NAACL 2025 Papers

Direct Preference Optimization (DPO) has become a widely used training method for the instruction fine-tuning of large language models (LLMs). In this work, we explore an under-investigated aspect of DPO – its dependency on the reference model or policy. Such reference policies, typically instantiated as the model to be further fine-tuned, are important since they can impose an upper limit on DPO’s effectiveness. Therefore, we address three related research questions in this work. First, we explore the optimal strength of the KL divergence constraint in DPO, which penalizes deviations from the reference policy, and find that DPO is sensitive to this strength. Next, we examine the necessity of the KL-constraint from the reference policies in DPO by providing both theoretical and empirical comparisons between DPO and related learning objectives, demonstrating DPO’s superiority in this controlled setting. Additionally, we investigate whether DPO benefits from stronger reference policies, finding that a stronger reference policy can lead to improved performance, but only when it is similar to the model being fine-tuned. Our findings highlight the confounding role of reference policies in DPO and offer insights for best practices, while also identifying open research questions for future studies.

While LLMs have exhibited strong performance on various NLP tasks, it is noteworthy that most of these tasks rely on utilizing the vast amount of knowledge encoded in LLMs’ parameters, rather than solving new problems without prior knowledge. In cognitive research, the latter ability is referred to as fluid intelligence, which is considered to be critical for assessing human intelligence. Recent research on fluid intelligence assessments has highlighted significant deficiencies in LLMs’ abilities. In this paper, we analyze the challenges LLMs face in demonstrating fluid intelligence through controlled experiments, using the most representative ARC task as an example. Our study revealed three major limitations in existing LLMs: limited ability for skill composition, unfamiliarity with abstract input formats, and the intrinsic deficiency of left-to-right decoding. Our data and code will be publicly released, and the data is also attached in the submission.

This paper conducts a longitudinal study over eleven months to address the limitations of prior research on the Open Ko-LLM Leaderboard, which have relied on empirical studies with restricted observation periods of only five months. By extending the analysis duration, we aim to provide a more comprehensive understanding of the progression in developing Korean large language models (LLMs). Our study is guided by three primary research questions: (1) What are the specific challenges in improving LLM performance across diverse tasks on the Open Ko-LLM Leaderboard over time? (2) How does model size impact task performance correlations across various benchmarks? (3) How have the patterns in leaderboard rankings shifted over time on the Open Ko-LLM Leaderboard?. By analyzing 1,769 models over this period, our research offers a comprehensive examination of the ongoing advancements in LLMs and the evolving nature of evaluation frameworks.

Large Vision-Language Models (VLMs) have demonstrated strong capabilities in tasks requiring a fine-grained understanding of literal meaning in images and text, such as visual question-answering or visual entailment. However, there has been little exploration of the capabilities of these models when presented with images and captions containing figurative meaning, such as metaphors or humor. To close this gap, we propose a new task framing the figurative meaning understanding problem as an explainable visual entailment task, where the model has to predict whether the image (premise) entails a caption (hypothesis) and justify the predicted label with a textual explanation. The figurative phenomena can be present in the image, in the caption, or both. Using a human-AI collaboration approach, we build the accompanying expert-verified dataset V-FLUTE, containing 6,027 image, caption, label, explanation instances spanning five diverse figurative phenomena: metaphors, similes, idioms, sarcasm, and humor. Through automatic evaluation, we find that VLMs struggle to generalize from literal to figurative meaning, particularly when it is present in images. Further, we identify common types of errors in VLM reasoning (hallucination and incomplete or unsound reasoning) across classes of models via human evaluation.

The widespread use of social media has led to a surge in popularity for automated methods of analyzing public opinion. Supervised methods are adept at text categorization, yet the dynamic nature of social media discussions poses a continual challenge for these techniques due to the constant shifting of the focus. On the other hand, traditional unsupervised methods for extracting themes from public discourse, such as topic modeling, often reveal overarching patterns that might not capture specific nuances. Consequently, a significant portion of research into social media discourse still depends on labor-intensive manual coding techniques and a human-in-the-loop approach, which are both time-consuming and costly. In this work, we study the problem of discovering arguments associated with a specific theme. We propose a generic **LLMs-in-the-Loop** strategy that leverages the advanced capabilities of Large Language Models (LLMs) to extract latent arguments from social media messaging. To demonstrate our approach, we apply our framework to contentious topics. We use two publicly available datasets: (1) the climate campaigns dataset of 14k Facebook ads with 25 themes and (2) the COVID-19 vaccine campaigns dataset of 9k Facebook ads with 14 themes. Additionally, we design a downstream task as stance prediction by leveraging talking points in climate debates. Furthermore, we analyze demographic targeting and the adaptation of messaging based on real-world events.

With the advent of Large Language Models (LLMs) possessing increasingly impressive capabilities, a number of Large Vision-Language Models (LVLMs) have been proposed to augment LLMs with visual inputs. Such models condition generated text on both an input image and a text prompt, enabling a variety of use cases such as visual question answering and multimodal chat. While prior studies have examined the social biases contained in text generated by LLMs, this topic has been relatively unexplored in LVLMs. Examining social biases in LVLMs is particularly challenging due to the confounding contributions of bias induced by information contained across the text and visual modalities. To address this challenging problem, we conduct a large-scale study of text generated by different LVLMs under counterfactual changes to input images, producing over 57 million responses from popular models. Our multi-dimensional bias evaluation framework reveals that social attributes such as perceived race, gender, and physical characteristics depicted in images can significantly influence the generation of toxic content, competency-associated words, harmful stereotypes, and numerical ratings of individuals.

To facilitate healthcare delivery, language models (LMs) have significant potential for clinical prediction tasks using electronic health records (EHRs). However, in these high-stakes applications, unreliable decisions can result in significant costs due to compromised patient safety and ethical concerns, thus increasing the need for good uncertainty modelling of automated clinical predictions. To address this, we consider uncertainty quantification of LMs for EHR tasks in both white-box and black-box settings. We first quantify uncertainty in white-box models, where we have access to model parameters and output logits. We show that an effective reduction of model uncertainty can be achieved by using the proposed multi-tasking and ensemble methods in EHRs. Continuing with this idea, we extend our approach to black-box settings, including popular proprietary LMs such as GPT-4. We validate our framework using longitudinal clinical data from over 6,000 patients across ten clinical prediction tasks. Results show that ensembling methods and multi-task prediction prompts reduce uncertainty across different scenarios. These findings increase model transparency in white-box and black-box settings, thereby advancing reliable AI healthcare.

Open relation extraction (OpenRE) aims to identify relational facts within open-domain corpora without relying on predefined relation types. A significant limitation of current state-of-the-art OpenRE approaches is their inability to accurately self-assess their performance. Which is caused by the reliance on pseudo-labels, that treats all points within a cluster equally, regardless of their actual relative position according to the cluster center. This leads to models that are often overconfident in their incorrect predictions , significantly undermining their reliability. In this paper, we introduce an approach that addresses this challenge by effectively modeling a part of the epistemic uncertainty within OpenRE. Instead of using pseudo labels that mask uncertainty, our approach is built to train a classifier directly with the clustering distribution. Our experimental results across various datasets demonstrate that the suggested approach improves reliability of OpenRE by preventing overconfident errors. Furthermore we show that by improving the reliability of the predictions, UOREX operates more efficiently in a generative active learning context where an LLM is the oracle, doubling the performance gain compared to the state-of-the-art.

Large Language Models (LLMs) excel in many Natural Language Processing tasks but are outperformed by specialized tools for certain tasks. This raises the question: Can we reduce redundant LLM parameters when using these tools? Given the size and high training costs of LLMs, it is essential to efficiently forget specific knowledge without retraining. This paper introduces UNLEARN, a novel method that uses subspace techniques to selectively remove knowledge without access to the original training data, without retraining, and with minimal impact to other tasks. Our results show that UNLEARN significantly outperforms previous methods for forgetting targeted (unwanted) knowledge while also preserving related (wanted) knowledge. We also propose LEARN, a complementary approach for targeted knowledge addition, which achieves fine-tuning accuracy comparable to Low-Rank Adaptation (LoRA) without degrading related task performance.

Mitigating the retention of sensitive or private information in large language models is essential for enhancing privacy and safety. Existing unlearning methods, like Gradient Ascent and Negative Preference Optimization, directly tune models to remove unwanted information. However, these methods often become unstable because they fine-tune by maximizing loss, which is the opposite of traditional loss minimization in learning. This reversal creates instability, especially on larger datasets, as the model struggles to balance unlearning with maintaining language capacity, leading to over-unlearning. In this paper, we introduce UnDIAL (Unlearning via Self-Distillation on Adjusted Logits), a novel and robust unlearning method. Our approach leverages self-distillation to adjust logits and selectively reduce the influence of targeted tokens. This technique ensures smooth convergence and avoids catastrophic forgetting, even in challenging unlearning tasks with large datasets and sequential unlearning requests. Extensive experiments show that UnDIAL is the first direct tuning method to achieve both robustness in unlearning and scalability, while maintaining stable training dynamics and resilience to hyperparameter tuning.

We introduce UFO, a UI-Fcused agent designed to fulfill user requests tailored to Windows OS applications by observing and analyzing the GUI and control information of these applications. UFO utilizes a hierarchical dual-agent framework that decomposes user requests using a divide-and-conquer approach, enabling seamless navigation and addressing sub-tasks across multiple applications. It also incorporates a control interaction module tailored for Windows OS, which detects control elements effectively and allows for fully automated execution. As a result, UFO simplifies complex and time-consuming processes into tasks that can be completed with natural language commands.We conducted testing of UFO across 9 popular Windows applications, encompassing a variety of scenarios. The results derived from both quantitative metrics and real-case studies, underscore the superior effectiveness of UFOin fulfilling user requests. To the best of our knowledge, UFO stands as the first UI agent specifically tailored for task completion within the Windows OS.

Existing legal benchmarks focusing on knowledge and logic effectively evaluate LLMs on various tasks in legal domain. However, few have explored the practical application of LLMs by actual users. To further assess whether LLMs meet the specific needs of legal practitioners in real-world scenarios, we introduce UCL-Bench, a Chinese User-Centric Legal Benchmark, comprising 22 tasks across 5 distinct legal scenarios.To build the UCL-Bench, we conduct a user survey targeting legal professionals to understand their needs and challenges. Based on the survey results, we craft tasks, verified by legal professionals, and categorized them according to Bloom’s taxonomy. Each task in UCL-Bench mirrors real-world legal scenarios, and instead of relying on pre-defined answers, legal experts provide detailed answer guidance for each task, incorporating both “information” and “needs” elements to mimic the complexities of legal practice. With the guidance, we use GPT-4 as the user simulator and evaluator, enabling multi-turn dialogues as a answer guidance based evaluation framework. Our findings reveal that many recent open-source general models achieve the highest performance, suggesting that they are well-suited to address the needs of legal practitioners. However, these legal LLMs do not outperform ChatGPT, indicating a need for training strategies aligned with users’ needs. Furthermore, we find that the most effective models are able to address legal issues within fewer dialogue turns, highlighting the importance of concise and accurate responses in achieving high performance. The code and dataset are available at https://github.com/wittenberg11/UCL-bench.

This paper introduces the UCFE: User-Centric Financial Expertise benchmark, an innovative framework designed to evaluate the ability of large language models (LLMs) to handle complex real-world financial tasks. UCFE benchmark adopts a hybrid approach that combines human expert evaluations with dynamic, task-specific interactions to simulate the complexities of evolving financial scenarios. Firstly, we conducted a user study involving 804 participants, collecting their feedback on financial tasks. Secondly, based on this feedback, we created our dataset that encompasses a wide range of user intents and interactions. This dataset serves as the foundation for benchmarking 11 LLMs services using the LLM-as-Judge methodology. Our results show a significant alignment between benchmark scores and human preferences, with a Pearson correlation coefficient of 0.78, confirming the effectiveness of the UCFE dataset and our evaluation approach. UCFE benchmark not only reveals the potential of LLMs in the financial domain but also provides a robust framework for assessing their performance and user satisfaction.

Large Vision-Language Models (LVLMs) are susceptible to typographic attacks, which are misclassifications caused by an attack text that is added to an image. In this paper, we introduce a multi-image setting for studying typographic attacks, broadening the current emphasis of the literature on attacking individual images. Specifically, our focus is on attacking image sets without repeating the attack query. Such non-repeating attacks are stealthier, as they are more likely to evade a gatekeeper than attacks that repeat the same attack text. We introduce two attack strategies for the multi-image setting, leveraging the difficulty of the target image, the strength of the attack text, and text-image similarity. Our text-image similarity approach improves attack success rates by 21% over random, non-specific methods on the CLIP model using ImageNet while maintaining stealth in a multi-image scenario. An additional experiment demonstrates transferability, i.e., text-image similarity calculated using CLIP transfers when attacking InstructBLIP.

Humans have the ability to reason about geometric patterns in images and scenes from a young age. However, developing large multimodal models (LMMs) capable of similar reasoning remains a challenge, highlighting the need for robust evaluation methods to assess these capabilities. We introduce TurtleBench, a benchmark designed to evaluate LMMs’ capacity to interpret geometric patterns—given visual examples, textual instructions, or both—and generate precise code outputs. Inspired by turtle geometry, a notion used to teach children foundational coding and geometric concepts, TurtleBench features tasks with patterned shapes that have underlying algorithmic logic. Our evaluation reveals that leading LMMs struggle significantly with these tasks, with GPT-4V achieving only 19% accuracy on the simplest tasks and few-shot prompting only marginally improves their performance (<2%). TurtleBench highlights the gap between human and AI performance in intuitive and visual geometrical understanding, setting the stage for future research in this area and stands as one of the few benchmarks to evaluate the integration of visual understanding and code generation capabilities in LMMs, setting the stage for future research.

Can advanced multi-modal models effectively tackle complex web-based tasks? Such tasks are often found on crowdsourcing platforms, where crowdworkers engage in challenging micro-tasks within web-based environments.Building on this idea, we present TurkingBench, a benchmark consisting of tasks presented as web pages with textual instructions and multi-modal contexts. Unlike previous approaches that rely on artificially synthesized web pages, our benchmark uses natural HTML pages originally designed for crowdsourcing workers to perform various annotation tasks. Each task’s HTML instructions are instantiated with different values derived from crowdsourcing tasks, creating diverse instances. This benchmark includes 32.2K instances spread across 158 tasks.To support the evaluation of TurkingBench, we have developed a framework that links chatbot responses to actions on web pages (e.g., modifying a text box, selecting a radio button). We assess the performance of cutting-edge private and open-source models, including language-only and vision-language models (such as GPT4 and InternVL), on this benchmark. Our results show that while these models outperform random chance, there is still significant room for improvement. We hope that this benchmark will drive progress in the evaluation and development of web-based agents.

Jailbreaking large-language models (LLMs) involves testing their robustness against adversarial prompts and evaluating their ability to withstand prompt attacks that could elicit unauthorized or malicious responses. In this paper, we present TurboFuzzLLM, a mutation-based fuzzing technique for efficiently finding a collection of effective jailbreaking templates that, when combined with harmful questions, can lead a target LLM to produce harmful responses through black-box access via user prompts. We describe the limitations of directly applying existing template-based attacking techniques in practice, and present functional and efficiency-focused upgrades we added to mutation-based fuzzing to generate effective jailbreaking templates automatically. TurboFuzzLLM achieves ≥ 95% attack success rates (ASR) on public datasets for leading LLMs (including GPT-4o & GPT-4 Turbo), shows impressive generalizability to unseen harmful questions, and helps in improving model defenses to prompt attacks.

Language models are aligned to the collective voice of many, resulting in generic outputs that do not align with specific users’ styles. In this work, we present Trial-Error-Explain In-Context Learning (TICL), a tuning-free method that personalizes language models for text generation tasks with fewer than 10 examples per user. TICL iteratively expands an in-context learning prompt via a trial-error-explain process, adding model-generated negative samples and explanations that provide fine-grained guidance towards a specific user’s style. TICL achieves favorable win rates on pairwise comparisons with LLM-as-a-judge up to 91.5% against the previous state-of-the-art and outperforms competitive tuning-free baselines for personalized alignment tasks of writing emails, essays and news articles. Both lexical and qualitative analyses show that the negative samples and explanations enable language models to learn stylistic context more effectively and overcome the bias towards structural and formal phrases observed in their zero-shot outputs. By front-loading inference compute to create a user-specific in-context learning prompt that does not require extra generation steps at test time, presents a novel yet simple approach for personalized alignment.

Retrieval-augmented generation (RAG) systems enhance large language models by incorporating external knowledge, addressing issues like outdated internal knowledge and hallucination. However, their reliance on external knowledge bases makes them vulnerable to corpus poisoning attacks, where adversarial passages can be injected to manipulate retrieval results. Existing methods for crafting such passages, such as random token replacement or training inversion models, are often slow and computationally expensive, requiring either access to retriever’s gradients or large computational resources. To address these limitations, we propose Dynamic Importance-Guided Genetic Algorithm (DIGA), an efficient black-box method that leverages two key properties of retrievers: insensitivity to token order and bias towards influential tokens. By focusing on these characteristics, DIGA dynamically adjusts its genetic operations to generate effective adversarial passages with significantly reduced time and memory usage. Our experimental evaluation shows that DIGA achieves superior efficiency and scalability compared to existing methods, while maintaining comparable or better attack success rates across multiple datasets.