of smaller (student) models with that of larger (teacher) models. Existing methods either train student models on the generated outputs of teacher models or imitate their token-level probability distributions. However, these distillation methods pay little to no attention to the input, which also plays a crucial role in ICL. Based on the finding that the performance of ICL is highly sensitive to the selection of demonstration examples, we propose Bidirectional Alignment (BiAlign) to fully leverage the models’ preferences for ICL examples to improve the ICL abilities of student models. Specifically, we introduce the alignment of input preferences between student and teacher models by incorporating a novel ranking loss, in addition to aligning the token-level output distribution. With extensive experiments and analysis, we demonstrate that BiAlign can consistently outperform existing baselines on a variety of tasks involving language understanding, reasoning, and coding.

Evaluating models on large benchmarks can be very resource-intensive, especially during a period of rapid model evolution. Existing efficient evaluation methods estimate the performance of target models by testing them on a small, static coreset derived from the publicly available evaluation results of source models, which are separate from the target models. However, these approaches rely on the assumption that target models have high prediction consistency with source models, which doesn’t generalize well in practice. To fill this gap, we propose TailoredBench, a method that conducts customized evaluation tailored to each target model. Specifically, a Global-coreset is first constructed as a probe to identify the most consistent source models for each target model with an adaptive source model selection strategy. Afterwards, a scalable K-Medoids clustering algorithm is proposed to extend the Global-coreset to a tailored Native-coreset for each target model. According to the predictions on respective Native-coreset, we estimate the overall performance of target models with a calibrated estimation strategy. Comprehensive experiments on five benchmarks across over 300 models demonstrate that compared to best performing baselines, TailoredBench achieves an average reduction of 31.4% in MAE of accuracy estimates under the same inference budgets, showcasing strong effectiveness and generalizability.

In-Context Reinforcement Learning (ICRL) is a frontier paradigm to solve Reinforcement Learning (RL) problems in the foundation-model era. While ICRL capabilities have been demonstrated in transformers through task-specific training, the potential of large language models (LLMs) out of the box remains largely unexplored. This paper investigates whether LLMs can generalize cross-domain to perform ICRL on the Dueling Bandits (DB) problem, a stateless preference-based RL setting. We find that top-performing LLMs exhibit a notable zero-shot capacity for relative decision-making, which translates to low short-term weak regret across all DB environments by quickly including the best arm in duels. However, an optimality gap still exists between LLMs and classic DB algorithms in terms of strong regret. LLMs struggle to converge and consistently exploit even when explicitly prompted to do so, and they are sensitive to prompt variations. To bridge this gap, we propose an agentic-flow framework—LLM with Enhanced Algorithmic Dueling (LEAD)—which integrates off-the-shelf DB algorithm support with LLM agents through fine-grained adaptive interplay. We show that LEAD inherits theoretical guarantees from classic DB algorithms on both weak and strong regret. We validate its efficacy and robustness even with noisy and adversarial prompts. The design of such an agentic framework sheds light on how to enhance the trustworthiness of general-purpose LLMs generalized to in-context decision-making tasks.

We study a subtype of implicitly abusive language, namely non-negative sentences about identity groups (e.g. “Women make good cooks”), and introduce a novel dataset of such utterances. Not only do we profile such abusive sentences, but since our dataset includes different semantic variants of the same characteristic attributed to an identity group, we can also systematically study the impact of varying degrees of generalization and perspective framing. Similarly, we switch identity groups to assess whether the characteristic described in a sentence is inherently abusive. We also report on classification experiments.

Automatic N-gram based metrics such as ROUGE are widely used for evaluating generative tasks such as summarization. While these metrics are considered indicative (even if imperfect), of human evaluation for English, their suitability for other languages remains unclear. To address this, in this paper we systematically assess evaluation metrics for generation — both n-gram-based and neural-based— to assess their effectiveness across languages and tasks. Specifically, we design a large-scale evaluation suite across eight languages from four typological families — agglutinative, isolating, low-fusional, and high-fusional — from both low- and high-resource languages, to analyze their correlations with human judgments. Our findings highlight the sensitivity of the evaluation metric to the language type at hand. For example, for fusional languages, n-gram-based metrics demonstrate a lower correlation with human assessments, compared to isolating and agglutinative languages. We also demonstrate that tokenization considerations can significantly mitigate this for fusional languages with rich morphology, up to reversing such negative correlations. Additionally, we show that neural-based metrics specifically trained for evaluation, such as COMET, consistently outperform other neural metrics and correlate better than ngrmas metrics with human judgments in low-resource languages. Overall, our analysis highlights the limitations of n-gram metrics for fusional languages and advocates for investment in neural-based metrics trained for evaluation tasks.

Knowledge distillation (KD) compresses large language models (LLMs), known as teacher models, into lightweight versions called student models, enabling efficient inference and downstream applications. However, prevailing approaches accomplish this by predominantly focusing on matching the final output distributions of student/teacher models. Drawing on the perspective that transformers can be viewed as discretizing ordinary differential equation (ODEs) on integer time steps (corresponding to layer indices), where intermediate features evolve across layers, we argue that effective KD requires aligning the entire feature dynamics between teacher and student models, which we call feature dynamics distillation (FDD). This alignment involves matching both the feature trajectory and its first-order derivative, rather than just the final states. Our approach extends the original KD objective with two additional loss terms: layer-wise feature KD, which matches discretized feature trajectory, and layer feature delta KD, which matches first-order changes in features across adjacent layers. Extensive experiments on various tasks validate the effectiveness of our distillation method.

In-context learning (ICL) is an important yet not fully understood ability of pre-trained large language models (LLMs). It can greatly enhance task performance using a few examples, termed demonstrations, without fine-tuning. Although effective in question answering, ICL often underperforms in long-form generation tasks such as summarization. Under appropriately realistic assumptions, we empirically and theoretically show that ICL demonstrations alone are insufficient to teach LLMs the task’s language and format distributions for generation. We argue for explicit exposure to the task distributions and hypothesize that defining them by prompting enhances model performance. To this end, we present LongGuide, which efficiently generates two parallel streams of guidelines capturing task language and format properties: (i) Metric Guidelines (MGs) that instruct models to optimize self-evaluated metrics; and (ii) Output Constraint Guidelines (OCGs) that constrain generation at both token and sentence levels. LongGuide automatically selects the best combination of guidelines, improving both strong open- and closed-source LLMs by over 5% in both zero- and few-shot settings. We show that LongGuide is generalizable, learnable by weak models to enhance strong ones, and integrates synergistically with automatic prompt optimizers.

Large Language Models (LLMs) exhibit impressive capabilities in In-Context Learning (ICL) but are prone to label bias—an undesirable tendency to favor certain answers. Existing calibration methods mitigate bias by leveraging in-domain data, yet such data is often unavailable in real-world scenarios. To address this limitation, we propose SDC (Synthetic Data Calibration), a simple-yet-effective approach that generates synthetic in-domain data from a few in-context demonstrations and utilizes it for calibration. By approximating the benefits of real in-domain data, SDC effectively reduces label bias without requiring access to actual domain-specific inputs. Experimental evaluations on 279 classification and multiple-choice tasks from the Super-NaturalInstructions benchmark. The results show that SDC significantly reduces label bias, achieving an average Bias Score reduction of 57.5%, and outperforming all competitive baselines. Moreover, when combined with Leave-One-Out Calibration (LOOC), further improves performance, underscoring its effectiveness and generalizability in enhancing the reliability of LLMs.

Multi-agent collaboration has emerged as a pivotal paradigm for addressing complex, distributed tasks in large language model (LLM)-driven applications. While prior research has focused on high-level architectural frameworks, the granular mechanisms governing agents—critical to performance and scalability—remain underexplored. This study systematically investigates four dimensions of collaboration strategies: (1) agent governance, (2) participation control, (3) interaction dynamics, and (4) dialogue history management. Through rigorous experimentation under two context-dependent scenarios—Distributed Evidence Integration (DEI) and Structured Evidence Synthesis (SES)—we quantify the impact of these strategies on both task accuracy and computational efficiency. Our findings reveal that centralized governance, instructor-led participation, ordered interaction patterns, and instructor-curated context summarization collectively optimize the trade-off between decision quality and resource utilization with the support of the proposed Token-Accuracy Ratio (TAR). This work establishes a foundation for designing adaptive, scalable multi-agent systems, shifting the focus from structural novelty to strategic interaction mechanics.

This paper presents ICAT, an evaluation framework for measuring coverage of diverse factual information in long-form text generation. ICAT breaks down a long output text into a list of atomic claims and not only verifies each claim through retrieval from a (reliable) knowledge source, but also computes the alignment between the atomic factual claims and various aspects expected to be presented in the output. We study three implementations of the ICAT framework, each with a different assumption on the availability of aspects and alignment method. By adopting data from the diversification task in the TREC Web Track and the ClueWeb corpus, we evaluate the ICAT framework. We demonstrate strong correlation with human judgments and provide comprehensive evaluation across multiple state-of-the-art LLMs. Our framework further offers interpretable and fine-grained analysis of diversity and coverage. Its modular design allows for easy adaptation to different domains and datasets, making it a valuable tool for evaluating the qualitative aspects of long-form responses produced by LLMs.

When exposed to complex queries containing multiple conditions, today’s large language models (LLMs) tend to produce responses that only partially satisfy the query while neglecting certain conditions. We, therefore, introduce the concept of Intent Hallucination, a phenomenon where LLMs either omit (failing to address certain parts) or misinterpret (responding to invented query parts) elements of the given query, leading to responses misaligned with the original query. To systematically evaluate intent hallucination, we introduce FAITHQA, a novel benchmark for intent hallucination that contains 20,068 problems, covering both query-only and retrieval-augmented generation (RAG) setups with varying topics and difficulty. FAITHQA is the first hallucination benchmark that goes beyond factual verification, tailored to identify the fundamental cause of intent hallucination. By evaluating various LLMs on FAITHQA, we find that (1) intent hallucination is a common issue even for state-of-the-art models, and (2) such a phenomenon stems from omission or misinterpretation of LLMs. To facilitate future research, we introduce an automatic LLM generation evaluation metric, named INTENT CONSTRAINT, for detecting intent hallucination. Human evaluation results demonstrate that INTENT CONSTRAINT is closer to human performance for intent hallucination compared to baselines.

The rapid advancement of large language models (LLMs) has revolutionized role-playing, enabling the development of general role-playing models. However, current role-playing training has two significant issues: (I) Using a predefined role profile to prompt dialogue training for specific scenarios usually leads to biases and even conflicts between the dialogue and the profile, resulting in training biases. (II) Models learn to imitate the role based solely on the profile, neglecting profile-dialogue alignment at the sentence level. To overcome the aforementioned hurdles, we propose a novel framework **Beyond Dialogue**, which introduces “beyond dialogue” tasks to align dialogue with profile traits for each scenario, eliminating biases during training. Furthermore, the framework achieves a sentence-level fine-grained alignment between profile and dialogue through an innovative prompting mechanism that generates reasoning data for training. Moreover, the aforementioned methods are fully automated and low-cost. Experimental results demonstrate our model excels in adhering to role profiles, outperforming most proprietary general and specialized role-playing baselines. The code and data are provided in https://github.com/yuyouyu32/BeyondDialogue.

People naturally vary in their annotations for subjective questions and some of this variation is thought to be due to the person’s sociodemographic characteristics. LLMs have also been used to label data, but recent work has shown that models perform poorly when prompted with sociodemographic attributes, suggesting limited inherent sociodemographic knowledge. Here, we ask whether LLMs can be trained to be accurate sociodemographic models of annotator variation. Using a curated dataset of five tasks with standardized sociodemographics, we show that models do improve in sociodemographic prompting when trained but that this performance gain is largely due to models learning annotator-specific behaviour rather than sociodemographic behaviours. Across all tasks, our results suggest that models learn little meaningful connection between sociodemographics and annotation, raising doubts about the current use of LLMs for simulating sociodemographic variation and behaviour.

The field of neural machine translation (NMT) has changed with the advent of large language models (LLMs). Much of the recent emphasis in natural language processing (NLP) has been on modeling machine translation and many other problems using a single pre-trained Transformer decoder, while encoder-decoder architectures, which were the standard in earlier NMT models, have received relatively less attention. In this paper, we explore translation models that are universal, efficient, and easy to optimize, by marrying the world of LLMs with the world of NMT. We apply LLMs to NMT encoding and leave the NMT decoder unchanged. We also develop methods for adapting LLMs to work better with the NMT decoder. Furthermore, we construct a new dataset involving multiple tasks to assess how well the machine translation system generalizes across various tasks. Evaluations on the WMT and our datasets show that results using our method match or surpass a range of baselines in terms of translation quality, but achieve 2.4 ∼ 6.5 × inference speedups and a 75% reduction in the memory footprint of the KV cache. It also demonstrates strong generalization across a variety of translation-related tasks.

Datasets used for emotion recognition tasks typically contain overt cues that can be used in predicting the emotions expressed in a text. However, one challenge is that texts sometimes contain covert contextual cues that are rich in affective semantics, which warrant higher-order reasoning abilities to infer emotional states, not simply the emotions conveyed. This study advances beyond surface-level perceptual features to investigate how large language models (LLMs) reason about others’ emotional states using contextual information, within a Theory-of-Mind (ToM) framework. Grounded in Cognitive Appraisal Theory, we curate a specialized ToM evaluation dataset to assess both forward reasoning—from context to emotion—and backward reasoning—from emotion to inferred context. We showed that LLMs can reason to a certain extent, although they are poor at associating situational outcomes and appraisals with specific emotions. Our work highlights the need for psychological theories in the training and evaluation of LLMs in the context of emotion reasoning.

In natural language processing (NLP) and computer vision (CV), the successful application of foundation models across diverse tasks has demonstrated their remarkable potential. However, despite the rich structural and textual information embedded in knowledge graphs (KGs), existing research of foundation model for KG has primarily focused on their structural aspects, with most efforts restricted to in-KG tasks (e.g., knowledge graph completion, KGC). This limitation has hindered progress in addressing more challenging out-of-KG tasks. In this paper, we introduce MERRY, a foundation model for general knowledge graph reasoning, and investigate its performance across two task categories: in-KG reasoning tasks (e.g., KGC) and out-of-KG tasks (e.g., KG question answering, KGQA). We not only utilize the structural information, but also the textual information in KGs. Specifically, we propose a multi-perspective Conditional Message Passing (CMP) encoding architecture to bridge the gap between textual and structural modalities, enabling their seamless integration. Additionally, we introduce a dynamic residual fusion module to selectively retain relevant textual information and a flexible edge scoring mechanism to adapt to diverse downstream tasks. Comprehensive evaluations on 28 datasets demonstrate that MERRY outperforms existing baselines in most scenarios, showcasing strong reasoning capabilities within KGs and excellent generalization to out-of-KG tasks such as KGQA.

Web browsers are a portal to the internet, where much of human activity is undertaken. Thus, there has been significant research work in AI agents that interact with the internet through web browsing.However, there is also another interface designed specifically for machine interaction with online content: application programming interfaces (APIs). In this paper we ask – *what if we were to take tasks traditionally tackled by Browsing Agents, and give AI agents access to APIs*?To do so, we propose two varieties of agents: (1) an API-calling agent that attempts to perform online tasks through APIs only, similar to traditional coding agents, and (2) a Hybrid Agent that can interact with online data through both web browsing and APIs.In experiments on WebArena, a widely-used and realistic benchmark for web navigation tasks, we find that API-Based Agents outperform web Browsing Agents.Hybrid Agents out-perform both others nearly uniformly across tasks, resulting in a more than 24.0% absolute improvement over web browsing alone, achieving a success rate of 38.9%, the SOTA performance among task-agnostic agents.These results strongly suggest that when APIs are available, they present an attractive alternative to relying on web browsing alone.

Role-playing enables large language models (LLMs) to engage users in immersive and personalized interactions, but it also introduces significant safety risks. Existing role-play fine-tuning techniques improve role adaptability but may degrade safety performance, particularly for villainous characters. In this work, we conduct the first comprehensive assessment of role-play fine-tuning risks by training 95 role-specific LLMs using RoleBench. Our experiments reveal that role-play fine-tuning leads to a noticeable decline in safety performance, with safety risks varying based on character traits. To tackle this challenge, we propose Safety-Aware Role-Play Fine-Tuning (SaRFT), a novel method designed to balance role-playing capabilities and safety. Extensive experiments on LLaMA-3-8B-Instruct, Gemma-2-9B-it, and Qwen2.5-7B-Instruct demonstrate that SaRFT consistently outperforms state-of-the-art baselines under both LoRA and full-parameter fine-tuning settings. Our findings highlight the necessity of role-adaptive safety measures and provide insights into mitigating role-specific safety risks in role-playing LLMs.