ACL 2025 Papers

Misinformation is a global issue that shapes public discourse, influencing opinions and decision-making across various domains. While automated fact-checking (AFC) has become essential in combating misinformation, most work in multilingual settings has focused on claim verification rather than generating explanatory verdicts (i.e. short texts discussing the veracity of the claim), leaving a gap in AFC resources beyond English.To this end, we introduce EuroVerdict, a multilingual dataset designed for verdict generation, covering eight European languages. Developed in collaboration with professional fact-checkers, the dataset comprises claims, manually written verdicts, and supporting evidence, including fact-checking articles and additional secondary sources. We evaluate EuroVerdict with Llama-3.1-8B-Instruct on verdict generation under different settings, varying the prompt language, input article language, and training approach. Our results show that fine-tuning consistently improves performance, with models fine-tuned on original-language articles achieving the highest scores in both automatic and human evaluations. Using articles in a different language from the claim slightly lowers performance; however, pairing them with language-specific prompts improves results. Zero-shot and Chain-of-Thought setups perform worse, reinforcing the benefits of fine-tuning for multilingual verdict generation.

In recent years, researchers have started analyzing the cultural sensitivity of LLMs. In this respect, Etiquettes have been an active area of research. Etiquettes are region-specific and are an essential part of the culture of a region; hence, it is imperative to make LLMs sensitive to etiquettes. However, there needs to be more resources in evaluating LLMs for their understanding and bias with regard to etiquettes. In this resource paper, we introduce EtiCor++, a corpus of etiquettes worldwide. We introduce different tasks for evaluating LLMs for knowledge about etiquettes across various regions. Further, we introduce various metrics for measuring bias in LLMs. Extensive experimentation with LLMs shows inherent bias towards certain regions.

Self-supervised learning (SSL) has reduced the reliance on expensive labeling in speech technologies by learning meaningful representations from unannotated data. Since most SSL-based downstream tasks prioritize content information in speech, ideal representations should disentangle content from unwanted variations like speaker characteristics in the SSL representations. However, removing speaker information often degrades other speech components, and existing methods either fail to fully disentangle speaker identity or require resource-intensive models. In this paper, we propose a novel disentanglement method that linearly decomposes SSL representations into speaker-specific and speaker-independent components, effectively generating speaker disentangled representations. Comprehensive experiments show that our approach achieves speaker independence and as such, when applied to content-driven tasks such as voice conversion, our representations yield significant improvements over state-of-the-art methods.

Language models (LMs) rely on their parametric knowledge augmented with relevant contextual knowledge for certain tasks, such as question answering. However, the contextual knowledge can contain private information that may be leaked when answering queries, and estimating this privacy leakage is not well understood. A straightforward approach of directly comparing an LM’s output to the contexts can overestimate the privacy risk, since the LM’s parametric knowledge might already contain the augmented contextual knowledge. To this end, we introduce context influence, a metric that builds on differential privacy, a widely-adopted privacy notion, to estimate the privacy leakage of contextual knowledge during decoding. Our approach effectively measures how each subset of the context influences an LM’s response while separating the specific parametric knowledge of the LM. Using our context influence metric, we demonstrate that context privacy leakage occurs when contextual knowledge is out of distribution with respect to parametric knowledge. Moreover, we experimentally demonstrate how context influence properly attributes the privacy leakage to augmented contexts, and we evaluate how factors– such as model size, context size, generation position, etc.– affect context privacy leakage. The practical implications of our results will inform practitioners of the privacy risk associated with augmented contextual knowledge.

The development of large language models (LLMs) depends on **trustworthy evaluation**. However, most current evaluations rely on public benchmarks, which are prone to data contamination issues that significantly compromise fairness. Previous researches have focused on constructing dynamic benchmarks to address contamination. However, continuously building new benchmarks is costly and cyclical.In this work, we aim to tackle contamination by analyzing the mechanisms of contaminated models themselves. Through our experiments, we discover that the overestimation of contaminated models is likely due to parameters acquiring shortcut solutions in training. We further propose a novel method for identifying shortcut neurons through **comparative and causal analysis**.Building on this, we introduce an evaluation method called **shortcut neuron patching** to suppress shortcut neurons. Experiments validate the effectiveness of our approach in mitigating contamination. Additionally, our evaluation results exhibit a strong linear correlation with MixEval, a recently released trustworthy benchmark, achieving a Spearman coefficient (𝜌) exceeding 0.95. This high correlation indicates that our method closely reveals true capabilities of the models and is trustworthy. We conduct further experiments to demonstrate the generalizability of our method across various benchmarks and hyperparameter settings. **Code**: https://github.com/GaryStack/Trustworthy-Evaluation.

Automated Essay Scoring (AES) plays a crucial role in educational assessment by providing scalable and consistent evaluations of writing tasks. However, traditional AES systems face three major challenges: (i) reliance on handcrafted features that limit generalizability, (ii) difficulty in capturing fine-grained traits like coherence and argumentation, and (iii) inability to handle multimodal contexts. In the era of Multimodal Large Language Models (MLLMs), we propose **EssayJudge**, the **first multimodal benchmark to evaluate AES capabilities across lexical-, sentence-, and discourse-level traits**. By leveraging MLLMs’ strengths in trait-specific scoring and multimodal context understanding, EssayJudge aims to offer precise, context-rich evaluations without manual feature engineering, addressing longstanding AES limitations. Our experiments with 18 representative MLLMs reveal gaps in AES performance compared to human evaluation, particularly in discourse-level traits, highlighting the need for further advancements in MLLM-based AES research. Our dataset and code will be available upon acceptance.

Language model agents excel in long-session planning and reasoning, but existing benchmarks primarily focus on goal-oriented tasks with explicit objectives, neglecting creative adaptation in unfamiliar environments. To address this, we introduce EscapeBench—a benchmark suite of room escape game environments designed to challenge agents with creative reasoning, unconventional tool use, and iterative problem-solving to uncover implicit goals. Our results show that current LM models, despite employing working memory and Chain-of-Thought reasoning, achieve only 15% average progress without hints, highlighting their limitations in creativity. To bridge this gap, we propose EscapeAgent, a framework designed to enhance creative reasoning through Foresight (innovative tool use) and Reflection (identifying unsolved tasks). Experiments show that EscapeAgent can execute action chains over 1,000 steps while maintaining logical coherence. It navigates and completes games with up to 40% fewer steps and hints, performs robustly across difficulty levels, and achieves higher action success rates with more efficient and innovative puzzle-solving strategies.

Large Multimodal Models (LMMs) have demonstrated impressive performance across numerous academic benchmarks. However, fine-tuning still remains essential to achieve satisfactory performance on downstream tasks, while the task-specific tuning samples are usually not readily available or expensive and time-consuming to obtain. To address this, we propose an error-driven data-efficient tuning framework that aims to efficiently adapt generic LMMs to newly emerging tasks without requiring extensive task-specific training samples. In our approach, a generic LMM, acting as a student model, is first evaluated on a small validation set of the target task, and then a more powerful model, acting as a teacher model, identifies the erroneous steps within the student model’s reasoning steps and analyzes its capability gaps from fully addressing the target task. Based on these gaps, targeted training samples are further retrieved from existing task-agnostic datasets to tune the student model and tailor it to the target task. We perform extensive experiments across three different training data scales and seven tasks, demonstrating that our training paradigm significantly and efficiently improves LMM’s performance on downstream tasks, achieving an average performance boost of 7.01%

Supervised fine-tuning (SFT) has enabled large language models (LLMs) to exhibit promising performance on various tasks. However, this fine-tuning process only compares current predictions and labels on each sample, yet fails to perceive and understand its error outputs from different degrees, which may potentially produce a large percentage of serious errors. This poses a problem for aspect-based sentiment analysis (ABSA) in that these serious errors bring a greater negative impact than acceptable ones. Humans tend to compare mistakes to understand the varying degrees of mistakes, thus avoiding major bad decisions. Inspired by this, we propose a simple yet effective framework that could perceive and understand the degree of different errors by learning from comparative error pairs. It utilizes the SFT model to yield multiple outputs on each sample and selects acceptable and severe errors based on the acceptable scores. Together with the labels, we construct two comparative error pairs and exploit their calibration losses to optimize parameters. We conduct comprehensive experiments on ABSA datasets to demonstrate the effectiveness of our framework over baselines.

The remarkable performance of Large language models (LLMs) relies heavily on the availability of abundant high-quality training data. However, the high cost of acquiring annotated data often prevents models from obtaining capabilities to tackle downstream tasks. In this paper, we introduce a novel method, EpiCoDe that boosts model performance in data-scarcity scenarios without extra training. We first employ model extrapolation to enhance a finetuned model with its inferior version, and then adopt contrastive decoding to further reduce predicted errors, by comparing the logit scores given by the extrapolated and the vanilla finetuned model. Experiments across three domains over four different LLMs show that EpiCoDe consistently outperforms existing methods with significant and robust improvement. We also propose a new theoretical framework to reveal the mechanism behind contrastive decoding in data-scarcity scenarios, which further helps better understand the effectiveness of our EpiCoDe.

Recent advances in Large Language Models (LLMs) have yielded impressive successes on many language tasks. However, efficient processing of long contexts using LLMs remains a significant challenge. We introduce **EpMAN** – a method for processing long contexts in an episodic memory module while holistically attending to semantically-relevant context chunks. Output from episodic attention is then used to reweigh the decoder’s self-attention to the stored KV cache of the context during training and generation. When an LLM decoder is trained using **EpMAN**, its performance on multiple challenging single-hop long-context recall and question-answering benchmarks is found to be stronger and more robust across the range from 16k to 256k tokens than baseline decoders trained with self-attention, and popular retrieval-augmented generation frameworks.

Self-reflection helps de-hallucinate Large Language Models (LLMs). However, the effectiveness of self-reflection remains insufficiently validated in the context of Small Language Models (SLMs), which exhibit limited semantic capacities. In particular, we demonstrate that the conventional self-reflection paradigm, such as Self-Refine, fails to deliver robust response refinement for models with parameter sizes of 10 billion or smaller, even when compared to generations elicited through Chain-of-Thought (CoT) prompting. To improve SLMs’ self-reflection, we redesign Self-Refine and introduce Entrospect (ENTROpy-aware IntroSPECTion), an information-theoretic framework based on prompt engineering.We evaluated Entrospect using accuracy and average time consumption metrics to comprehensively assess its precision and computational efficiency. Experiments conducted across four distinct SLMs and four baseline methods demonstrate that Entrospect achieves state-of-the-art performance on validation tasks. Notably, under identical model and data settings, Entrospect delivers a remarkable improvement of up to 36.2 in reasoning accuracy while enhancing computational efficiency by as much as 10 times compared to its predecessor, Self-Refine.

Multi-step processes via large language models (LLMs) have proven effective for solving complex reasoning tasks. However, the depth of exploration of the reasoning procedure can significantly affect the task performance. Existing methods to automatically decide the depth often lead to high cost and a lack of flexibility. To address these issues, we propose Entropy-based Exploration Depth Conduction (Entro-duction), a novel method that dynamically adjusts the exploration depth during multi-step reasoning by monitoring LLM’s output entropy and variance entropy. We employ these two features to capture the model’s uncertainty of the current step and the fluctuation of uncertainty across consecutive reasoning steps. Based on the observed entropy changes, the LLM selects whether to deepen, expand, or stop exploration according to the probability, which facilitates the trade-off between the reasoning accuracy and exploration effectiveness. Experimental results across four benchmark datasets demonstrate the efficacy of Entro-duction.

The retriever, which retrieves relevant knowledge pieces from a knowledge base given a context, is an important component in many natural language processing (NLP) tasks. Retrievers have been introduced in knowledge-grounded dialog systems to improve knowledge acquisition. In knowledge-grounded dialog systems, when conditioning on a given context, there may be multiple relevant and correlated knowledge pieces. However, knowledge pieces are usually assumed to be conditionally independent in current retriever models. To address this issue, we propose Entriever, an energy-based retriever. The Entriever directly models the candidate retrieval results as a whole instead of modeling the knowledge pieces separately, with the relevance score defined by an energy function. We explore various architectures of energy functions and different training methods for Entriever, and show that Entriever substantially outperforms the strong cross-encoder baseline in knowledge retrieval tasks. Furthermore, we show that in semi-supervised training of knowledge-grounded dialog systems, Entriever enables the effective scoring of retrieved knowledge pieces and leads to a significant improvement in the end-to-end performance of the dialog system.

We introduce a novel multilingual and hierarchical corpus annotated for entity framing and role portrayal in news articles. The dataset uses a unique taxonomy inspired by storytelling elements, comprising 22 fine-grained roles, or archetypes, nested within three main categories: protagonist, antagonist, and innocent. Each archetype is carefully defined, capturing nuanced portrayals of entities such as guardian, martyr, and underdog for protagonists; tyrant, deceiver, and bigot for antagonists; and victim, scapegoat, and exploited for innocents. The dataset includes 1,378 recent news articles in five languages (Bulgarian, English, Hindi, European Portuguese, and Russian) focusing on two critical domains of global significance: the Ukraine-Russia War and Climate Change. Over 5,800 entity mentions have been annotated with role labels. This dataset serves as a valuable resource for research into role portrayal and has broader implications for news analysis. We describe the characteristics of the dataset and the annotation process, and we report evaluation results on fine-tuned state-of-the-art multilingual transformers and hierarchical zero-shot learning using LLMs at the level of a document, a paragraph, and a sentence.

First-order logic (FOL) is often used to represent logical entailment, but determining natural language (NL) entailment using FOL remains a challenge. To address this, we propose the Entailment-Preserving FOL representations (EPF) task and introduce reference-free evaluation metrics for EPF (Entailment-Preserving Rate (EPR) family). In EPF, one should generate FOL representations from multi-premise NL entailment data (e.g., EntailmentBank) so that the automatic prover’s result preserves the entailment labels. Furthermore, we propose a training method specialized for the task, iterative learning-to-rank, which trains an NL-to-FOL translator by using the natural language entailment labels as verifiable rewards. Our method achieves a 1.8–2.7% improvement in EPR and a 17.4–20.6% increase in EPR@16 compared to diverse baselines in three datasets. Further analyses reveal that iterative learning-to-rank effectively suppresses the arbitrariness of FOL representation by reducing the diversity of predicate signatures, and maintains strong performance across diverse inference types and out-of-domain data.

Much of human communication depends on implication, conveying meaning beyond literal words to express a wider range of thoughts, intentions, and feelings. For models to better understand and facilitate human communication, they must be responsive to the text’s implicit meaning. We focus on Natural Language Inference (NLI), a core tool for many language tasks, and find that state-of-the-art NLI models and datasets struggle to recognize a range of cases where entailment is implied, rather than explicit from the text. We formalize implied entailment as an extension of the NLI task and introduce the Implied NLI dataset (INLI) to help today’s LLMs both recognize a broader variety of implied entailments and to distinguish between implicit and explicit entailment. We show how LLMs fine-tuned on INLI understand implied entailment and can generalize this understanding across datasets and domains.

As large language models (LLMs) reach human-like fluency, reliably distinguishing AI-generated text from human authorship becomes increasingly difficult. While watermarks already exist for LLMs, they often lack flexibility and struggle with attacks such as paraphrasing. To address these issues, we propose a multi-feature method for generating watermarks that combines multiple distinct watermark features into an ensemble watermark. Concretely, we combine acrostica and sensorimotor norms with the established red-green watermark to achieve a 98% detection rate. After a paraphrasing attack, the performance remains high with 95% detection rate. In comparison, the red-green feature alone as a baseline achieves a detection rate of 49% after paraphrasing. The evaluation of all feature combinations reveals that the ensemble of all three consistently has the highest detection rate across several LLMs and watermark strength settings. Due to the flexibility of combining features in the ensemble, various requirements and trade-offs can be addressed. Additionally, the same detection function can be used without adaptations for all ensemble configurations. This method is particularly of interest to facilitate accountability and prevent societal harm.

Large language models (LLMs) exhibit the ability to generalize given few-shot examples in their input prompt, an emergent capability known as in-context learning (ICL). We investigate whether LLMs use ICL to perform structured reasoning in ways that are consistent with a Bayesian framework or rely on pattern matching. Using a controlled setting of biased coin flips, we find that: (1) LLMs often possess biased priors, causing initial divergence in zero-shot settings, (2) in-context evidence outweighs explicit bias instructions, (3) LLMs broadly follow Bayesian posterior updates, with deviations primarily due to miscalibrated priors rather than flawed updates, and (4) attention magnitude has negligible effect on Bayesian inference. With sufficient demonstrations of biased coin flips via ICL, LLMs update their priors in a Bayesian manner. Code and visualizations are available on the [project page](https://ai-climate.berkeley.edu/llm-coin-flips/).