ACL 2025 Papers

To adapt large language models (LLMs) to ranking tasks, existing list-wise methods, represented by list-wise Direct Preference Optimization (DPO), focus on optimizing partial-order or full-order list ranking consistency for LLMs to enhance their ranking abilities.However, we argue that optimizing top-K ranking consistency could be more appropriate for real-world applications. There are two main reasons: (1) users are typically concerned with only the top-K results, making top-K ranking more important, and (2) tail items often lack precise feedback, making top-K ranking more reliable. Based on this, we propose K-order Ranking Preference Optimization (KPO) by extending the DPO’s Plackett-Luce model to accommodate top-K rankings. Additionally, recognizing that the number of important items can vary across queries, we extend KPO to dynamically determine appropriate K for different samples and introduce a curriculum learning strategy to boost training efficiency. Extensive experiments demonstrate the effectiveness of KPO, highlighting its high sample efficiency and robustness to noise. The code is available at https://github.com/Lanyu0303/KPO.

Self-harm detection on social media is critical for early intervention and mental health support, yet remains challenging due to the subtle, context-dependent nature of such expressions. Identifying self-harm intent aids suicide prevention by enabling timely responses, but current large language models (LLMs) struggle to interpret implicit cues in casual language and emojis. This work enhances LLMs’ comprehension of self-harm by distinguishing intent through nuanced language–emoji interplay. We present the Centennial Emoji Sensitivity Matrix (CESM-100)—a curated set of 100 emojis with contextual self-harm interpretations—and the Self-Harm Identification aNd intent Extraction with Supportive emoji sensitivity (SHINES) dataset, offering detailed annotations for self-harm labels, casual mentions (CMs), and serious intents (SIs). Our unified framework:a) enriches inputs using CESM-100;b) fine-tunes LLMs for multi-task learning—self-harm detection (primary) and CM/SI span detection (auxiliary);c) generate explainable rationales for self-harm predictions. We evaluate the framework on three state-of-the-art LLMs—Llama 3, Mental-Alpaca, and MentalLlama—across zero-shot, few-shot, and fine-tuned scenarios. By coupling intent differentiation with contextual cues, our approach commendably enhances LLM performance in both detection and explanation tasks, effectively addressing the inherent ambiguity in self-harm signals. The SHINES dataset, CESM-100 and codebase are publicly available at: https://www.iitp.ac.in/%7eai-nlp-ml/resources.html#SHINES

LLM use in annotation is becoming widespread, and given LLMs’ overall promising performance and speed, putting humans in the loop to simply “review” LLM annotations can be tempting. In subjective tasks with multiple plausible answers, this can impact both evaluation of LLM performance, and analysis using these labels in a social science task downstream. In a pre-registered experiment with 350 unique annotators and 7,000 annotations across 4 conditions, 2 models, and 2 datasets, we find that presenting crowdworkers with LLM-generated annotation suggestions did not make them faster annotators, but did improve their self-reported confidence in the task. More importantly, annotators strongly took the LLM suggestions, significantly changing the label distribution compared to the baseline. We show that when these labels created with LLM assistance are used to evaluate LLM performance, reported model performance significantly increases. We show how changes in label distributions as a result of LLM assistance can affect conclusions drawn by analyzing even “human-approved” LLM-annotated datasets. We believe our work underlines the importance of understanding the impact of LLM-assisted annotation on subjective, qualitative tasks, on the creation of gold data for training and testing, and on the evaluation of NLP systems on subjective tasks.

Detecting toxicity in online multimodal environments, such as memes, remains a challenging task due to the complex contextual connections across modalities (e.g., text and visual), which demand both common-sense reasoning and contextual awareness. To bridge this gap, we propose a hybrid neurosymbolic framework that unifies (1) distillation of implicit contextual knowledge (e.g., sarcasm, cultural references) from Large Vision-Language Models (LVLMs) and (2) infusion of explicit relational semantics through sub-graphs from Knowledge Graphs (KGs). Experimental results on two benchmark datasets show the superior performance of our approach, Knowledge-Infused Distilled Vision-Language Model (KID-VLM), over the state-of-the-art baselines across AUC and F1, with improvements of 0.5%, and 10.6%, respectively, in HatefulMemes Benchmark across variants. Further, KID-VLM demonstrates better generalizability and achieves the best performance across all baselines in the HarMeme Dataset with a 6.3% and 3.2% in F1 and AUC.Given the contextual complexity of the toxicity detection, KID-VLM showcases the significance of learning compact models (~500M parameters) from both explicit (i.e., KG) and implicit (i.e., LVLMs) contextual cues incorporated through a hybrid neurosymbolic approach. Our codes and pretrained models are publicly available.

Large language models (LLMs) have demonstrated impressive translation capabilities even without being explicitly trained on parallel data. This remarkable property has led some to believe that parallel data is no longer necessary for building multilingual language models. While some attribute this to the emergent abilities of LLMs due to scale, recent work suggests that it is actually caused by incidental bilingual signals present in the training data. Various methods have been proposed to maximize the utility of parallel data to enhance the multilingual capabilities of multilingual encoder-based and encoder-decoder language models. However, some decoder-based LLMs opt to ignore parallel data instead. In this work, we conduct a systematic study on the impact of adding parallel data on LLMs’ multilingual capabilities, focusing specifically on translation and multilingual common-sense reasoning. Through controlled experiments, we demonstrate that parallel data can significantly improve LLMs’ multilingual capabilities.

Retrieval-Augmented Generation (RAG) has proven its effectiveness in alleviating hallucinations for Large Language Models (LLMs). However, existing automated evaluation metrics cannot fairly evaluate the outputs generated by RAG models during training and evaluation. LLM-based judgment models provide the potential to produce high-quality judgments, but they are highly sensitive to evaluation prompts, leading to inconsistencies when judging the output of RAG models. This paper introduces the Judge-Consistency (ConsJudge) method, which aims to enhance LLMs to generate more accurate evaluations for RAG models. Specifically, ConsJudge prompts LLMs to generate different judgments based on various combinations of judgment dimensions, utilizes the judge-consistency to evaluate these judgments, and selects the chosen and rejected judgments for DPO training. Our experiments show that ConsJudge can effectively provide more accurate judgments for optimizing RAG models across various RAG models and datasets. Further analysis reveals that judgments generated by ConsJudge have a high agreement with the superior LLM. All codes are available at https://github.com/OpenBMB/ConsJudge.

Given the rapid progress of generative AI, there is a pressing need to systematically compare and choose between the numerous models and configurations available. The scale and versatility of such evaluations make the use of LLM-based judges a compelling solution for this challenge. Crucially, this approach requires first to validate the quality of the LLM judge itself. Previous work has focused on instance-based assessment of LLM judges, where a judge is evaluated over a set of responses, or response pairs, while being agnostic to their source systems. We argue that this setting overlooks critical factors affecting system-level ranking, such as a judge’s positive or negative bias towards certain systems. To address this gap, we conduct the first large-scale study of LLM judges as system rankers. System scores are generated by aggregating judgment scores over multiple system outputs, and the judge’s quality is assessed by comparing the resulting system ranking to a human-based ranking. Beyond overall judge assessment, our analysis provides a fine-grained characterization of judge behavior, including their decisiveness and bias.

Instruction tuning is vital for enhancing the performance of large language models (LLMs), but existing text-to-text methods, referred to as TextTuning, struggle with issues such as generalization, robustness, and controllability due to their lack of explicit task structures. We introduce JsonTuning, a structure-to-structure approach that uses JSON structures to represent tasks. This method improves generalization by clarifying task elements and their relations, boosts robustness by minimizing ambiguity, and enhances controllability by allowing precise control over outputs. We conduct an extensive comparative analysis between JsonTuning and TextTuning using various language models and benchmarks. Our findings reveal that JsonTuning consistently surpasses TextTuning in terms of performance, robustness, and controllability across different scenarios. By overcoming the limitations of TextTuning, JsonTuning demonstrates significant potential for developing more effective and reliable LLMs capable of handling diverse scenarios.

Large Language Models (LLMs) have demonstrated impressive performances in complex text generation tasks. However, the contribution of the input prompt to the generated content still remains obscure to humans, underscoring the necessity of understanding the causality between input and output pairs. Existing works for providing prompt-specific explanation often confine model output to be classification or next-word prediction. Few initial attempts aiming to explain the entire language generation often treat input prompt texts independently, ignoring their combinatorial effects on the follow-up generation. In this study, we introduce a counterfactual explanation framework based on joint prompt attribution, JoPA, which aims to explain how a few prompt texts collaboratively influences the LLM’s complete generation. Particularly, we formulate the task of prompt attribution for generation interpretation as a combinatorial optimization problem, and introduce a probabilistic algorithm to search for the casual input combination in the discrete space. We define and utilize multiple metrics to evaluate the produced explanations, demonstrating both the faithfulness and efficiency of our framework.

Large language models (LLMs) excel in various tasks but remain vulnerable to jailbreak attacks, where adversaries manipulate prompts to generate harmful outputs. Examining jailbreak prompts helps uncover the shortcomings of LLMs. However, current jailbreak methods and the target model’s defenses are engaged in an independent and adversarial process, resulting in the need for frequent attack iterations and redesigning attacks for different models. To address these gaps, we propose a Reverse Embedded Defense Attack (REDA) mechanism that disguises the attack intention as the “defense”. intention against harmful content. Specifically, REDA starts from the target response, guiding the model to embed harmful content within its defensive measures, thereby relegating harmful content to a secondary role and making the model believe it is performing a defensive task. The attacking model considers that it is guiding the target model to deal with harmful content, while the target model thinks it is performing a defensive task, creating an illusion of cooperation between the two. Additionally, to enhance the model’s confidence and guidance in “defensive” intentions, we adopt in-context learning (ICL) with a small number of attack examples and construct a corresponding dataset of attack examples. Extensive evaluations demonstrate that the REDA method enables cross-model attacks without the need to redesign attack strategies for different models, enables successful jailbreak in one iteration, and outperforms existing methods on both open-source and closed-source models.

Jailbreak attacks aim to bypass the LLMs’ safeguards. While researchers have proposed different jailbreak attacks in depth, they have done so in isolation—either with unaligned settings or comparing a limited range of methods. To fill this gap, we present a large-scale evaluation of various jailbreak attacks. We collect 17 representative jailbreak attacks, summarize their features, and establish a novel jailbreak attack taxonomy. Then we conduct comprehensive measurement and ablation studies across nine aligned LLMs on 160 forbidden questions from 16 violation categories. Also, we test jailbreak attacks under eight advanced defenses. Based on our taxonomy and experiments, we identify some important patterns, such as heuristic-based attacks, which could achieve high attack success rates but are easy to mitigate by defenses. Our study offers valuable insights for future research on jailbreak attacks and defenses and serves as a benchmark tool for researchers and practitioners to evaluate them effectively.

With the rapid advancement of Large Vision-Language Models (VLMs), concerns about their ‌potential misuse and abuse have grown rapidly. Prior research has exposed VLMs’ vulnerability to jailbreak attacks, where carefully crafted inputs can lead the model to produce content that violates ethical and legal standards. However, current jailbreak methods often fail against cutting-edge models such as GPT-4o. We attribute this to the over-exposure of harmful content and the absence of stealthy malicious guidance. In this work, we introduce a novel jailbreak framework: Multi-Modal Linkage (MML) Attack. Drawing inspiration from cryptography, MML employs an encryption-decryption process across text and image modalities to mitigate the over-exposure of malicious information. To covertly align the model’s output with harmful objectives, MML leverages a technique we term evil alignment, framing the attack within the narrative context of a video game development scenario. Extensive experiments validate the effectiveness of MML. Specifically, MML jailbreaks GPT-4o with attack success rates of 99.40% on SafeBench, 98.81% on MM-SafeBench, and 99.07% on HADES-Dataset. Our code is available at https://github.com/wangyu-ovo/MML.

Though online medical literature has made health information more available than ever, the barrier of complex medical jargon prevents the general public from understanding it. Though parallel and comparable corpora for Biomedical Text Simplification have been introduced, these conflate the many syntactic and lexical operations involved in simplification. To enable more targeted development and evaluation, we present a fine-grained lexical simplification task and dataset, Jargon Explanations for Biomedical Simplification (JEBS). The JEBS task involves identifying complex terms, classifying how to replace them, and generating replacement text. The JEBS dataset contains 21,595 replacements for 10,314 terms across 400 biomedical abstracts and their manually simplified versions. Additionally, we provide baseline results for a variety of rule-based and transformer-based systems for the three subtasks. The JEBS task, data, and baseline results pave the way for development and rigorous evaluation of systems for replacing or explaining complex biomedical terms.

Recently, action-based decision-making in open-world environments has gained significant attention. Visual Language Action (VLA) models, pretrained on large-scale web datasets, have shown promise in decision-making tasks. However, previous work has primarily focused on action post-training, often neglecting enhancements to the foundation model itself. In response, we introduce Act from Visual Language Post-Training (ActVLP), a novel training paradigm. ActVLP distinctively enhances the foundation model prior to action-specific tuning by first post-training it on a curated set of environment-specific visual and linguistic tasks using self-supervised learning. This initial stage significantly improves the model’s capabilities in world knowledge, visual recognition, and spatial grounding. Subsequently, this strengthened VLM undergoes action post-training via imitation learning on trajectory datasets.Following this paradigm, we develop JARVIS-VLA, the first VLA model in Minecraft that can follow human instructions on over 1k different atomic tasks, including crafting, smelting, cooking, mining, and killing. Our experiments demonstrate that our ActVLP paradigm leads to a significant 40% improvement over the best agent baseline on a diverse set of atomic tasks. Furthermore, JARVIS-VLA surpasses traditional imitation learning-based policies in Minecraft, achieving state-of-the-art performance. We have open-sourced the code, models, and datasets to foster further research.The project page can be found at https://craftjarvis.github.io/JarvisVLA.

Idioms are defined as a group of words with a figurative meaning not deducible from their individual components. Although modern machine translation systems have made remarkable progress, translating idioms remains a major challenge, especially for speech-to-text systems, where research on this topic is notably sparse. In this paper, we systematically evaluate idiom translation as compared to conventional news translation in both text-to-text machine translation (MT) and speech-to-text translation (SLT) systems across two language pairs (German to English, Russian to English). We compare state-of-the-art end-to-end SLT systems (SeamlessM4T SLT-to-text, Whisper Large v3) with MT systems (SeamlessM4T SLT-to-text, No Language Left Behind), Large Language Models (DeepSeek, LLaMA) and cascaded alternatives. Our results reveal that SLT systems experience a pronounced performance drop on idiomatic data, often reverting to literal translations even in higher layers, whereas MT systems and Large Language Models demonstrate better handling of idioms. These findings underscore the need for idiom-specific strategies and improved internal representations in SLT architectures.

Bragging, as a pervasive social-linguistic phenomenon, reflects complex human interaction patterns. However, the understanding and generation of appropriate bragging behavior in large language models (LLMs) remains underexplored. In this paper, we propose a comprehensive study that combines analytical and controllable approaches to examine bragging in LLMs. We design three tasks, bragging recognition, bragging explanation, and bragging generation, along with novel evaluation metrics to assess the models’ ability to identify bragging intent, social appropriateness, and account for context sensitivity. Our analysis reveals the challenges of bragging in the social context, such as recognizing bragging and responding appropriately with bragging in conversation. This work provides new insights into how LLMs process bragging and highlights the need for more research on generating contextually appropriate behavior in LLMs.

Real-world applications of KBQA require models to detect different types of unanswerable questions with a limited volume of in-domain labeled training data. We propose the novel task of few-shot transfer for KBQA with unanswerable questions. The state-of-the-art KBQA few-shot transfer model (FuSIC-KBQA) uses an iterative repair strategy that assumes that all questions are answerable. As a remedy, we present FUn-FuSIC – a novel solution for our task that extends FuSIC-KBQA with Feedback for Unanswerability (FUn), which is an iterative repair strategy for answerable as well as unanswerable questions. FUn uses feedback from a suite of strong and weak verifiers, and an adaptation of self-consistency for unanswerability for assessing answerability of questions. Our experiments show that FUn-FuSIC significantly outperforms suitable adaptations of multiple LLM-based and supervised SoTA models on our task, while establishing a new SoTA performance for answerable few-shot transfer as well. We have made datasets and other resources publicly available at https://github.com/dair-iitd/funfusic/.

Data augmentation, a widely-employed technique for addressing data scarcity, involves generating synthetic data examples which are then used to augment available training data. Researchers have seen surprising success from simple methods, such as random perturbations from natural examples, where models seem to benefit even from data with nonsense words, or data that doesn’t conform to the rules of the language. A second line of research produces synthetic data that does in fact follow all linguistic constraints; these methods require some linguistic expertise and are generally more challenging to implement. No previous work has done a systematic, empirical comparison of both linguistically-naive and linguistically-motivated data augmentation strategies, leaving uncertainty about whether the additional time and effort of linguistically-motivated data augmentation work in fact yields better downstream performance.In this work, we conduct a careful and comprehensive comparison of augmentation strategies (both linguistically-naive and linguistically-motivated) for two low-resource languages with different morphological properties, Uspanteko and Arapaho. We evaluate the effectiveness of many different strategies and their combinations across two important sequence-to-sequence tasks for low-resource languages: machine translation and interlinear glossing. We find that linguistically-motivated strategies can have benefits over naive approaches, but only when the new examples they produce are not significantly unlike the training data distribution.

We investigate static and contextualized embeddings for English pseudowords across a variety of Large Language Models (LLMs), to study (i) how these models represent semantic attributes of strings they encounter for the very first time and how (ii) these representations interact with sentence context. We zoom in on a key semantic attribute, valence, which plays an important role in theories of language processing, acquisition, and evolution. Across three experiments, we show that pseudoword valence is encoded in meaningful ways both in isolation and in context, and that, in some LLMs, pseudowords affect the representation of whole sentences similarly to words. This highlights how, at least for most LLMs we surveyed, pseudowords and words are not qualitatively different constructs. Our study confirms that LLMs capture systematic mappings between form and valence, and shows how different LLMs handle the contextualisation of pseudowords differently. Our findings provide a first computational exploration of how sub-lexical distributional patterns influence the valence of novel strings in context, offering useful insights for theories on the form-meaning interface and how it affects language learning and processing.