ACL 2025 论文列表

Large Language Models (LLMs) have been evaluated using diverse question types, e.g., multiple-choice, true/false, and short/long answers. This study answers an unexplored question about the impact of different question types on LLM accuracy on reasoning tasks. We investigate the performance of five LLMs on three different types of questions using quantitative and deductive reasoning tasks. The performance metrics include accuracy in the reasoning steps and choosing the final answer. Key Findings: (1) Significant differences exist in LLM performance across different question types. (2) Reasoning accuracy does not necessarily correlate with the final selection accuracy. (3) The number of options and the choice of words, influence LLM performance.

Large Language Models (LLMs) have become essential for offensive language detection, yet their ability to handle annotation disagreement remains underexplored. Disagreement samples, which arise from subjective interpretations, pose a unique challenge due to their ambiguous nature. Understanding how LLMs process these cases, particularly their confidence levels, can offer insight into their alignment with human annotators. This study systematically evaluates the performance of multiple LLMs in detecting offensive language at varying levels of annotation agreement. We analyze binary classification accuracy, examine the relationship between model confidence and human disagreement, and explore how disagreement samples influence model decision-making during few-shot learning and instruction fine-tuning. Our findings reveal that LLMs struggle with low-agreement samples, often exhibiting overconfidence in these ambiguous cases. However, utilizing disagreement samples in training improves both detection accuracy and model alignment with human judgment. These insights provide a foundation for enhancing LLM-based offensive language detection in real-world moderation tasks.

Discourse particles are crucial elements that subtly shape the meaning of text. These words, often polyfunctional, give rise to nuanced and often quite disparate semantic/discourse effects,as exemplified by the diverse uses of the particle *just* (e.g., exclusive, temporal, emphatic). This work investigates the capacity of LLMs to distinguish the fine-grained senses of English *just*, a well-studied example in formal semantics, using data meticulously created and labeled by expert linguists. Our findings reveal that while LLMs exhibit some ability to differentiate between broader categories, they struggle to fully capture more subtle nuances, highlighting a gap in their understanding of discourse particles.

In the stance detection task, a text is classified as either favorable, opposing, or neutral towards a target. Prior work suggests that the use of external information, e.g., excerpts from Wikipedia, improves stance detection performance. However, whether or not such information can benefit large language models (LLMs) remains an unanswered question, despite their wide adoption in many reasoning tasks. In this study, we conduct a systematic evaluation on how Wikipedia and web search external information can affect stance detection across eight LLMs and in three datasets with 12 targets. Surprisingly, we find that such information degrades performance in most cases, with macro F1 scores dropping by up to 27.9%. We explain this through experiments showing LLMs’ tendency to align their predictions with the stance and sentiment of the provided information rather than the ground truth stance of the given text. We also find that performance degradation persists with chain-of-thought prompting, while fine-tuning mitigates but does not fully eliminate it. Our findings, in contrast to previous literature on BERT-based systems which suggests that external information enhances performance, highlight the risks of information biases in LLM-based stance classifiers.

Machine-generated Text (MGT) detection is crucial for regulating and attributing online texts. While the existing MGT detectors achieve strong performance, they remain vulnerable to simple perturbations and adversarial attacks. To build an effective defense against malicious perturbations, we view MGT detection from a threat modeling perspective, that is, analyzing the model’s vulnerability from an adversary’s point of view and exploring effective mitigations. To this end, we introduce an adversarial framework for training a robust MGT detector, named GREedy Adversary PromoTed DefendER (GREATER). The GREATER consists of two key components: an adversary GREATER-A and a detector GREATER-D. The GREATER-D learns to defend against the adversarial attack from GREATER-A and generalizes the defense to other attacks. GREATER-A identifies and perturbs the critical tokens in embedding space, along with greedy search and pruning to generate stealthy and disruptive adversarial examples. Besides, we update the GREATER-A and GREATER-D synchronously, encouraging the GREATER-D to generalize its defense to different attacks and varying attack intensities. Our experimental results across 10 text perturbation strategies and 6 adversarial attacks show that our GREATER-D reduces the Attack Success Rate (ASR) by 0.67% compared with SOTA defense methods while our GREATER-A is demonstrated to be more effective and efficient than SOTA attack approaches. Codes and dataset are available in https://github.com/Liyuuuu111/GREATER.

Large language models (LLMs) exhibit remarkable multilingual capabilities despite the extreme language imbalance in the pre-training data. In this paper, we closely examine the reasons behind this phenomenon, focusing on the pre-training corpus. We find that the existence of code-switching, alternating between different languages within a context, is key to multilingual capabilities. We conduct an analysis to investigate code-switching in the pre-training corpus, examining its presence and categorizing it into four types within two quadrants. We then assess its impact on multilingual performance. These types of code-switching data are unbalanced in proportions and demonstrate different effects on facilitating language transfer. To better explore the power of code-switching for language alignment during pre-training, we investigate the strategy of synthetic code-switching. We continuously scale up the synthetic code-switching data and observe remarkable improvements in both benchmarks and representation space. Extensive experiments indicate that incorporating synthetic code-switching data enables better language alignment and generalizes well to high, medium, and low-resource languages with pre-training corpora of varying qualities.

Homans’ Social Exchange Theory (SET) is widely recognized as a basic framework for understanding the formation and emergence of human civilizations and social structures. In social science, this theory is typically studied based on simple simulation experiments or real-world human studies, both of which either lack realism or are too expensive to control. In artificial intelligence, recent advances in large language models (LLMs) have shown promising capabilities in simulating human behaviors. Inspired by these insights, we adopt an interdisciplinary research perspective and propose using LLM-based agents to study Homans’ SET. Specifically, we construct a virtual society composed of three LLM agents and have them engage in a social exchange game to observe their behaviors. Through extensive experiments, we found that Homans’ SET is well validated in our agent society, demonstrating the consistency between the agent and human behaviors. Building on this foundation, we intentionally alter the settings of the agent society to extend the traditional Homans’ SET, making it more comprehensive and detailed. To the best of our knowledge, this paper marks the first step in studying Homans’ SET with LLM-based agents. More importantly, it introduces a novel and feasible research paradigm that bridges the fields of social science and computer science through LLM-based agents. Code is available at https://github.com/Paitesanshi/SET .

Large Multimodal Models (LMMs) have recently demonstrated impressive performance on general video comprehension benchmarks. Nevertheless, for broader applications, the robustness of their temporal analysis capability needs to be thoroughly investigated yet predominantly ignored. Motivated by this, we propose a novel temporal robustness benchmark (TemRobBench), which introduces temporal inconsistency perturbations separately at the visual and textual modalities to assess the robustness of models. We evaluate 16 mainstream LMMs and find that they exhibit over-reliance on prior knowledge and textual context in adversarial environments, while ignoring the actual temporal dynamics in the video. To mitigate this issue, we design panoramic direct preference optimization (PanoDPO), which encourages LMMs to incorporate both visual and linguistic feature preferences simultaneously. Experimental results show that PanoDPO can effectively enhance the model’s robustness and reliability in temporal analysis.

Omnimodal Large Language Models (OLLMs) have shown significant progress in integrating vision and text, but still struggle with integrating vision and audio, often exhibiting suboptimal performance when processing audio queries compared to text queries. This disparity is primarily due to insufficient alignment between vision and audio modalities during training, leading to inadequate attention to visual information when using audio queries. To mitigate this issue, we propose a Self-Knowledge Distillation (Self-KD) training method where the vision-text component of the OLLM serves as the teacher and the vision-audio component as the student. This enables the model to process audio in a manner analogous to its text processing. Our experimental results demonstrate that Self-KD is an effective method for enhancing the vision-audio capabilities of OLLMs by learning from the vision-text components, which subsequently improves the interaction between audio and images and results in improved performance on multimodal tasks.

In this work, we investigate various state-of-the-art (SOTA) speech pre-trained models (PTMs) for their capability to capture prosodic sig-natures of the generative sources for audio deepfake source attribution (ADSD). These prosodic characteristics can be considered oneof major signatures for ADSD, which is unique to each source. So better is the PTM at capturing prosodic signs better the ADSD per-formance. We consider various SOTA PTMs that have shown top performance in different prosodic tasks for our experiments on benchmark datasets, ASVSpoof 2019 and CFAD. x-vector (speaker recognition PTM) attains the highest performance in comparison to allthe PTMs considered despite consisting lowest model parameters. This higher performance can be due to its speaker recognition pre-training that enables it for capturing unique prosodic characteristics of the sources in a better way. Further, motivated from tasks suchas audio deepfake detection and speech recognition, where fusion of PTMs representations lead to improved performance, we explorethe same and propose FINDER for effective fusion of such representations. With fusion of Whisper and x-vector representations through FINDER, we achieved the topmost performance in comparison to all the individual PTMs as well as baseline fusion techniques and attaining SOTA performance.

Multilingual Retrieval-Augmented Generation (mRAG) systems enhance language models by integrating external multilingual information to produce context-aware responses. However, mRAG systems struggle with retrieving relevant information due to linguistic variations between queries and documents, generating inconsistent responses when multilingual sources conflict. In this work, we systematically investigate language preferences in both retrieval and generation of mRAG through a series of experiments. Our analysis indicates that retrievers tend to prefer high-resource and query languages, yet this preference does not consistently improve generation performance. Moreover, we observe that generators prefer the query language or Latin scripts, leading to inconsistent outputs. To overcome these issues, we propose Dual Knowledge Multilingual RAG (DKM-RAG), a simple yet effective framework that fuses translated multilingual passages with complementary model knowledge. Empirical results demonstrate that DKM-RAG mitigates language preference in generation and enhances performance across diverse linguistic settings. Code is available at https://github.com/jeonghyunpark2002/LanguagePreference.git

Recently, inference-time scaling of chain-of-thought (CoT) has been demonstrated as a promising approach for addressing multi-modal reasoning tasks.While existing studies have predominantly centered on text-based thinking, the integration of both visual and textual modalities within the reasoning process remains unexplored.In this study, we pioneer the exploration of inference-time scaling with multi-modal thought, aiming to bridge this gap.To provide a comprehensive analysis, we systematically investigate popular sampling-based and tree search-based inference-time scaling methods on 10 challenging tasks spanning various domains.Besides, we uniformly adopt a consistency-enhanced verifier to ensure effective guidance for both methods across different thought paradigms.Results show that multi-modal thought promotes better performance against conventional text-only thought, and blending the two types of thought fosters more diverse thinking.Despite these advantages, multi-modal thoughts necessitate higher token consumption for processing richer visual inputs, which raises concerns in practical applications.We hope that our findings on the merits and drawbacks of this research line will inspire future works in the field. The code will be released upon acceptance.

Retrieval-augmented generation (RAG) improves Large Language Models (LLMs) by incorporating external information into the response generation process. However, how context-faithful LLMs are and what factors influence LLMs’ context faithfulness remain largely unexplored. In this study, we investigate the impact of memory strength and evidence presentation on LLMs’ receptiveness to external evidence. We quantify the memory strength of LLMs by measuring the divergence in LLMs’ responses to different paraphrases of the same question, which is not considered by previous works. We also generate evidence in various styles to examine LLMs’ behavior. Our results show that for questions with high memory strength, LLMs are more likely to rely on internal memory. Furthermore, presenting paraphrased evidence significantly increases LLMs’ receptiveness compared to simple repetition or adding details. These findings provide key insights for improving retrieval-augmented generation and context-aware LLMs. Our code is available at https://github.com/liyp0095/ContextFaithful.

Aligning Large Language Models (LLMs) with investor decision-making processes under herd behavior is a critical challenge in behavioral finance, which grapples with a fundamental limitation: the scarcity of real-user data needed for Supervised Fine-Tuning (SFT). While SFT can bridge the gap between LLM outputs and human behavioral patterns, its reliance on massive authentic data imposes substantial collection costs and privacy risks. We propose **InvestAlign**, a novel framework that constructs high-quality SFT datasets by leveraging theoretical solutions to similar and simple optimal investment problems rather than the complex scenarios. Our theoretical analysis demonstrates that training LLMs with **InvestAlign**-generated data achieves faster parameter convergence than using real-user data, suggesting superior learning efficiency. Furthermore, we develop **InvestAgent**, an LLM agent fine-tuned with **InvestAlign**, which shows significantly closer alignment to real-user data than pre-SFT models in both simple and complex investment problems. This highlights our proposed **InvestAlign** as a promising approach with the potential to address complex optimal investment problems and align LLMs with investor decision-making processes under herd behavior. Our code is publicly available at https://github.com/thu-social-network-research-group/InvestAlign.

Supervised Fine-Tuning (SFT) and Preference Optimization (PO) are key processes for aligning Language Models (LMs) with human preferences post pre-training. While SFT excels in efficiency and PO in effectiveness, they are often combined sequentially without integrating their optimization objectives. This approach ignores the opportunities to bridge their paradigm gap and take the strengths from both. In this paper, we interpret SFT and PO with two sub-processes — *Preference Estimation* and *Transition Optimization* — defined at token level within the Markov Decision Process (MDP). This modeling shows that SFT is only a special case of PO with inferior estimation and optimization. PO estimates the model’s preference by its entire generation, while SFT only scores model’s subsequent predicted tokens based on prior tokens from ground truth answer. These priors deviates from model’s distribution, hindering the preference estimation and transition optimization. Building on this view, we introduce **Intuitive Fine-Tuning (IFT)** to integrate SFT and PO into a single process. Through a temporal residual connection, IFT brings better estimation and optimization by capturing LMs’ intuitive sense of its entire answers. But it solely relies on a single policy and the same volume of non-preference-labeled data as SFT. Our experiments show that IFT performs comparably or even superiorly to SFT and some typical PO methods across several tasks, particularly those requires generation, reasoning, and fact-following abilities. An explainable Frozen Lake game further validates the effectiveness of IFT for getting competitive policy.

Examining logical inconsistencies among multiple statements (such as collections of sentences or question-answer pairs) is a crucial challenge in machine learning, particularly for ensuring the safety and reliability of models. Traditional methods that rely on 1:1 pairwise comparisons often fail to capture inconsistencies that only emerge when more than two statements are evaluated collectively. To address this gap, we introduce the task of set-consistency verification, an extension of natural language inference (NLI) that assesses the logical coherence of entire sets rather than isolated pairs. Building on this task, we present the Set-Consistency Energy Network (SC-Energy), a novel model that employs a margin-based loss to learn the compatibility among a collection of statements. Our approach not only efficiently verifies inconsistencies and pinpoints the specific statements responsible for logical contradictions, but also significantly outperforms existing methods, including prompting-based LLM models. Furthermore, we release two new datasets: Set-LConVQA and Set-SNLI for set-consistency verification task.

Lexical relation refers to the way words are related within a language. Prior work has demonstrated that pretrained language models (PLMs) can effectively mine lexical relations between word pairs. However, they overlook the potential of graph structures composed of lexical relations, which can be integrated with the semantic knowledge of PLMs. In this work, we propose a parameter-efficient fine-tuning method through graph context, which integrates graph features and semantic representations for lexical relation classification (LRC) and lexical entailment (LE) tasks. Our experiments show that graph features can help PLMs better understand more complex lexical relations, establishing a new state-of-the-art for LRC and LE. Finally, we perform an error analysis, identifying the bottlenecks of language models in lexical relation mining tasks and providing insights for future improvements.

Large language models (LLMs) excel at downstream NLP tasks through in-context learning (ICL) with a few demonstrations of input–label pairs. However, the internal mechanisms behind ICL remain under-explored, particularly the mappings between inputs and labels. In this work, we reverse-engineer ICL by examining input-label mappings: what they are within LLMs, where they function, and how LLMs utilize them. (1) what: We discover input-label mappings stored within a few specific layers in the form of principal components (PCs), which capture human-interpretable and task-related words. (2) where: We propose a PC patching approach to identify the modules where input-label mappings function. Specifically, PC patching automatically crafts counterfactual representations using identified semantic PCs, rather than manually designing counterfactual text, to suppress the behavior related to LLM capability for ICL-related modules. Utilizing PC patching, we identify LLMs apply input-label mappings in a small fraction of attention heads. (3) how: We observe and verify that the identified key heads utilize input-label mappings from demonstrations to generate target labels for new queries. Based on these discoveries, we further show that precisely fine-tuning key ICL-related modules leads to significant improvements across diverse tasks.

We investigate the impacts of NLP research published in top-tier conferences (i.e., ACL, EMNLP, and NAACL) from 1979 to 2024. By analyzing citations from research articles and external sources such as patents, media, and policy documents, we examine how different NLP topics are consumed both within the academic community and by the broader public. Our findings reveal that language modeling has the widest internal and external influence, while linguistic foundations have lower impacts. We also observe that internal and external impacts generally align, but topics like ethics, bias, and fairness show significant attention in policy documents with much fewer academic citations. Additionally, external domains exhibit distinct preferences, with patents focusing on practical NLP applications and media and policy documents engaging more with the societal implications of NLP models.