NAACL 2025 Papers

We study retrieving a set of documents that covers various perspectives on a complex and contentious question (e.g., will ChatGPT do more harm than good?). We curate a Benchmark for Retrieval Diversity for Subjective questions (BERDS), where each example consists of a question and diverse perspectives associated with the question, sourced from survey questions and debate websites. On this data, retrievers paired with a corpus are evaluated to surface a document set that contains diverse perspectives. Our framing diverges from most retrieval tasks in that document relevancy cannot be decided by simple string matches to references. Instead, we build a language model-based automatic evaluator that decides whether each retrieved document contains a perspective. This allows us to evaluate the performance of three different types of corpus (Wikipedia, web snapshot, and corpus constructed on the fly with retrieved pages from the search engine) paired with retrievers. Retrieving diverse documents remains challenging, with the outputs from existing retrievers covering all perspectives on only 33.74% of the examples. We further study the impact of query expansion and diversity-focused reranking approaches and analyze retriever sycophancy. Together, we lay the foundation for future studies in retrieval diversity handling complex queries.

The Open Ko-LLM Leaderboard has been instrumental in benchmarking Korean Large Language Models (LLMs), yet it has certain limitations. Notably, the disconnect between quantitative improvements on the overly academic leaderboard benchmarks and the qualitative impact of the models should be addressed. Furthermore, the benchmark suite is largely composed of translated versions of their English counterparts, which may not fully capture the intricacies of the Korean language. To address these issues, we propose Open Ko-LLM Leaderboard2, an improved version of the earlier Open Ko-LLM Leaderboard. The original benchmarks are entirely replaced with new tasks that are more closely aligned with real-world capabilities. Additionally, four new native Korean benchmarks are introduced to better reflect the distinct characteristics of the Korean language. Through these refinements, Open Ko-LLM Leaderboard2 seeks to provide a more meaningful evaluation for advancing Korean LLMs.

Open domain question answering systems frequently rely on information retrieved from large collections of text (such as the Web) to answer questions. However, such collections of text often contain conflicting information, and indiscriminately depending on this information may result in untruthful and inaccurate answers. To understand the gravity of this problem, we collect a human-annotated dataset, Question Answering with Conflicting Contexts (QACC), and find that as much as 25% of unambiguous, open domain questions can lead to conflicting contexts when retrieved using Google Search. We evaluate and benchmark three powerful Large Language Models (LLMs) with our dataset QACC and demonstrate their limitations in effectively addressing questions with conflicting information. To explore how humans reason through conflicting contexts, we request our annotators to provide explanations for their selections of correct answers. We demonstrate that by finetuning LLMs to explain their answers, we can introduce richer information into their training that guide them through the process of reasoning with conflicting contexts. We publicly release our dataset and code to promote research along this line.

Researchers in social science and psychology have recently proposed using large language models (LLMs) as replacements for humans in behavioral research. In addition to arguments about whether LLMs accurately capture population-level patterns, this has raised questions about whether LLMs capture human-like conceptual diversity. Separately, it is debated whether post-training alignment (RLHF or RLAIF) affects models’ internal diversity. Inspired by human studies, we use a new way of measuring the conceptual diversity of synthetically-generated LLM “populations” by relating the internal variability of simulated individuals to the population-level variability. We use this approach to evaluate non-aligned and aligned LLMs on two domains with rich human behavioral data. While no model reaches human-like diversity, aligned models generally display less diversity than their instruction fine-tuned counterparts. Our findings highlight potential trade-offs between increasing models’ value alignment and decreasing the diversity of their conceptual representations.

Emotion cause analysis is a critical topic in natural language processing. Key tasks include emotion cause extraction (ECE), emotion-cause pair extraction (ECPE), social emotion cause identification (SECI) as well as social emotion mining and its cause identification (SEMCI). While current emotion cause analysis methods often focus on task-specific model design, they tend to overlook the underlying common ground across these tasks rooted in cognitive emotion theories, in particular, the cognitive structure of emotions. Drawing inspiration from this theory, in this paper, we propose a unified model capable of tackling diverse emotion cause analysis tasks, which constructs the emotion cognitive structure through LLM-based in-context learning. To mitigate the hallucination inherent in LLMs, we introduce a self-promote mechanism built on iterative refinement. It dynamically assesses the reliability of substructures based on their cognitive consistency and leverages the more reliable substructures to promote the inconsistent ones. Experimental results on multiple emotion cause analysis tasks ECE, ECPE, SECI and SEMCI demonstrate the superiority of our unified model over existing SOTA methods and LLM-based baselines.

Text sanitization, which employs differential privacy to replace sensitive tokens with new ones, represents a significant technique for privacy protection. Typically, its performance in preserving privacy is evaluated by measuring the attack success rate (ASR) of reconstruction attacks, where attackers attempt to recover the original tokens from the sanitized ones. However, current reconstruction attacks on text sanitization are developed empirically, making it challenging to accurately assess the effectiveness of sanitization. In this paper, we aim to provide a more accurate evaluation of sanitization effectiveness. Inspired by the works of Palamidessi et al., we implement theoretically optimal reconstruction attacks targeting text sanitization. We derive their bounds on ASR as benchmarks for evaluating sanitization performance. For real-world applications, we propose two practical reconstruction attacks based on these theoretical findings. Our experimental results underscore the necessity of reassessing these overlooked risks. Notably, one of our attacks achieves a 46.4% improvement in ASR over the state-of-the-art baseline, with a privacy budget of 𝜖=4.0 on the SST-2 dataset. Our code is available at: https://github.com/mengtong0110/On-the-Vulnerability-of-Text-Sanitization.

Text toxicity detection systems exhibit significant biases, producing disproportionate rates of false positives on samples mentioning demographic groups. But what about toxicity detection in speech? To investigate the extent to which text-based biases are mitigated by speech-based systems, we produce a set of high-quality group annotations for the multilingual MuTOX dataset, and then leverage these annotations to systematically compare speech- and text-based toxicity classifiers. Our findings indicate that access to speech data during inference supports reduced bias against group mentions, particularly for ambiguous and disagreement-inducing samples. Our results also suggest that improving classifiers, rather than transcription pipelines, is more helpful for reducing group bias. We publicly release our annotations and provide recommendations for future toxicity dataset construction.

Argument mining (AM) focuses on analyzing argumentative structures such as Argument Components (ACs) and Argumentative Relations (ARs). Modeling dependencies between ACs and ARs is challenging due to the complex interactions between ACs. Existing approaches often overlook crucial conceptual links, such as key phrases that connect two related ACs, and tend to rely on cartesian product methods to model these dependencies, which can result in class imbalances. To extract key phrases from the AM benchmarks, we employ a prompt-based strategy utilizing an open-source Large Language Model (LLM). Building on this, we propose a unified text-to-text generation framework that leverages Augmented Natural Language (ANL) formatting and integrates the extracted key phrases inside the ANL itself to efficiently solve multiple AM tasks in a joint formulation. Our method sets new State-of-the-Art (SoTA) on three structurally distinct standard AM benchmarks, surpassing baselines by up to 9.5% F1 score, demonstrating its strong potential.

Retrieval-augmented generation (RAG) has emerged as an approach to augment large language models (LLMs) by reducing their reliance on static knowledge and improving answer factuality. RAG retrieves relevant context snippets and generates an answer based on them. Despite its increasing industrial adoption, systematic exploration of RAG components is lacking, particularly regarding the ideal size of provided context, and the choice of base LLM and retrieval method. To help guide development of robust RAG systems, we evaluate various context sizes, BM25 and semantic search as retrievers, and eight base LLMs. Moving away from the usual RAG evaluation with short answers, we explore the more challenging long-form question answering in two domains, where a good answer has to utilize the entire context. Our findings indicate that final QA performance improves steadily with up to 15 snippets but stagnates or declines beyond that. Finally, we show that different general-purpose LLMs excel in the biomedical domain than the encyclopedic one, and that open-domain evidence retrieval in large corpora is challenging.

CodeLLMs are widely used for code generation, yet their ability to handle repository-level dependencies remains underexplored. We introduce RepoExec, a benchmark for evaluating repository-level code generation, focusing on executability, functional correctness, and dependency utilization. Our study evaluates 18 models, revealing that retaining full dependency context yields the best performance, while smaller context sizes can be misleading. Pretrained LLMs excel in correctness but often reimplement dependencies, while instruction-tuned models better utilize dependencies but sometimes introduce unnecessary complexity. We propose an instruction-tuning dataset that improves dependency handling and introduce a new metric, Dependency Invocation Rate (DIR), to measure context utilization. Experiments show that instruction-tuned models improve DIR by over 10%, and multi-round debugging further enhances both correctness and dependency use. RepoExec provides a comprehensive framework to advance CodeLLMs for real-world applications. The dataset and source code are available at https://github.com/FSoft-AI4Code/RepoExec.

The field of text privatization often leverages the notion of *Differential Privacy* (DP) to provide formal guarantees in the rewriting or obfuscation of sensitive textual data. A common and nearly ubiquitous form of DP application necessitates the addition of calibrated noise to vector representations of text, either at the data- or model-level, which is governed by the privacy parameter 𝜀. However, noise addition almost undoubtedly leads to considerable utility loss, thereby highlighting one major drawback of DP in NLP. In this work, we introduce a new sentence infilling privatization technique, and we use this method to explore the effect of noise in DP text rewriting. We empirically demonstrate that non-DP privatization techniques excel in utility preservation and can find an acceptable empirical privacy-utility trade-off, yet cannot outperform DP methods in empirical privacy protections. Our results highlight the significant impact of noise in current DP rewriting mechanisms, leading to a discussion of the merits and challenges of DP in NLP as well as the opportunities that non-DP methods present.

Large language models have emerged as powerful tools for general intelligence, showcasing advanced natural language processing capabilities that find applications across diverse domains. Despite their impressive performance, recent studies have highlighted the potential for significant enhancements in LLMs’ task-specific performance through fine-tuning strategies like Reinforcement Learning with Human Feedback (RLHF), supervised fine-tuning (SFT), and Quantized Low-Rank Adapters (Q-LoRA) method. However, previous works have shown that while fine-tuning offers significant performance gains, it also leads to challenges such as catastrophic forgetting and privacy and safety risks. To this end, there has been little to no work in *understanding the impact of fine-tuning on the reasoning capabilities of LLMs*. Our research investigates the effect of fine-tuning on the reasoning abilities of LLMs, addressing critical questions regarding the impact of task-specific fine-tuning on overall reasoning capabilities, the influence of fine-tuning on Chain-of-Thought (CoT) reasoning performance, and the implications for the faithfulness of CoT reasonings. By exploring these dimensions, our study shows the impact of fine-tuning on LLM reasoning capabilities, where the faithfulness of CoT reasoning, on average across four datasets, decreases, highlighting potential shifts in internal mechanisms of the LLMs resulting from fine-tuning processes.

Data attribution methods are used to measure the contribution of training data towards model outputs, and have several important applications in areas such as dataset curation and model interpretability. However, many standard data attribution methods, such as influence functions, utilize model gradients and are computationally expensive. In our paper, we show in-context probing (ICP) – prompting a LLM – can serve as a fast proxy for gradient-based data attribution for data selection under conditions contingent on data similarity. We study this connection empirically on standard NLP tasks, and show that ICP and gradient-based data attribution are well-correlated in identifying influential training data for tasks that share similar task type and content as the training data. Additionally, fine-tuning models on influential data selected by both methods achieves comparable downstream performance, further emphasizing their similarities. We then examine the connection between ICP and gradient-based data attribution using synthetic data on linear regression tasks. Our synthetic data experiments show similar results with those from NLP tasks, suggesting that this connection can be isolated in simpler settings, which offers a pathway to bridging their differences.

A small subset of dimensions within language Transformers’ representation spaces emerge as “outliers” during pretraining, encoding critical knowledge sparsely. We extend previous findings on emergent outliers to Encoder-Decoder Transformers and instruction-finetuned models, and tackle the problem of distilling a student Transformer from a larger teacher Transformer. Knowledge distillation reduces model size and cost by transferring knowledge from a larger teacher to a smaller student, necessitating a trade-off among representation dimensions. We show that emergent outlier dimensions contribute significantly more to zero-shot performance than non-outlier dimensions. Based on this, we propose the Emergent Outlier Focused Distillation (EOFD) method, which prioritizes critical outlier dimensions in distillation using a weighted MSE loss. We empirically demonstrate that EOFD outperforms state-of-the-art distillation methods and generalizes well across Encoder-only BERT, Decoder-only GPT-2, and Encoder-Decoder T5 architectures.

While natural language processing (NLP) techniques have been applied to user reviews in recommendation systems, the potential of leveraging Arabic dialects in this context remains unexplored. Arabic is spoken by over 420 million people, with significant dialectal variation across regions. These dialects, often classified as low-resource languages, present both challenges and opportunities for machine learning applications. This paper represents the first attempt to incorporate Arabic dialects as a signal in recommendation systems. We explore both explicit and implicit approaches for integrating Arabic dialect information from user reviews, demonstrating its impact on improving recommendation performance. Our findings highlight the potential for leveraging dialectal diversity in Arabic to enhance recommendation systems and encourage further research at the intersection of NLP and recommendation systems within the Arab multicultural world.

Language Models (LMs) have been shown to exhibit a strong preference towards entities associated with Western culture when operating in non-Western languages. In this paper, we aim to uncover the origins of entity-related cultural biases in LMs by analyzing several contributing factors, including the representation of entities in pre-training data and the impact of variations in linguistic phenomena across languages. We introduce CAMeL-2, a parallel Arabic-English benchmark of 58,086 entities associated with Arab and Western cultures and 367 masked natural contexts for entities. Our evaluations using CAMeL-2 reveal reduced performance gaps between cultures by LMs when tested in English compared to Arabic. We find that LMs struggle in Arabic with entities that appear at high frequencies in pre-training, where entities can hold multiple word senses. This also extends to entities that exhibit high lexical overlap with languages that are not Arabic but use the Arabic script. Further, we show how frequency-based tokenization leads to this issue in LMs, which gets worse with larger Arabic vocabularies. We will make CAMeL-2 available at: https://github.com/tareknaous/camel2

We revisit the reference determinacy (RD) assumption in the task of natural language inference (NLI), i.e., the premise and hypothesis are assumed to refer to the same context when human raters annotate a label. While RD is a practical assumption for constructing a new NLI dataset, we observe that current NLI models—which are typically trained solely on hypothesis-premise pairs created with the RD assumption—fail in downstream applications such as fact verification, where the input premise and hypothesis may refer to different contexts. To highlight the impact of this phenomenon in real-world use cases, we introduce RefNLI, a diagnostic benchmark for identifying reference ambiguity in NLI examples. In RefNLI, the premise is retrieved from a knowledge source (i.e. Wikipedia) and does not necessarily refer to the same context as the hypothesis. With RefNLI, we demonstrate that finetuned NLI models and few-shot prompted LLMs both fail to recognize context mismatch, leading to >80% false contradiction and >50% entailment predictions. We discover that the existence of reference ambiguity in NLI examples can in part explain the inherent human disagreements in NLI, and provide insight into how the RD assumption impacts NLI dataset creation process.

Large Language Models (LLMs) often exhibit positional bias in long-context settings, under-attending to information in the middle of inputs. We investigate the presence of this bias in long-form summarization, its impact on faithfulness, and various techniques to mitigate this bias. To consistently evaluate faithfulness, we first compile a benchmark of eight human-annotated long-form summarization datasets and perform a meta-evaluation of faithfulness metrics. We show that LLM-based faithfulness metrics, though effective with full-context inputs, remain sensitive to document order, indicating positional bias. Analyzing LLM-generated summaries across six datasets, we find a “U-shaped” trend in faithfulness, where LLMs faithfully summarize the beginning and end of documents but neglect middle content. Perturbing document order similarly reveals models are less faithful when important documents are placed in the middle of the input. We find that this behavior is partly due to shifting focus with context length: as context increases, summaries become less faithful, but beyond a certain length, faithfulness improves as the model focuses on the end. Finally, we experiment with different generation techniques to reduce positional bias and find that prompting techniques effectively direct model attention to specific positions, whereas more sophisticated approaches offer limited improvements. Our data and code will be publicly available.

Warning: This paper contains offensive language.Ensuring that large language models (LLMs) do not generate harmful text is critical for their safe deployment. A common failure mode involves producing toxic responses to otherwise innocuous prompts. While various detoxification methods have been proposed, the underlying mechanisms that drive toxic generation in LLMs are not yet fully understood. Our work aims to provide a mechanistic understanding of toxic generation against innocuous-seeming adversarial prompts through the lens of memory localization. We find evidence of localization of toxic memories in the early Multilayer Perceptron (MLP) layers of GPT-2-XL. We further investigate the effects of editing and deleting these toxic memories in MLP layers to reduce toxic generation. Editing significantly reduces toxic generation, from 62.86% to 28.61%. However, this reduction comes with a trade-off in generation quality as perplexity increases from 78.18 on GPT2-XL against the adversarial prompts to 106.06 after editing. Localization-informed deletion achieves a better toxicity-perplexity tradeoff compared to random early layer editing, which reduces toxicity but leads to greater perplexity increases.