NAACL 2025 Papers

Whilst large language models (LLMs) have made significant advances in every natural language processing task, studies have shown that these models are vulnerable to small perturbations in the inputs, raising concerns about their robustness in the real-world. Given the rise of misinformation online and its significant impact on society, fact verification is one area in which assessing the robustness of models developed for this task is crucial. However, the robustness of LLMs in fact verification remains largely unexplored. In this paper, we introduce FactEval, a novel large-scale benchmark for extensive evaluation of LLMs in the fact verification domain covering 17 realistic word-level and character-level perturbations and 4 types of subpopulations. We investigate the robustness of several LLMs in zero-shot, few-shot, and chain-of-thought prompting. Our analysis using FEVER, one of the largest and most widely-used datasets for fact verification, reveals that LLMs are brittle to small input changes and also exhibit performance variations across different subpopulations.

Prior research on training grounded factuality classification models to detect hallucinations in large language models (LLMs) has relied on public natural language inference (NLI) data and synthetic data. However, conventional NLI datasets are not well-suited for document-level reasoning, which is critical for detecting LLM hallucinations. Recent approaches to document-level synthetic data generation involve iteratively removing sentences from documents and annotating factuality using LLM-based prompts. While effective, this method is computationally expensive for long documents and limited by the LLM’s capabilities. In this work, we analyze the differences between existing synthetic training data used in state-of-the-art models and real LLM output claims. Based on our findings, we propose a novel approach for synthetic data generation, CG2C, that leverages multi-hop reasoning on context graphs extracted from documents. Our fact checker model, FactCG, demonstrates improved performance with more connected reasoning, using the same backbone models. Experiments show it even outperforms GPT-4-o on the LLM-Aggrefact benchmark with much smaller model size.

Multimodal foundation models hold significant potential for automating radiology report generation, thereby assisting clinicians in diagnosing cardiac diseases. However, generated reports often suffer from serious factual inaccuracy. In this paper, we introduce a fact-aware multimodal retrieval-augmented pipeline in generating accurate radiology reports (FactMM-RAG). We first leverage RadGraph to mine factual report pairs, then integrate factual knowledge to train a universal multimodal retriever. Given a radiology image, our retriever can identify high-quality reference reports to augment multimodal foundation models, thus enhancing the factual completeness and correctness of report generation. Experiments on two benchmark datasets demonstrate that our multimodal retriever significantly outperforms other state-of-the-art retrievers on both language generation and radiology-specific metrics, up to 6.5% and 2% score in F1CheXbert and F1RadGraph. Further analysis indicates that employing our factually-informed training strategy imposes an effective supervision signal, without relying on explicit diagnostic label guidance, and successfully propagate fact-aware capabilities from the multimodal retriever to the multimodal foundation model in radiology report generation.

Large Language Models (LLMs) have demonstrated significant performance improvements across various cognitive tasks. An emerging application is using LLMs to enhance retrieval-augmented generation (RAG) capabilities. These systems require LLMs to understand user queries, retrieve relevant information, and synthesize coherent and accurate responses. Given the increasing real-world deployment of such systems, comprehensive evaluation becomes crucial. To this end, we propose FRAMES (Factuality, Retrieval, And reasoning MEasurement Set), a high-quality evaluation dataset designed to test LLMs’ ability to provide factual responses, assess retrieval capabilities, and evaluate the reasoning required to generate final answers. While previous work has provided datasets and benchmarks to evaluate these abilities in isolation, FRAMES offers a unified framework that provides a clearer picture of LLM performance in end-to-end RAG scenarios. Our dataset comprises challenging multi-hop questions that require the integration of information from multiple sources. We present baseline results demonstrating that even state-of-the-art LLMs struggle with this task, achieving 0.40 accuracy with no retrieval. The accuracy is significantly improved with our proposed multi-step retrieval pipeline, achieving an accuracy of 0.66 (>50% improvement). We hope our work will help bridge evaluation gaps and assist in developing more robust and capable RAG systems.

Tracking a patient’s cognitive status early in the onset of the disease provides an opportunity to diagnose and intervene in Alzheimer’s disease (AD). However, relying solely on magnetic resonance imaging (MRI) images with traditional classification and regression models may not fully extract finer-grained information. This study proposes a multi-task Fusion Language Image Question Answering model (FLIQA-AD) to perform AD identification and Mini Mental State Examination (MMSE) prediction. Specifically, a 3D Adapter is introduced in Vision Transformer (ViT) model for image feature extraction. The patient electronic health records (EHR) information and questions related to the disease work as text prompts to be encoded. Then, an ADFormer model, which combines self-attention and cross-attention mechanisms, is used to capture the correlation between EHR information and structure features. After that, the extracted brain structural information and textual content are combined as input sequences for the large language model (LLM) to identify AD and predict the corresponding MMSE score. Experimental results demonstrate the strong discrimination and MMSE prediction performance of the model, as well as question-answer capabilities.

Text-to-SQL systems have become crucial for translating natural language into SQL queries in various industries, enabling non-technical users to perform complex data operations. The need for accurate evaluation methods has increased as these systems have grown more sophisticated. However, the Execution Accuracy (EX), the most prevalent evaluation metric, still shows many false positives and negatives. Thus, this paper introduces **FLEX(False-Less EXecution)**, a novel approach to evaluating text-to-SQL systems using large language models (LLMs) to emulate human expert-level evaluation of SQL queries. Our metric improves agreement with human experts (from 62 to 87.04 in Cohen’s kappa) with comprehensive context and sophisticated criteria. Our extensive experiments yield several key insights: (1) Models’ performance increases by over 2.6 points on average, substantially affecting rankings on Spider and BIRD benchmarks; (2) The underestimation of models in EX primarily stems from annotation quality issues; and (3) Model performance on particularly challenging questions tends to be overestimated. This work contributes to a more accurate and nuanced evaluation of text-to-SQL systems, potentially reshaping our understanding of state-of-the-art performance in this field.

Recent advancements in Large Language Models (LLMs) have significantly enhanced interactions between users and models. These advancements concurrently underscore the need for rigorous safety evaluations due to the manifestation of social biases, which can lead to harmful societal impacts. Despite these concerns, existing benchmarks may overlook the intrinsic weaknesses of LLMs, which can generate biased responses even with simple adversarial instructions. To address this critical gap, we introduce a new benchmark, Fairness Benchmark in LLM under Extreme Scenarios (FLEX), designed to test whether LLMs can sustain fairness even when exposed to prompts constructed to induce bias. To thoroughly evaluate the robustness of LLMs, we integrate prompts that amplify potential biases into the fairness assessment. Comparative experiments between FLEX and existing benchmarks demonstrate that traditional evaluations may underestimate the inherent risks in models. This highlights the need for more stringent LLM evaluation benchmarks to guarantee safety and fairness.

Sign language translation has historically been peripheral to mainstream machine translation research. In order to help converge the fields, we introduce FLEURS-ASL, an extension of the multiway parallel benchmarks FLORES (for text) and FLEURS (for speech) to support their first sign language (as video), American Sign Language, translated by 5 Certified Deaf Interpreters. FLEURS-ASL can be used to evaluate a variety of tasks—primarily sentence- and discourse-level translation—between ASL and 200 other languages as text, or 102 languages as speech. We provide baselines for tasks from ASL to English text using a unified modeling approach that incorporates timestamp tokens and previous text tokens in a 34-second context window, trained on random video clips from YouTube-ASL. This model meets or exceeds the performance of phrase-level baselines while supporting a multitude of new tasks. We also use FLEURS-ASL to show that multimodal frontier models have virtually no understanding of ASL, underscoring the importance of including sign languages in standard evaluation suites.

Fact-checking long-form text is challenging, and it is therefore common practice to break it down into multiple atomic claims. The typical approach to fact-checking these atomic claims involves retrieving a fixed number of pieces of evidence, followed by a verification step. However, this method is usually not cost-effective, as it underutilizes the verification model’s internal knowledge of the claim and fails to replicate the iterative reasoning process in human search strategies. To address these limitations, we propose FIRE, a novel agent-based framework that integrates evidence retrieval and claim verification in an iterative manner. Specifically, FIRE employs a unified mechanism to decide whether to provide a final answer or generate a subsequent search query, based on its confidence in the current judgment. We compare FIRE with other strong fact-checking frameworks and find that it achieves slightly better performance while reducing large language model (LLM) costs by an average of 7.6 times and search costs by 16.5 times. These results indicate that FIRE holds promise for application in large-scale fact-checking operations.

The rapid expansion of text data has increased the need for effective methods to distill meaningful information from large datasets. Traditional and state-of-the-art approaches have made significant strides in topic modeling, yet they fall short in generating contextually specific and semantically intuitive topics, particularly in dynamic environments and low-resource languages. Additionally, multi-document summarization systems often struggle with issues like redundancy, scalability, and maintaining readability. We introduce FIDELITY (Fine-grained Interpretable Distillation for Effective Language Insights and Topic Yielding), a hybrid method that combines topic modeling and text summarization to produce fine-grained, semantically rich, and contextually relevant output. FIDELITY enhances dataset accessibility and interpretability, outperforming traditional models in topic diversity, similarity, and in the ability to process new, unseen documents. Additionally, it demonstrates robust multilingual capabilities, effectively handling low-resource languages like Tagalog. This makes FIDELITY a powerful tool for distilling and understanding complex textual data, providing detailed insights while maintaining the necessary granularity for practical applications.

With the widespread consumption of AI-generated content, there has been an increased focus on developing automated tools to verify the factual accuracy of such content. However, prior research and tools developed for fact verification treat it as a binary classification or a linear regression problem. Although this is a useful mechanism as part of automatic guardrails in systems, we argue that such tools lack transparency in the prediction reasoning and diversity in source evidence to provide a trustworthy user experience.We develop FACTS&EVIDENCE—an interactive and transparent tool for user-driven verification of complex text. The tool facilitates the intricate decision-making involved in fact-verification, presenting its users a breakdown of complex input texts to visualize the credibility of individual claims along with explanation of model decisions and attribution to multiple, diverse evidence sources. FACTS&EVIDENCE aims to empower consumers of machine-generated text and give them agency to understand, verify, selectively trust and use such text.

Large Language Models (LLMs) are proficient at retrieving single facts from extended contexts, yet they struggle with tasks requiring the simultaneous retrieval of multiple facts, especially during generation. This paper identifies a novel “lost-in-the-middle” phenomenon, where LLMs progressively lose track of critical information throughout the generation process, resulting in incomplete or inaccurate retrieval. To address this challenge, we introduce Find All Crucial Texts (FACT), an iterative retrieval method that refines context through successive rounds of rewriting. This approach enables models to capture essential facts incrementally, which are often overlooked in single-pass retrieval. Experiments demonstrate that FACT substantially enhances multi-fact retrieval performance across various tasks, though improvements are less notable in general-purpose QA scenarios. Our findings shed light on the limitations of LLMs in multi-fact retrieval and underscore the need for more resilient long-context retrieval strategies.

Alignment of large language models (LLMs) with human values and preferences, often achieved through fine-tuning based on human feedback, is essential for ensuring safe and responsible AI behaviors. However, the process typically requires substantial data and computation resources. Recent studies have revealed that alignment might be attainable at lower costs through simpler methods, such as in-context learning. This leads to the question: Is alignment predominantly superficial? In this paper, we delve into this question and provide a quantitative analysis. We formalize the concept of superficial knowledge, defining it as knowledge that can be acquired through easily token restyling, without affecting the model’s ability to capture underlying causal relationships between tokens. We propose a method to extract and isolate those superficial knowledge from aligned models, focusing on the shallow modifications to the final token selection process. By comparing models augmented only with superficial knowledge to fully aligned models, we quantify the superficial portion of alignment. Our findings reveal that while superficial knowledge constitutes a significant portion of alignment, particularly in safety and detoxification tasks, it is not the whole story. Tasks requiring reasoning and contextual understanding still rely on deeper knowledge. Additionally, we demonstrate two practical advantages of isolated superficial knowledge: (1) it can be transferred between models, enabling efficient offsite alignment of larger models using extracted superficial knowledge from smaller models, and (2) it is recoverable, allowing for the restoration of alignment in compromised models without sacrificing performance.

Extracting temporal relationships between events in the text is crucial for understanding how events unfold over time, especially in the information-dense and precision-demanding military field. Existing models for extracting event temporal relations typically compare the relative times of events directly, neglecting the contextual information between event pairs. This can lead to difficulties in handling uncertain temporal boundaries expressed in text. In this paper, we propose an event temporal relationship extraction model for the military field, based on relative event time prediction and virtual adversarial training, MFRV. The relative event time prediction as an auxiliary task enhances the model’s ability to capture and infer temporal relationships. Virtual adversarial training increases the model’s generalization by generating adversarial samples. Additionally, we adopt the MoCo (Multi-objective gradient correction) method to balance the losses from relative event time prediction and virtual adversarial training, effectively resolving the gradient bias issue in multi-objective optimization. Furthermore, we have constructed a new dataset, TRMF, specifically for event temporal relationship extraction in the military field. Experiments conducted on TRMF, as well as widely used public datasets MATRES and TCR, demonstrate the effectiveness of MFRV.

Large language models (LLMs) have presented significant opportunities to enhance various machine learning applications, including graph neural networks (GNNs). By leveraging the vast open-world knowledge within LLMs, we can more effectively interpret and utilize textual data to better characterize heterophilic graphs, where neighboring nodes often have different labels. However, existing approaches for heterophilic graphs overlook the rich textual data associated with nodes, which could unlock deeper insights into their heterophilic contexts. In this work, we explore the potential of LLMs for modeling heterophilic graphs and propose a novel two-stage framework: LLM-enhanced edge discriminator and LLM-guided edge reweighting. In the first stage, we fine-tune the LLM to better identify homophilic and heterophilic edges based on the textual content of their nodes. In the second stage, we adaptively manage message propagation in GNNs for different edge types based on node features, structures, and heterophilic or homophilic characteristics. To cope with the computational demands when deploying LLMs in practical scenarios, we further explore model distillation techniques to fine-tune smaller, more efficient models that maintain competitive performance. Extensive experiments validate the effectiveness of our framework, demonstrating the feasibility of using LLMs to enhance node classification on heterophilic graphs.

Crowdsourcing has been increasingly utilized to gather subjective assessment, such as evaluating the toxicity of texts. Since there doesnot exist a single “ground truth” answer for subjective annotations, obtaining annotations to accurately reflect the opinions of differentsubgroups becomes a key objective for these subjective assessment tasks. Traditionally, this objective is accomplished by directly soliciting a large number of annotations from each subgroup, which can be costly especially when annotators of certain subgroups are hard to access. In this paper, using toxicity evaluation as an example, we explore the feasibility of using perspective taking—that is, asking annotators to take the point of views of a certain subgroup and estimate opinions within that subgroup—as a way to achieve this objective cost-efficiently. Our results show that compared to the baseline approach of directly soliciting annotations from the target subgroup, perspective taking could lead to better estimates of the subgroup-level opinion when annotations from the target subgroup is costly while the budget is limited. Moreover, prompting annotators to take the perspectives of contrasting subgroups simultaneously can further improve the quality of the estimates. Finally, we find that aggregating multiple perspective-taking annotations while soliciting a small number of annotations directly from the target subgroup for calibration leads to the highest-quality estimates under limited budget.

Large language models (LLMs) are increasingly used in business dialogue systems but they also pose security and ethical risks. Multi-turn conversations, in which context influences the model’s behavior, can be exploited to generate undesired responses. In this paper, we investigate the use of off-the-shelf LLMs in conversational red-teaming settings, where an attacker LLM attempts to elicit undesired outputs from a target LLM. Our experiments address critical questions and offer valuable insights regarding the effectiveness of using LLMs as automated red-teamers, shedding light on key strategies and usage approaches that significantly impact their performance.Our findings demonstrate that off-the-shelf models can serve as effective red-teamers, capable of adapting their attack strategies based on prior attempts. Allowing these models to freely steer conversations and conceal their malicious intent further increases attack success. However, their effectiveness decreases as the alignment of the target model improves.

As large language models (LLMs) become increasingly integrated into daily applications, it is essential to ensure they function fairly across diverse user demographics. In this work, we show that LLMs suffer from personalization bias, where their performance is impacted when they are personalized to a user’s identity. We quantify personalization bias by evaluating the performance of LLMs along two axes - safety and utility. We measure safety by examining how benign LLM responses are to unsafe prompts. We measure utility by evaluating the LLM’s performance on various tasks, including general knowledge, mathematical abilities, programming, and reasoning skills. We find that various LLMs, ranging from open-source models like Llama-3.1 and Mistral to API-based ones like GPT-3.5 and GPT-4o, exhibit significant variance in performance in terms of safety and utility when personalized with different user identities. Finally, we discuss several strategies to mitigate personalization bias and investigate the origin of personalization bias.

Hate speech detection deals with many language variants, slang, slurs, expression modalities, and cultural nuances. This outlines the importance of working with specific corpora, when addressing hate speech within the scope of Natural Language Processing, recently revolutionized by the irruption of Large Language Models. This work presents a brief analysis of the performance of large language models in the detection of Hate Speech for Rioplatense Spanish. We performed classification experiments leveraging chain-of-thought reasoning with ChatGPT 3.5, Mixtral, and Aya, comparing their results with those of a state-of-the-art BERT classifier. These experiments outline that, even if large language models show a lower precision compared to the fine-tuned BERT classifier and, in some cases, they find hard-to-get slurs or colloquialisms, they still are sensitive to highly nuanced cases (particularly, homophobic/transphobic hate speech). We make our code and models publicly available for future research.