NAACL 2025 Papers

The lottery ticket hypothesis posits the existence of “winning tickets” within a randomly initialized neural network. Do winning tickets exist for LLMs in fine-tuning scenarios? How can we find such winning tickets? In this paper, we propose KS-Lottery, a method to identify a small subset of LLM parameters highly effective in multilingual fine-tuning. Our key idea is to use Kolmogorov-Smirnov Test to analyze the distribution shift of parameters before and after fine-tuning. We further theoretically prove that KS-Lottery can find the certified winning tickets in the embedding layer, fine-tuning on the found parameters is guaranteed to perform as well as full fine-tuning. Comparing KS-Lottery with other tuning algorithms on translation tasks, the experimental results show that KS-Lottery finds a much smaller set of parameters for fine-tuning while achieving the comparable performance as full fine-tuning LLM. Surprisingly, we find that fine-tuning 18 tokens’ embedding of LLaMA suffices to reach the fine-tuning translation performance .

We propose KMMLU, a Korean benchmark with 35,030 expert-level multiple-choice questions across 45 subjects ranging from humanities to STEM. While prior Korean evaluation tools heavily rely on translated versions of existing English benchmarks, KMMLU is collected from original Korean exams, thereby capturing linguistic and cultural aspects of the Korean language. Recent models struggle to show performance over 60%, significantly below the pass mark of the source exams (80%), highlighting the room for improvement. Notably, one-fifth of the questions in KMMLU require knowledge of Korean culture for accurate resolution. KMMLU thus provides a more accurate reflection of human preferences compared to translated versions of MMLU and offers deeper insights into LLMs’ shortcomings in Korean knowledge. The dataset and codes are made publicly available for future research.

The increasing demand for mental health services has led to the rise of AI-driven mental health chatbots, though challenges related to privacy, data collection, and expertise persist. Motivational Interviewing (MI) is gaining attention as a theoretical basis for boosting expertise in the development of these chatbots. However, existing datasets are showing limitations for training chatbots, leading to a substantial demand for publicly available resources in the field of MI and psychotherapy. These challenges are even more pronounced in non-English languages, where they receive less attention. In this paper, we propose a novel framework that simulates MI sessions enriched with the expertise of professional therapists. We train an MI forecaster model that mimics the behavioral choices of professional therapists and employ Large Language Models (LLMs) to generate utterances through prompt engineering. Then, we present KMI, the first synthetic dataset theoretically grounded in MI, containing 1,000 high-quality Korean Motivational Interviewing dialogues. Through an extensive expert evaluation of the generated dataset and the dialogue model trained on it, we demonstrate the quality, expertise, and practicality of KMI. We also introduce novel metrics derived from MI theory in order to evaluate dialogues from the perspective of MI.

Strategic reasoning is a complex yet essential capability for intelligent agents. It requires Large Language Model (LLM) agents to adapt their strategies dynamically in multi-agent environments. Unlike static reasoning tasks, success in these contexts depends on anticipating other agents’ beliefs and actions while continuously adjusting strategies to achieve individual goals. LLMs and LLM agents often struggle with strategic reasoning due to the absence of a reasoning framework that enables them to dynamically infer others’ perspectives and adapt to changing environments. Inspired by the Level-K framework from game theory and behavioral economics, which extends reasoning from simple reactions to structured strategic depth, we propose a novel framework: “K-Level Reasoning with Large Language Models (K-R).” This framework employs recursive mechanisms to enable LLMs to achieve varying levels of strategic depth, allowing agents to form higher order beliefs—beliefs about others’ beliefs. We validate this framework through rigorous testing on four testbeds: two classical game theory problems and two social intelligence tasks. The results demonstrate the advantages of K-R in strategic reasoning. Our work presents the first recursive implementation of strategic depth in large language models (LLMs). It establishes a foundation for future research into theory of mind and strategic reasoning in LLMs.

Document-level multi-event extraction aims to identify a list of event types and corresponding arguments from the document. However, most of the current methods neglect the fine-grained difference among events in multi-event documents, which leads to event confusion and missing. This is also one of the reasons why the recall and F1-score of multi-event recognition are lower compared to single-event recognition. In this paper, we propose an event-specific probe-based method to sniff multiple events by querying each corresponding argument library, which uses a novel probe-label alignment method for differential optimization. In addition, the role contrastive loss and probe consistent loss are designed to fine-tune the fine-grained role differences and probe differences in each event. The experimental results on two general datasets show that our method outperforms the state-of-the-art method in the F1-score, especially in the recall of multi-events.

Large language models (LLMs) have seen rapid development in recent years, revolutionizing various applications and significantly enhancing convenience and productivity. However, alongside their impressive capabilities, ethical concerns and new types of attacks, such as jailbreaking, have emerged. While most prompting techniques focus on optimizing adversarial inputs for individual cases, resulting in higher computational costs when dealing with large datasets. Less research has addressed the more general setting of training a universal attacker that can transfer to unseen tasks. In this paper, we introduce JUMP, a prompt-based method designed to jailbreak LLMs using universal multi-prompts. We also adapt our approach for defense, which we term DUMP. Experimental results demonstrate that our method for optimizing universal multi-prompts outperforms existing techniques.

Text-to-image (T2I) models can be maliciously used to generate harmful content such as sexually explicit, unfaithful, and misleading or Not-Safe-for-Work (NSFW) images. Previous attacks largely depend on the availability of the diffusion model or involve a lengthy optimization process. In this work, we investigate a more practical and universal attack that does not require the presence of a target model and demonstrate that the high-dimensional text embedding space inherently contains NSFW concepts that can be exploited to generate harmful images. We present the Jailbreaking Prompt Attack (JPA). JPA first searches for the target malicious concepts in the text embedding space using a group of antonyms generated by ChatGPT. Subsequently, a prefix prompt is optimized in the discrete vocabulary space to align malicious concepts semantically in the text embedding space.We further introduce a soft assignment with gradient masking technique that allows us to perform gradient ascent in the discrete vocabulary space.We perform extensive experiments with open-sourced T2I models, e.g. stable-diffusion-v1-4 and closed-sourced online services, e.g. DALL·E 2 and Midjourney with black-box safety checkers. Results show that (1) JPA bypasses both text and image safety checkers, (2) while preserving high semantic alignment with the target prompt. (3) JPA demonstrates a much faster speed than previous methods and can be executed in a fully automated manner. These merits render it a valuable tool for robustness evaluation in future text-to-image generation research.

Science journalism reports current scientific discoveries to non-specialists, aiming to enable public comprehension of the state of the art. This task is challenging as the audience often lacks specific knowledge about the presented research. We propose JRE-L, a framework that integrates three LLMs mimicking the writing-reading-feedback-revision loop. In JRE-L, one LLM acts as the journalist, another LLM as the general public reader, and the third LLM as an editor. The journalist’s writing is iteratively refined by feedback from the reader and suggestions from the editor. Our experiments demonstrate that by leveraging the collaboration of two 7B and one 1.8B open-source LLMs, we can generate articles that are more accessible than those generated by existing methods, including prompting single advanced models such as GPT-4 and other LLM-collaboration strategies. Our code is publicly available at github.com/Zzoay/JRE-L.

Recent advancements in instruction fine-tuning, alignment methods such as reinforcement learning from human feedback (RLHF), and optimization techniques like direct preference optimization (DPO), have significantly enhanced the adaptability of large language models (LLMs) to user preferences. However, despite these innovations, many LLMs continue to exhibit biases toward Western, Anglo-centric, or American cultures, with performance on English data consistently surpassing that of other languages. This reveals a persistent cultural gap in LLMs, which complicates their ability to accurately process culturally rich and diverse figurative language, such as proverbs. To address this, we introduce *Jawaher*, a benchmark designed to assess LLMs’ capacity to comprehend and interpret Arabic proverbs. *Jawaher* includes proverbs from various Arabic dialects, along with idiomatic translations and explanations. Through extensive evaluations of both open- and closed-source models, we find that while LLMs can generate idiomatically accurate translations, they struggle with producing culturally nuanced and contextually relevant explanations. These findings highlight the need for ongoing model refinement and dataset expansion to bridge the cultural gap in figurative language processing.

Recent research has shown that Large Language Models (LLMs) are vulnerable to automated jailbreak attacks, where adversarial suffixes crafted by algorithms appended to harmful queries bypass safety alignment and trigger unintended responses. Current methods for generating these suffixes are computationally expensive and have low Attack Success Rates (ASR), especially against well-aligned models like Llama2 and Llama3. To overcome these limitations, we introduce **ADV-LLM**, an iterative self-tuning process that crafts adversarial LLMs with enhanced jailbreak ability. Our framework significantly reduces the computational cost of generating adversarial suffixes while achieving nearly 100% ASR on various open-source LLMs. Moreover, it exhibits strong attack transferability to closed-source models, achieving 99% ASR on GPT-3.5 and 49% ASR on GPT-4, despite being optimized solely on Llama3. Beyond improving jailbreak ability, ADV-LLM provides valuable insights for future safety alignment research through its ability to generate large datasets for studying LLM safety.

Moderation is essential for maintaining and improving the quality of online discussions. This involves: (1) countering negativity, e.g. hate speech and toxicity, and (2) promoting positive discourse, e.g. broadening the discussion to involve other users and perspectives. While significant efforts have focused on addressing negativity, driven by an urgency to address such issues, this left moderation promoting positive discourse (henceforth PositiveModeration) under-studied. With the recent advancements in LLMs, Positive Moderation can potentially be scaled to vast conversations, fostering more thoughtful discussions and bridging the increasing divide in online interactions.We advance the understanding of Positive Moderation by annotating a dataset on 13 moderation properties, e.g. neutrality, clarity and curiosity. We extract instructions from professional moderation guidelines and use them to prompt LLaMA to generate such moderation. This is followed by extensive evaluation showing that (1) annotators rate generated higher than professional moderation, but still slightly prefer professional moderation in pairwise comparison, and (2) LLMs can be used to estimate human evaluation as an efficient alternative.

Adversarial datasets should validate AI robustness by providing samples on which humans perform well, but models do not. However, as models evolve, datasets can become obsolete. Measuring whether a dataset remains adversarial is hindered by the lack of a standardized metric for measuring adversarialness. We propose ADVSCORE, a human-grounded evaluation metric that assesses a dataset’s adversarialness by capturing models’ and humans’ varying abilities, while also identifying poor examples. We then use ADVSCORE to motivate a new dataset creation pipeline for realistic and high-quality adversarial samples, enabling us to collect an adversarial question answering (QA) dataset, ADVQA. We apply ADVSCORE using 9,347 human responses and ten language models’ predictions to track model improvement over five years (2020–2024). ADVSCORE thus provides guidance for achieving robustness comparable with human capabilities. Furthermore, it helps determine to what extent adversarial datasets continue to pose challenges, ensuring that, rather than reflecting outdated or overly artificial difficulties, they effectively test model capabilities.

Conceptual combination is a cognitive process that merges basic concepts, enabling the creation of complex expressions. During this process, the properties of combination (e.g., the whiteness of a peeled apple) can be inherited from basic concepts, newly emerge, or be canceled. However, previous studies have evaluated a limited set of properties and have not examined the generative process.To address this gap, we introduce the Conceptual Combination with Property Type dataset (CCPT), which consists of 12.3K annotated triplets of noun phrases, properties, and property types. Using CCPT, we establish three types of tasks to evaluate LLMs for conceptual combination thoroughly.Our key findings are threefold:(1) Our automatic metric grading property emergence and cancellation closely corresponds with human judgments.(2) LLMs, including OpenAI’s o1, struggle to generate noun phrases which possess given emergent properties.(3) Our proposed method, inspired by cognitive psychology model that explains how relationships between concepts are formed, improves performances in all generative tasks.The dataset and experimental code are available at https://github.com/seokwon99/CCPT.git.

The rapid advancement of Large Language Models (LLMs) has led to the development of benchmarks that consider temporal dynamics, however, there remains a gap in understanding how well these models can generalize across temporal contexts due to the inherent dynamic nature of language and information. This paper introduces the concept of temporal generalization in LLMs, including bias in past and future generalizations. Then we introduce FreshBench, a new evaluation framework that employs fresh text and event prediction for assessing LLMs’ temporal adaptability, ensuring the evaluation process free from data leakage and subjective bias. The experiment shows significant temporal biases and a decline in performance over time.

Large language models (LLMs) have demonstrated multilingual capabilities, yet they are mostly English-centric due to the imbalanced training corpora. While prior works have leveraged this bias to enhance multilingual performance through translation, they have been largely limited to natural language processing (NLP) tasks. In this work, we extend the evaluation to real-world user queries and non-English-centric LLMs, offering a broader examination of multilingual performance. Our key contribution lies in demonstrating that while translation into English can boost the performance of English-centric LLMs on NLP tasks, it is not universally optimal. For culture-related tasks that need deep language understanding, prompting in the native language proves more effective as it better captures the nuances of culture and language. Our experiments expose varied behaviors across LLMs and tasks in the multilingual context, underscoring the need for a more comprehensive approach to multilingual evaluation. Therefore, we call for greater efforts in developing and evaluating LLMs that go beyond English-centric paradigms.

Recent advances in Retrieval-Augmented Generation (RAG) systems have popularized semantic chunking, which aims to improve retrieval performance by dividing documents into semantically coherent segments. Despite its growing adoption, the actual benefits over simpler fixed-size chunking, where documents are split into consecutive, fixed-size segments, remain unclear. This study systematically evaluates the effectiveness of semantic chunking using three common retrieval-related tasks: document retrieval, evidence retrieval, and retrieval-based answer generation. The results show that the computational costs associated with semantic chunking are not justified by consistent performance gains. These findings challenge the previous assumptions about semantic chunking and highlight the need for more efficient chunking strategies in RAG systems.

Endangered languages, such as Navajo—the most widely spoken Native American language—are significantly underrepresented in contemporary language technologies, exacerbating the challenges of their preservation and revitalization. This study evaluates Google’s Language Identification (LangID) tool, which does not currently support any Native American languages. To address this, we introduce a random forest classifier trained on Navajo and twenty erroneously suggested languages by LangID. Despite its simplicity, the classifier achieves near-perfect accuracy (97-100%). Additionally, the model demonstrates robustness across other Athabaskan languages—a family of Native American languages spoken primarily in Alaska, the Pacific Northwest, and parts of the Southwestern United States—suggesting its potential for broader application. Our findings underscore the pressing need for NLP systems that prioritize linguistic diversity and adaptability over centralized, one-size-fits-all solutions, especially in supporting underrepresented languages in a multicultural world. This work directly contributes to ongoing efforts to address cultural biases in language models and advocates for the development of culturally localized NLP tools that serve diverse linguistic communities.

Large language models (LLMs) have shown the emergent capability of in-context learning (ICL). One line of research has claimed that ICL is functionally equivalent to gradient descent, a type of error-driven learning mechanism. In this paper, we introduce a new way of diagnosing whether ICL is functionally performing error-driven learning. Our approach is based on the inverse frequency effect (IFE)—a phenomenon in which an agent’s behavior is influenced to a greater degree when presented with improbable examples as compared to more likely ones. The IFE has previously been identified in psycholinguistics where humans exhibit the IFE in the context of structural priming (the tendency for people to produce sentence structures they have encountered recently). In that context, the IFE has been used as evidence that human structural priming must involve error-driven learning mechanisms. In our experiments, we simulated structural priming with ICL and found that LLMs indeed display the IFE, with the effect being stronger in larger models. We conclude that at least in the case we studied, ICL is indeed a type of error-driven learning, supporting the hypothesis that an error signal is implicitly computed in the forward pass during ICL. Our results suggest that both humans and LLMs make use of error-driven processing mechanisms in on-line processing.

Despite the widespread adoption of Large language models (LLMs), their remarkable capabilities remain limited to a few high-resource languages. Additionally, many low-resource languages (e.g. African languages) are often evaluated only on basic text classification tasks due to the lack of appropriate or comprehensive benchmarks outside of high-resource languages. In this paper, we introduce IrokoBench—a human-translated benchmark dataset for 17 typologically-diverse low-resource African languages covering three tasks: natural language inference(AfriXNLI), mathematical reasoning(AfriMGSM), and multi-choice knowledge-based QA(AfriMMLU). We use IrokoBench to evaluate zero-shot, few-shot, and translate-test settings(where test sets are translated into English) across 10 open and four proprietary LLMs. Our evaluation reveals a significant performance gap between high-resource languages (such as English and French) and low-resource African languages. We observe a significant performance gap between open and proprietary models, with the highest performing open model, Gemma 2 27B only at 63% of the best-performing proprietary model GPT-4o performance. Machine translating the test set to English before evaluation helped to close the gap for larger models that are English-centric, like Gemma 2 27B and LLaMa 3.1 70B. These findings suggest that more efforts are needed to develop and adapt LLMs for African languages.