ACL 2025 论文列表

While Large Language Models (LLMs) excel at many natural language processing tasks, they often suffer from factual inaccuracies in knowledge-intensive scenarios. Integrating external knowledge resources, particularly knowledge graphs (KGs), provides a transparent and updatable foundation for more reliable reasoning. Knowledge Base Question Answering (KBQA), which queries and reasons over KGs, is central to this effort, especially for complex, multi-hop queries. However, multi-hop reasoning poses two key challenges: (1) maintaining coherent reasoning paths, and (2) avoiding prematurely discarding critical multi-hop connections. To address these issues, we introduce iQUEST, a question-guided KBQA framework that iteratively decomposes complex queries into simpler sub-questions, ensuring a structured and focused reasoning trajectory. Additionally, we integrate a Graph Neural Network (GNN) to look ahead and incorporate 2-hop neighbor information at each reasoning step. This dual approach strengthens the reasoning process, enabling the model to explore viable paths more effectively. Detailed experiments demonstrate the consistent improvement delivered by iQUEST across four benchmark datasets and four LLMs.

Understanding how individuals perceive and react to information is fundamental for advancing social and behavioral sciences and developing human-centered AI systems. Current approaches often lack the granular data needed to model these personalized responses, relying instead on aggregated labels that obscure the rich variability driven by individual differences. We introduce iNews, a novel large-scale dataset specifically designed to facilitate the modeling of personalized affective responses to news content. Our dataset comprises annotations from 291 demographically diverse UK participants across 2,899 multimodal Facebook news posts from major UK outlets, with an average of 5.18 annotators per sample. For each post, annotators provide multifaceted labels including valence, arousal, dominance, discrete emotions, content relevance judgments, sharing likelihood, and modality importance ratings. Crucially, we collect comprehensive annotator persona information covering demographics, personality, media trust, and consumption patterns, which explain 15.2% of annotation variance - substantially higher than existing NLP datasets. Incorporating this information yields a 7% accuracy gain in zero-shot prediction and remains beneficial even with 32-shot in-context learning.

Enhancing the fine-grained instance spatiotemporal motion perception capabilities of Video Large Language Models is crucial for improving their temporal and general video understanding. However, current models struggle to perceive detailed and complex instance motions. To address these challenges, we have made improvements from both data and model perspectives. In terms of data, we have meticulously curated iMOVE-IT, the first large-scale instance-motion-aware video instruction-tuning dataset. This dataset is enriched with comprehensive instance motion annotations and spatiotemporal mutual-supervision tasks, providing extensive training for the model’s instance-motion-awareness. Building on this foundation, we introduce iMOVE, an instance-motion-aware video foundation model that utilizes Event-aware Spatiotemporal Efficient Modeling to retain informative instance spatiotemporal motion details while maintaining computational efficiency. It also incorporates Relative Spatiotemporal Position Tokens to ensure awareness of instance spatiotemporal positions. Evaluations indicate that iMOVE excels not only in video temporal understanding and general video understanding but also demonstrates significant advantages in long-term video understanding. We will release the data, code, and model weights after acceptance.

Traditional recommender systems usually take the user-platform paradigm, where users are directly exposed under the control of the platform’s recommendation algorithms. However, the defect of recommendation algorithms may put users in very vulnerable positions under this paradigm. First, many sophisticated models are often designed with commercial objectives in mind, focusing on the platform’s benefits, which may hinder their ability to protect and capture users’ true interests. Second, these models are typically optimized using data from all users, which may overlook individual user’s preferences. Due to these shortcomings, users may experience several disadvantages under the traditional user-platform direct exposure paradigm, such as lack of control over the recommender system, potential manipulation by the platform, echo chamber effects, or lack of personalization for less active users due to the dominance of active users during collaborative learning. Therefore, there is an urgent need to develop a new paradigm to protect user interests and alleviate these issues. Recently, some researchers have introduced LLM agents to simulate user behaviors, these approaches primarily aim to optimize platform-side performance, leaving core issues in recommender systems unresolved. To address these limitations, we propose a new user-agent-platform paradigm, where agent serves as the protective shield between user and recommender system that enables indirect exposure. To this end, we first construct four recommendation datasets, denoted as InstructRec, along with user instructions for each record. To understand user’s intention, we design an Instruction-aware Agent capable of using tools to acquire knowledge from external environments. Moreover, we introduce an Individual Instruction-aware Agent, which incorporates a dynamic memory mechanism to optimize from individual feedback. Results on four datasets demonstrate that consistently achieves an average improvement of 16.6% over SOTA baselines across ranking metrics. Moreover, iAgent mitigates echo chamber effects and effectively alleviates the model bias in disadvantaged users (less-active), serving as a shield between user and recommender systems.

Mixed Boolean-Arithmetic (MBA) obfuscation protects intellectual property by converting programs into forms that are more complex to analyze. However, MBA has been increasingly exploited by malware developers to evade detection and cause significant real-world problems. Traditional MBA deobfuscation methods often consider these expressions as part of a black box and overlook their internal semantic information. To bridge this gap, we propose a truth table, which is an automatically constructed semantic representation of an expression’s behavior that does not rely on external resources. The truth table is a mathematical form that represents the output of expression for all possible combinations of input. We also propose a general and extensible guided MBA deobfuscation framework (gMBA) that modifies a Transformer-based neural encoder-decoder Seq2Seq architecture to incorporate this semantic guidance. Experimental results and in-depth analysis show that integrating expression semantics significantly improves performance and highlights the importance of internal semantic expressions in recovering obfuscated code to its original form.

Current studies have exposed the risk of Large Language Models (LLMs) generating harmful content by jailbreak attacks. However, they overlook that the direct generation of harmful content from scratch is more difficult than inducing LLM to calibrate benign content into harmful forms.In our study, we introduce a novel attack framework that exploits AdVersArial meTAphoR (AVATAR) to induce the LLM to calibrate malicious metaphors for jailbreaking.Specifically, to answer harmful queries, AVATAR adaptively identifies a set of benign but logically related metaphors as the initial seed.Then, driven by these metaphors, the target LLM is induced to reason and calibrate about the metaphorical content, thus jailbroken by either directly outputting harmful responses or calibrating residuals between metaphorical and professional harmful content.Experimental results demonstrate that AVATAR can effectively and transferably jailbreak LLMs and achieve a state-of-the-art attack success rate across multiple advanced LLMs.

Large language models (LLMs) have demonstrated exceptional performance across various applications, but their conversational abilities decline sharply as model size decreases, presenting a barrier to their deployment in resource-constrained environments. Knowledge distillation (KD) with Direct Preference Optimization (DPO) has emerged as a promising approach to enhance the conversational abilities of smaller models using a larger teacher model. However, current methods primarily focus on “black-box” KD, which only uses the teacher’s responses, overlooking the rich distributional information within the teacher’s probability distribution. This paper addresses this gap by introducing daDPO (Distillation-Aware DPO), a novel framework that integrates the teacher’s distributional information into DPO distillation while preserving theoretical guarantees. Our framework offers a unified objective that enhances both preference optimization and distribution-based distillation. We provide rigorous theoretical analysis and empirical validation, showing that daDPO outperforms existing methods in restoring performance for pruned models and enhancing smaller models within the same LLM family. Notably, in in-domain evaluation, our method enables a 20% pruned Vicuna1.5-7B to achieve near-teacher performance (-7.3% preference rate), and allows Qwen2.5-1.5B to occasionally outperform its 7b teacher model (14.0% win rate).

What happens when a named entity recognition (NER) system encounters entities it has never seen before? In practical applications, models must generalize to unseen entity types where labeled training data is either unavailable or severely limited—a challenge that demands zero-shot learning capabilities. While large language models (LLMs) offer extensive parametric knowledge, they fall short in cost-effectiveness compared to specialized small encoders. Existing zero-shot methods predominantly adopt a relaxed definition of the term with potential leakage issues and rely on entity type names for generalization, overlooking the value of richer descriptions for disambiguation. In this work, we introduce ZeroNER, a description-driven framework that enhances hard zero-shot NER in low-resource settings. By leveraging general-domain annotations and entity type descriptions with LLM supervision, ZeroNER enables a BERT-based student model to successfully identify unseen entity types. Evaluated on three real-world benchmarks, ZeroNER consistently outperforms LLMs by up to 16% in F1 score, and surpasses lightweight baselines that use type names alone. Our analysis further reveals that LLMs derive significant benefits from incorporating type descriptions in the prompts.

The advancements in large language models (LLMs) have brought significant progress in NLP tasks. However, if a task cannot be fully described in prompts, the models could fail to carry out the task. In this paper, we propose a simple yet effective method to contextualize a task toward a LLM. The method utilizes (1) open-ended zero-shot inference from the entire dataset, (2) aggregate the inference results, and (3) finally incorporate the aggregated meta-information for the actual task. We show the effectiveness in text clustering tasks, empowering LLMs to perform text-to-text-based clustering and leading to improvements on several datasets. Furthermore, we explore the generated class labels for clustering, showing how the LLM understands the task through data.

This paper introduces PhoAudiobook, a newly curated dataset comprising 941 hours of high-quality audio for Vietnamese text-to-speech. Using PhoAudiobook, we conduct experiments on three leading zero-shot TTS models: VALL-E, VoiceCraft, and XTTS-V2. Our findings demonstrate that PhoAudiobook consistently enhances model performance across various metrics. Moreover, VALL-E and VoiceCraft exhibit superior performance in synthesizing short sentences, highlighting their robustness in handling diverse linguistic contexts. We publicly release PhoAudiobook to facilitate further research and development in Vietnamese text-to-speech.

Stance detection, which aims to identify public opinion towards specific targets using social media data, is an important yet challenging task. With the increasing number of online debates among social media users, conversational stance detection has become a crucial research area. However, existing conversational stance detection datasets are restricted to a limited set of specific targets, which constrains the effectiveness of stance detection models when encountering a large number of unseen targets in real-world applications. To bridge this gap, we manually curate a large-scale, high-quality zero-shot conversational stance detection dataset, named ZS-CSD, comprising 280 targets across two distinct target types. Leveraging the ZS-CSD dataset, we propose SITPCL, a speaker interaction and target-aware prototypical contrastive learning model, and establish the benchmark performance in the zero-shot setting. Experimental results demonstrate that our proposed SITPCL model achieves state-of-the-art performance in zero-shot conversational stance detection. Notably, the SITPCL model attains only an F1-macro score of 43.81%, highlighting the persistent challenges in zero-shot conversational stance detection.

We present ZIPA, a family of efficient speech models that advances the state-of-the-art performance of crosslinguistic phone recognition. We first curated IPA PACK++, a large-scale multilingual speech corpus with 17,000+ hours of normalized phone transcriptions and a novel evaluation set capturing unseen languages and sociophonetic variation. ZIPA, including transducer (ZIPA-T) and CTC-based (ZIPA-CR) variants, leverages the efficient Zipformer backbones and outperforms existing phone recognition systems with much fewer parameters. Further scaling via noisy student training on 11,000+ hours of pseudo-labeled multilingual data yields further improvement. While ZIPA achieves strong performance on benchmarks, error analysis reveals persistent limitations in modeling sociophonetic diversity, underscoring challenges for future research.

Due to the immense resource demands and the involved complex techniques, it is still challenging for successfully pre-training a large language models (LLMs) with state-of-the-art performance. In this paper, we explore the key bottlenecks and designs during pre-training, and make the following contributions: (1) a comprehensive investigation into the factors contributing to training instability; (2) a robust optimization approach designed to mitigate training instability effectively; (3) an elaborate data pipeline that integrates data synthesis, data curriculum, and data selection. By integrating the above techniques, we create a rather low-cost training recipe and use it to pre-train YuLan-Mini, a fully-open base model with 2.4B parameters on 1.08T tokens. Remarkably, YuLan-Mini achieves top-tier performance among models of similar parameter scale, with comparable performance to industry-leading models that require significantly more data. To facilitate reproduction, we release the full details of training recipe and data composition. Project details can be accessed at the following link: https://anonymous.4open.science/r/YuLan-Mini/README.md.

The difficulty intrinsic to a given example, rooted in its inherent ambiguity, is a key yet often overlooked factor in evaluating neural NLP models. We investigate the interplay and divergence among various metrics for assessing intrinsic difficulty, including annotator dissensus, training dynamics, and model confidence. Through a comprehensive analysis using 29 models on three datasets, we reveal that while correlations exist among these metrics, their relationships are neither linear nor monotonic. By disentangling these dimensions of uncertainty, we aim to refine our understanding of data complexity and its implications for evaluating and improving NLP models.

Large Language Models (LLMs) have demonstrated strong reasoning capabilities across various tasks. However, even minor variations in query phrasing, despite preserving the underlying semantic meaning, can significantly affect their performance. To address this, we focus on enhancing LLMs’ awareness of symmetry in query variations and propose syMmetry-ENhanceD (MEND) data augmentation, a data-centric approach that improves the model’s ability to extract useful information from context. Unlike existing methods that emphasize reasoning chain augmentation, our approach improves model robustness at the knowledge extraction stage through query augmentation, enabling more data-efficient training and stronger generalization to Out-of-Distribution (OOD) settings. Extensive experiments on both logical and arithmetic reasoning tasks show that MEND enhances reasoning performance across diverse query variations, providing new insights into improving LLM robustness through structured dataset curation.

Meeting summarization suffers from limited high-quality data, mainly due to privacy restrictions and expensive collection processes. We address this gap with FAME, a dataset of 500 meetings in English and 300 in German produced by MIMIC, our new multi-agent meeting synthesis framework that generates meeting transcripts on a given knowledge source by defining psychologically grounded participant profiles, outlining the conversation, and orchestrating a large language model (LLM) debate. A modular post-processing step refines these outputs, mitigating potential repetitiveness and overly formal tones, ensuring coherent, credible dialogues at scale. We also propose a psychologically grounded evaluation framework assessing naturalness, social behavior authenticity, and transcript difficulties. Human assessments show that FAME approximates real-meeting spontaneity (4.5/5 in naturalness), preserves speaker-centric challenges (3/5 in spoken language), and introduces richer information-oriented difficulty (4/5 points in difficulty). These findings show FAME is a good and scalable proxy for real-world meeting conditions. It enables new test scenarios for meeting summarization research and other conversation-centric applications in tasks requiring conversation data or simulating social scenarios under behavioral constraints.

Precise alignment in Text-to-Image (T2I) systems is crucial for generating visuals that reflect user intent while adhering to ethical and policy standards. Recent controversies, such as the Google Gemini-generated Pope image backlash, highlight the urgent need for robust alignment mechanisms. Building on alignment successes in Large Language Models (LLMs), this paper introduces YinYangAlign, a benchmarking framework designed to evaluate and optimize T2I systems across six inherently contradictory objectives. These objectives highlight core trade-offs, such as balancing faithfulness to prompts with artistic freedom and maintaining cultural sensitivity without compromising creativity. Alongside this benchmark, we propose the Contradictory Alignment Optimization (CAO) framework, an extension of Direct Preference Optimization (DPO), which employs multi-objective optimization techniques to address these competing goals. By leveraging per-axiom loss functions, synergy-driven global preferences, and innovative tools like the Synergy Jacobian, CAO achieves superior alignment across all objectives. Experimental results demonstrate significant improvements in fidelity, diversity, and ethical adherence, setting new benchmarks for the field. This work provides a scalable, effective approach to resolving alignment challenges in T2I systems while offering insights into broader AI alignment paradigms.

Large Language Models (LLMs) drive scientific question-answering on modern search engines, yet their evaluation robustness remains underexplored. We introduce YESciEval, an open-source framework that combines fine-grained rubric-based assessment with reinforcement learning to mitigate optimism bias in LLM evaluators. We release multidisciplinary scienceQ&A datasets, including adversarial variants, with evaluation scores from multiple LLMs. Independent of proprietary models and human feedback, our approach enables scalable, cost-free evaluation. By advancing reliable LLM-as-a-judge models, this work supports AI alignment and fosters robust, transparent evaluation essential for scientific inquiry.

Solving financial problems demands complex reasoning, multimodal data processing, and a broad technical understanding, presenting unique challenges for current large language models (LLMs). We introduce **XFinBench**, a novel benchmark with 4,235 examples designed to evaluate LLM’s ability in solving comple**X**, knowledge-intensive **Fin**ancial problems across diverse graduate-level finance topics with multi-modal context. We identify five core capabilities of LLMs using XFinBench, i.e., _terminology understanding_, _temporal reasoning_, _future forecasting_, _scenario planning_, and _numerical modelling_. Upon XFinBench, we conduct extensive experiments on 18 leading models. The result shows that o1 is the best-performing text-only model with an overall accuracy of 67.3%, but still lags significantly behind human experts with 12.5%, especially in temporal reasoning and scenario planning capabilities. We further construct a knowledge bank with 3,032 finance terms for knowledge augmentation analysis, and find that relevant knowledge to the question only brings consistent accuracy improvements to small open-source model. Additionally, our error analysis reveals that rounding errors during calculation and blindness to position and intersection of curves in the image are two primary issues leading to model’s poor performance in calculating and visual-context questions, respectively.