ACL 2025 Papers

Non-collaborative dialogue involves two participants with conflicting interests engaging in a multi-round dialogue to achieve their own goals. Strategy planning is the key to guiding both participants towards a consensus. Most LLMs-based methods use stimulus prompts or external strategy planners for strategy planning. However, stimulus prompts fail to teach LLMs to plan dialogue strategies explicitly. Moreover, training external strategy planners doesn’t fully account for adversarial interactions, thereby limiting their effectiveness against tough resisters. In this paper, to mitigate the above issues, we propose GAIA, a Game-based Adversarial self-play InterActive training paradigm, which constructs an adversarial two-player (a persuader and a resister) zero-sum game and guides the game to approximate Nash Equilibrium (NE) via reinforcement learning (RL) for the non-collaborative dialogues. First, we design a Chain-of-Mind prompt to reason the resister’s dialogue act step-by-step to plan the persuasive strategies. Secondly, to adversarially improve the persuader, we construct diverse resistant planners and theoretically improve the persuader’s optimal lower bound. Finally, we iteratively optimise their policies via adversarial self-play interactive RL and design an 𝜖-NE verification algorithm to approximate the game’s NE. Experiments on three datasets show that our model obtains state-of-the-art performance.

Recent advances in large language models (LLMs) have demonstrated remarkable capabilities on widely benchmarked high-resource languages. However, linguistic nuances of under-resourced languages remain unexplored. We introduce Batayan, a holistic Filipino benchmark that systematically evaluates LLMs across three key natural language processing (NLP) competencies: understanding, reasoning, and generation. Batayan consolidates eight tasks, three of which have not existed prior for Filipino corpora, covering both Tagalog and code-switched Taglish utterances. Our rigorous, native-speaker-driven adaptation and validation processes ensures fluency and authenticity to the complex morphological and syntactic structures of Filipino, alleviating the pervasive translationese bias in existing Filipino corpora. We report empirical results on a variety of open-source and commercial LLMs, highlighting significant performance gaps that signal the under-representation of Filipino in pre-training corpora, the unique hurdles in modeling Filipino’s rich morphology and construction, and the importance of explicit Filipino language support. Moreover, we discuss the practical challenges encountered in dataset construction and propose principled solutions for building culturally and linguistically-faithful resources in under-represented languages. We also provide a public evaluation suite as a clear foundation for iterative, community-driven progress in Filipino NLP.

Large language models (LLMs) have demonstrated remarkable abilities in various natural language processing areas, but they demand high computation resources which limits their deployment in real-world. Distillation is one technique to solve this problem through either knowledge distillation or task distillation. Both distillation approaches train small models to imitate specific features of LLMs, but they all neglect basic reading education for small models on generic texts that are unrelated to downstream tasks. In this paper, we propose basic reading distillation (BRD) which educates a small model to imitate LLMs basic reading behaviors, such as named entity recognition, question raising and answering, on each sentence. After such basic education, we apply the small model on various tasks including language inference benchmarks and BIG-bench tasks. It shows that the small model can outperform or perform comparable to over 20x bigger LLMs. Analysis reveals that BRD effectively influences the probability distribution of the small model, and has orthogonality to either knowledge distillation or task distillation.

Prompt optimization aims to search for effective prompts that enhance the performance of large language models (LLMs). Although existing prompt optimization methods have discovered effective prompts, they often differ from sophisticated prompts carefully designed by human experts. Prompt design strategies, representing best practices for improving prompt performance, can be key to improving prompt optimization. Recently, a method termed the Autonomous Prompt Engineering Toolbox (APET) has incorporated various prompt design strategies into the prompt optimization process. In APET, the LLM is needed to implicitly select and apply the appropriate strategies because prompt design strategies can have negative effects. This implicit selection may be suboptimal due to the limited optimization capabilities of LLMs. This paper introduces Optimizing Prompts with sTrategy Selection (OPTS), which implements explicit selection mechanisms for prompt design. We propose three mechanisms, including a Thompson sampling-based approach, and integrate them into EvoPrompt, a well-known prompt optimizer. Experiments optimizing prompts for two LLMs, Llama-3-8B-Instruct and GPT-4o mini, were conducted using BIG-Bench Hard. Our results show that the selection of prompt design strategies improves the performance of EvoPrompt, and the Thompson sampling-based mechanism achieves the best overall results. Our experimental code is provided at https://github.com/shiralab/OPTS.

This study presents ***BanStereoSet***, a dataset designed to evaluate stereotypical social biases in multilingual LLMs for the Bangla language. In an effort to extend the focus of bias research beyond English-centric datasets, we have localized the content from the StereoSet, IndiBias, and kamruzzaman-etal’s datasets, producing a resource tailored to capture biases prevalent within the Bangla-speaking community. Our BanStereoSet dataset consists of 1,194 sentences spanning 9 categories of bias: race, profession, gender, ageism, beauty, beauty in profession, region, caste, and religion. This dataset not only serves as a crucial tool for measuring bias in multilingual LLMs but also facilitates the exploration of stereotypical bias across different social categories, potentially guiding the development of more equitable language technologies in *Bangladeshi* contexts. Our analysis of several language models using this dataset indicates significant biases, reinforcing the necessity for culturally and linguistically adapted datasets to develop more equitable language technologies.

Post-training of Large Language Models often involves a pipeline of Supervised Finetuning (SFT) followed by Preference Finetuning (PFT) using methods like Direct Preference Optimization. Both stages require annotated data that are very different in structure and costs. We study how to optimally allocate a fixed training data budget between the two stages, through extensive experiments spanning four diverse tasks, multiple model sizes and various data annotation costs. Our findings reveal that just SFT on the base model dominates performance in low-data regimes (<1,000 annotated examples). With larger data-budgets, we observe that a combination of SFT and PFT, often with increasing portions allocated towards preference data yields optimal performance. However, completely eliminating SFT and running PFT directly on the base model yields suboptimal performance, described as the cold start problem on tasks like mathematics. We observe that this is due to the distribution shift arising from using DPO directly on the base model to elicit step-by-step reasoning. This limitation can be effectively addressed by allocating even a small portion (<10%) of the budget to SFT first, resulting in performance improvements of 15-20% on analytical benchmarks like GSM8k. These results provide actionable insights for researchers and practitioners optimizing model development under budget constraints, where high-quality data curation often represents a significant portion of the total costs of model development.

Sampling-based decoding strategies have been widely adopted for Large Language Models (LLMs) in numerous applications, targeting a balance between diversity and quality via temperature tuning and tail truncation. Considering the strong dependency of the candidate next tokens on different prefixes, recent studies propose to adaptively truncate the tail of LLMs’ predicted distribution. Although improved results have been reported with these methods on open-ended text generation tasks, the results are highly dependent on the curated parameters and the limited exemplar text. In this paper, we propose a systematic way to estimate the intrinsic capacity of a truncation sampling method by considering the trade-off between diversity and risk at each decoding step, based on our collected prefix tree which preserves the context of a full sentence. Our work offers a comprehensive comparison of existing truncation sampling methods and serves as a practical user guideline for their parameter selection. Our code is available at https://anonymous.4open.science/r/Truncation-Sampling-Evaluation-251F.

Current backdoor attack defenders in Natural Language Processing (NLP) typically involve data reduction or model pruning, risking losing crucial information. To address this challenge, we introduce a novel backdoor defender, i.e., BadWindtunnel, in which we build a high-noise simulated training environment, similar to the wind tunnel, which allows precise control over training conditions to model the backdoor learning behavior without affecting the final model. We also use the confidence variance as a learning behavior quantification metric in the simulated training, which is based on the characteristics of backdoor-poisoned data (shorted in poisoned data): higher learnability and robustness. In addition, we propose a two-step strategy to further model poisoned data, including target label identification and poisoned data revealing. Extensive experiments demonstrate BadWindtunnel’s superiority, with a 21% higher average reduction in attack success rate than the second-best defender.

With Large Language Models (LLMs) becoming increasingly multilingual, effective knowledge editing (KE) needs to propagate edits across languages. Evaluation of the existing methods for cross-lingual knowledge editing (CKE) is limited both w.r.t. edit effectiveness: benchmarks do not account for entity aliases and use faulty entity translations; as well as robustness: existing work fails to report on downstream generation and task-solving abilities of LLMs after editing. In this work, we aim to (i) maximize the effectiveness of CKE while at the same time (ii) minimizing the extent of downstream model collapse due to the edits. To accurately measure the effectiveness of CKE methods, we introduce BabelEdits, a new CKE benchmark covering 60 languages that combines high-quality multilingual synsets from BabelNet with marker-based translation to ensure entity translation quality. Unlike existing CKE benchmarks, BabelEdits accounts for the rich variety of entity aliases within and across languages. We then propose BabelReFT, a modular CKE approach based on representation fine-tuning (ReFT) which learns entity-scope ReFT modules, applying them to all multilingual aliases at inference. Our experimental results show that not only is BabelReFT more effective in CKE than state-of-the-art methods, but, owing to its modular design, much more robust against downstream model collapse when subjected to many sequential edits.

People worldwide use language in subtle and complex ways to express emotions. Although emotion recognition–an umbrella term for several NLP tasks–impacts various applications within NLP and beyond, most work in this area has focused on high-resource languages. This has led to significant disparities in research efforts and proposed solutions, particularly for under-resourced languages, which often lack high-quality annotated datasets.In this paper, we present BRIGHTER–a collection of multi-labeled, emotion-annotated datasets in 28 different languages and across several domains. BRIGHTER primarily covers low-resource languages from Africa, Asia, Eastern Europe, and Latin America, with instances labeled by fluent speakers. We highlight the challenges related to the data collection and annotation processes, and then report experimental results for monolingual and crosslingual multi-label emotion identification, as well as emotion intensity recognition. We analyse the variability in performance across languages and text domains, both with and without the use of LLMs, and show that the BRIGHTER datasets represent a meaningful step towards addressing the gap in text-based emotion recognition.

A large part of human communication relies on nonverbal cues such as facial expressions, eye contact, and body language. Unlike language or sign language, such nonverbal communication lacks formal rules, requiring complex reasoning based on commonsense understanding.Enabling current Video Large Language Models (VideoLLMs) to accurately interpret body language is a crucial challenge, as human unconscious actions can easily cause the model to misinterpret their intent.To address this, we propose a dataset, BQA, a body language question answering dataset, to validate whether the model can correctly interpret emotions from short clips of body language comprising 26 emotion labels of videos of body language.We evaluated various VideoLLMs on the BQA with and without Multimodal Chain of Thought (CoT) and revealed that understanding body language is challenging, and our analyses of the wrong answers by VideoLLMs show that certain VideoLLMs made largely biased answers depending on the age group and ethnicity of the individuals. We also found consistent error patterns in VideoLLMs.

LLMs exhibit advanced reasoning capabilities, offering the potential to transform natural language questions into mathematical models. However, existing open-source datasets in operations research domain lack detailed annotations of the modeling process, such as variable definitions, focusing solely on objective values, which hinders reinforcement learning applications. To address this, we release the StructuredOR dataset, annotated with comprehensive labels that capture the complete mathematical modeling process. We further propose BPP-Search, an algorithm that integrates reinforcement learning into a tree-of-thought structure using Beam search, a Process reward model, and a pairwise Preference algorithm. This approach enables efficient exploration of tree structures, avoiding exhaustive search while improving accuracy. Extensive experiments on StructuredOR, NL4OPT, and MAMO-ComplexLP datasets show that BPP-Search significantly outperforms state-of-the-art methods. In tree-based reasoning, BPP-Search excels in accuracy and efficiency, enabling faster retrieval of correct solutions. The StructuredOR dataset is available on Huggingface https://huggingface.co/datasets/LLM4OR/StructuredOR and GitHub https://github.com/LLM4OR/StructuredOR.

This paper poses two critical issues in evaluating base models (without post-training): (1) Unstable evaluation during training: in the early stages of pre-training, the models lack the capability to answer questions as required, leading to unstable evaluation results. This instability makes it difficult to provide solid conclusions to guide the training, especially for key experiments such as data ablation and scaling law. (2) Inconsistency between base and instruct models: base models generally exhibit poorer evaluation performance compared to corresponding instruct models. This gap poses a challenge for assessing whether a base model with better evaluation can truly lead to a better instruct model. To address these issues, we propose **B**ase model **O**riented **S**ystematic **E**valuation (**BOSE**), a method specifically designed to optimize the evaluation of base models. Specifically, BOSE introduces two key innovations: In-Context Light-instruction Prompt (**ICLiP**) for open-ended tasks and **Blank-ppl** for multi-choice tasks with candidate options, which transforms the standard perplexity (ppl) metric into a fill-in-the-blank format to mitigate early-stage evaluation fluctuations. Furthermore, we are the first to propose Kendall’s rank correlation to quantitatively measure the evaluation stability and consistency. Experimental results demonstrate that BOSE significantly enhances both the stability of evaluations during pre-training and the consistency between base and instruct models, thereby providing more reliable guidance for the LLMs’ training.

Recent advances in large language models (LLMs) have enabled social simulation through multi-agent systems. Prior efforts focus on agent societies created from scratch, assigning agents with newly defined personas. However, simulating established fictional worlds and characters remain largely underexplored, despite its significant practical value. In this paper, we introduce BookWorld, a comprehensive system for constructing and simulating book-based multi-agent societies. BookWorld’s design covers comprehensive real-world intricacies, including diverse and dynamic characters, fictional worldviews, geographical constraints and changes, e.t.c. BookWorld enables diverse applications including story generation, interactive games and social simulation, offering novel ways to extend and explore beloved fictional works. Through extensive experiments, we demonstrate that BookWorld generates creative, high-quality stories while maintaining fidelity to the source books, surpassing previous methods with a win rate of 75.36%. The code and demo of this paper can be found at the project page: https://bookworld2025.github.io/.

Coreference Resolution systems are typically evaluated on benchmarks containing small- to medium-scale documents.When it comes to evaluating long texts, however, existing benchmarks, such as LitBank, remain limited in length and do not adequately assess system capabilities at the book scale, i.e., when co-referring mentions span hundreds of thousands of tokens.To fill this gap, we first put forward a novel automatic pipeline that produces high-quality Coreference Resolution annotations on full narrative texts. Then, we adopt this pipeline to create the first book-scale coreference benchmark, BOOKCOREF, with an average document length of more than 200,000 tokens.We carry out a series of experiments showing the robustness of our automatic procedure and demonstrating the value of our resource, which enables current long-document coreference systems to gain up to +20 CoNLL-F1 points when evaluated on full books.Moreover, we report on the new challenges introduced by this unprecedented book-scale setting, highlighting that current models fail to deliver the same performance they achieve on smaller documents.We release our data and code to encourage research and development of new book-scale Coreference Resolution systems at https://github.com/sapienzanlp/bookcoref.

This paper introduces BMIKE-53, a comprehensive benchmark for cross-lingual in-context knowledge editing (IKE), spanning 53 languages and three KE datasets: zsRE, CounterFact, and WikiFactDiff. Cross-lingual KE, which requires knowledge edited in one language to generalize across diverse languages while preserving unrelated knowledge, remains underexplored. To address this, we systematically evaluate IKE under zero-shot, one-shot, and few-shot setups, including tailored metric-specific demonstrations. Our findings reveal that model scale and demonstration alignment critically govern cross-lingual editing efficacy, with larger models and tailored demonstrations significantly improving performance. Linguistic properties, particularly script type, strongly influence outcomes, with non-Latin languages underperforming due to issues like language confusion.

Recently, generative pre-trained models have made significant strides, particularly highlighted by the release of ChatGPT and GPT-4, which exhibit superior cross-domain capabilities. However, these models still face challenges on constrained writing tasks like poem generation under open-domain titles via direct generation.In response to this challenge, we introduce Block Inverse Prompting (BIPro) constrained generation framework. BIPro leverages two block inverse prompting methods, revise and rewrite. This inference scaling approach mimics the process of human text writing using block generative models. It significantly improves the zero-shot generation quality on the constrained generation task of open-domain traditional-form Chinese poem generation. Based on a less powerful block generative model GLM-10B-Chinese, poems composed via BIPro without priming or additional training outperform both much larger direct generative systems like GPT-4 or GLM-4 and domain-specific systems such as Yusheng, Shisanbai, or Baidu Poetry Helper in human evaluation by proficient poets. BIPro considerably narrows the gap between AI-generated works and short-listed human literary arts in another human evaluation, unveiling the promising potential of inference scaling in improving the quality of constrained generation. It is open-sourced and available as an agent in chatglm app.

In this work, we tackle the challenge of embedding realistic human personality traits into LLMs. Previous approaches have primarily focused on prompt-based methods that describe the behavior associated with the desired personality traits, suffering from realism and validity issues. To address these limitations, we introduce BIG5-CHAT, a large-scale dataset containing 100,000 dialogues designed to ground models in how humans express their personality in text. Leveraging this dataset, we explore Supervised Fine-Tuning and Direct Preference Optimization as training-based methods to align LLMs more naturally with human personality patterns. Our methods outperform prompting on personality assessments such as BFI and IPIP-NEO, with trait correlations more closely matching human data. Furthermore, our experiments reveal that models trained to exhibit higher conscientiousness, higher agreeableness, lower extraversion, and lower neuroticism display better performance on reasoning tasks, aligning with psychological findings on how these traits impact human cognitive performance. To our knowledge, this work is the first comprehensive study to demonstrate how training-based methods can shape LLM personalities through learning from real human behaviors.

Current benchmarks for large language model (LLM) reasoning predominantly focus on mathematical and coding abilities, leaving a gap in evaluating broader reasoning proficiencies. One particular exception is the BIG-Bench dataset, which has served as a crucial benchmark for evaluating the general reasoning capabilities of LLMs, thanks to its diverse set of challenging tasks that allowed for a comprehensive assessment of general reasoning across various skills within a unified framework. However, recent advances in LLMs have led to saturation on BIG-Bench, and its harder version BIG-Bench Hard (BBH). State-of-the-art models achieve near-perfect scores on many tasks in BBH, thus diminishing its utility. To address this limitation, we introduce BIG-Bench Extra Hard (BBEH), a new benchmark designed to push the boundaries of LLM reasoning evaluation. BBEH replaces each task in BBH with a novel task that probes a similar reasoning capability but exhibits significantly increased difficulty. We evaluate various general-purpose and reasoning-specialized models on BBEH and observe an accuracy of 23.9% for the best general-purpose model and 54.2% for the best reasoning-specialized model, indicating substantial room for improvement and highlighting the ongoing challenge of achieving robust general reasoning in LLMs. We release BBEH publicly at: https://github.com/google-deepmind/bbeh.