NAACL 2025 论文列表

The distinction between function and content units of the lexicon has been somewhat neglected in recent NLP work, but it could still be useful when working with low-resource languages, and, in particular, to improve cross-lingual transfer. In this paper, we investigate to what extent BPE subword tokenization can be used to identify units of the functional lexicon in a language without any annotated data. We analyze subword tokens in terms of their productivity and attempt to find thresholds that best distinguish function from content tokens. On a sample of seven diverse languages, we find that the best results are obtained with 50 BPE merges. We also show that this subword tokenization setting can be beneficial for the interlinear glossing task.

The manipulation of the personality traits of large language models (LLMs) has emerged as a key area of research. Methods like prompt-based In-Context Knowledge Editing (IKE) and gradient-based Model Editor Networks (MEND) have been explored but show irregularity and variability; IKE depends on the prompt, leading to variability and sensitivity, while MEND yields inconsistent and gibberish outputs. To address this, we employed Opinion QA Based Parameter-Efficient Fine-Tuning (PEFT), specifically Quantized Low-Rank Adaptation (QLoRA), to manipulate the Big Five personality traits: Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism. After PEFT, models such as Mistral-7B-Instruct and LLaMA-2-7B-chat showed a latent behaviour by generating emojis for certain traits, despite no emojis being present in the PEFT data. For instance, LLaMA-2-7B-chat generated emojis in 99.5% of extraversion-related test instances, while Mistral-7B-Instruct did so in 92.5% of openness-related test instances. ICL Explainability analysis indicated that the LLMs used emojis intentionally to express these traits. Mechanistic Interpretability analysis showed that this latent behaviour of LLMs could be traced to specific neurons that became activated or amplified after PEFT. This paper provides a number of novel contributions. First, introducing an Opinion QA dataset for PEFT-driven personality manipulation; second, developing metric models to benchmark LLM personality traits; third, demonstrating PEFT’s superiority over IKE in personality manipulation; and finally, analysing and validating emoji usage through explainability methods such as Mechanistic Interpretability and In-context learning Explainability methods.

Although many studies have investigated and reduced hallucinations in large language models (LLMs) for single-document tasks, research on hallucination in multi-document summarization (MDS) tasks remains largely unexplored. Specifically, it is unclear how the challenges arising from handling multiple documents (e.g., repetition and diversity of information) affect models outputs. In this work, we investigate how hallucinations manifest in LLMs when summarizing topic-specific information from a set of documents. Since no benchmarks exist for investigating hallucinations in MDS, we leverage existing news and conversation datasets, annotated with topic-specific insights, to create two novel multi-document benchmarks. When evaluating 5 LLMs on our benchmarks, we observe that on average, up to 75% of the content in LLM-generated summary is hallucinated, with hallucinations more likely to occur towards the end of the summaries. Moreover, when summarizing non-existent topic-related information, GPT-3.5-turbo and GPT-4o still generate summaries about 79.45% and 44% of the time, raising concerns about their tendency to fabricate content. To better understand the characteristics of these hallucinations, we conduct a human evaluation of 700+ insights and discover that most errors stem from either failing to follow instructions or producing overly generic insights. Motivated by these observations, we investigate the efficacy of simple post-hoc baselines in mitigating hallucinations but find them only moderately effective. Our results underscore the need for more effective approaches that systematically mitigate hallucinations in MDS.

Large Vision Language Models (LVLMs) achieve great performance on visual-language reasoning tasks, however, the black-box nature of LVLMs hinders in-depth research on the reasoning mechanism. As all images need to be converted into image tokens to fit the input format of large language models (LLMs) along with natural language prompts, sequential visual representation is essential to the performance of LVLMs, and the information flow analysis approach can be an effective tool for determining interactions between these representations. In this paper, we propose integrating attention analysis with LLaVA-CAM, concretely, attention scores highlight relevant regions during forward propagation, while LLaVA-CAM captures gradient changes through backward propagation, revealing key image features. By exploring the information flow from the perspective of visual representation contribution, we observe that it tends to converge in shallow layers but diversify in deeper layers. To validate our analysis, we conduct comprehensive experiments with truncation strategies across various LVLMs for visual question answering and image captioning tasks, and experimental results not only verify our hypothesis but also reveal a consistent pattern of information flow convergence in the corresponding layers, and the information flow cliff layer will be different due to different contexts.

Motivated by in-context learning (ICL) capabilities of Large Language Models (LLMs), multimodal LLMs with additional visual modality are also exhibited with similar ICL abilities when multiple image-text pairs are provided as demonstrations. However, relatively less work has been done to investigate the principles behind how and why multimodal ICL works. We conduct a systematic and principled evaluation of multimodal ICL for models of different scales on a broad spectrum of new yet critical tasks. Through perturbations over different modality information, we show that modalities matter differently across tasks in multimodal ICL. Guided by task-specific modality impact, we recommend modality-driven demonstration strategies to boost ICL performance. We also find that models may follow inductive biases from multimodal ICL even if they are rarely seen in or contradict semantic priors from pretraining data. Our principled analysis provides a comprehensive way of understanding the role of demonstrations in multimodal in-context learning, and sheds light on effectively improving multimodal ICL on a wide range of tasks.

With the rapid growth of Large Language Models (LLMs), safeguarding textual content against unauthorized use is crucial. Watermarking offers a vital solution, protecting both - LLM-generated and plain text sources. This paper presents a unified overview of different perspectives behind designing watermarking techniques through a comprehensive survey of the research literature. Our work has two key advantages: (1) We analyze research based on the specific intentions behind different watermarking techniques, evaluation datasets used, and watermarking addition and removal methods to construct a cohesive taxonomy. (2) We highlight the gaps and open challenges in text watermarking to promote research protecting text authorship. This extensive coverage and detailed analysis sets our work apart, outlining the evolving landscape of text watermarking in Language Models.

Application of LLMs for complex causal question answering can be stymied by their opacity and propensity for hallucination. Although recent approaches such as Retrieval Augmented Generation and Chain of Thought prompting have improved reliability, we argue current approaches are insufficient and further fail to satisfy key criteria humans use to select and evaluate causal explanations. Inspired by findings from the social sciences, we present an implemented causal QA approach that combines iterative RAG with guidance from a formal model of causation. Our causal model is backed by the Cogent reasoning engine, allowing users to interactively perform counterfactual analysis and refine their answer. Our approach has been integrated into a deployed Collaborative Research Assistant (Cora) and we present a pilot evaluation in the life sciences domain.

This paper investigates the knowledge of language models from the perspective of Bayesian epistemology. We explore how language models adjust their confidence and responses when presented with evidence with varying levels of informativeness and reliability. To study these properties, we create a dataset with various types of evidence and analyze language models’ responses and confidence using verbalized confidence, token probability, and sampling. We observed that language models do not consistently follow Bayesian epistemology: language models follow the Bayesian confirmation assumption well with true evidence but fail to adhere to other Bayesian assumptions when encountering different evidence types. Also, we demonstrated that language models can exhibit high confidence when given strong evidence, but this does not always guarantee high accuracy. Our analysis also reveals that language models are biased toward golden evidence and show varying performance depending on the degree of irrelevance, helping explain why they deviate from Bayesian assumptions.

The alignment process changes several properties of a large language model’s (LLM’s) output distribution. We analyze two aspects of post-alignment distributional shift of LLM responses. First, we re-examine previously reported reductions in response diversity post-alignment. Our analysis suggests that an apparent drop in the diversity of responses is largely explained by quality control and information aggregation. Alignment suppresses irrelevant and unhelpful content while shifting the output distribution toward longer responses that cover information spanning several responses from the base LLM, essentially presenting diverse information in a single response. Finding little evidence that alignment suppresses useful information, it is natural to ask the opposite question: do aligned models surface information that cannot be recovered from base models? Our second investigation shows this is not the case and the behavior of aligned models is recoverable from base models without fine-tuning. A combination of in-context examples and lower-resolution semantic hints about response content can elicit responses from base LLMs that are as similar to alignment-tuned LLM responses as alignment-tuned LLM responses are to each other. Taken together, these results indicate that current alignment techniques capture but do not extend the useful subset of assistant-like base LLM behavior, providing further evidence for the Superficial Alignment Hypothesis. They also show that in-context alignment can go surprisingly far as a strategy for imitating aligned LLMs without fine-tuning. Our code and data is available at [github.com/thomlake/investigating-alignment](https://github.com/thomlake/investigating-alignment).

Humans attempt to understand the real world by asking the fundamental question ”Why?” when faced with incomprehensible situations in everyday life. Such why-questions provide essential knowledge that can help in understanding these situations. In this study, we conducted an end-to-end process to verify the utility of consecutive why-questions, from constructing a large language model (LLM)-based dataset to performing quantitative evaluation and analysis. Firstly, we created a WHY-Chain dataset, consisting of answers generated by an LLM in response to chain-of-why-questions, including a validity check. We also incorporated objectives that effectively capture the ”consecutive” characteristic of the data. Using the WHY-Chain dataset and two types of self-supervised objectives, we trained the pre-trained model. As a result, the refined model demonstrated improved performance on downstream tasks that require commonsense reasoning. Additionally, we conducted various ablation studies to assess the impact of different factors, confirming the scalability of the proposed approach. Lastly, we confirmed the consistency of the logical information by reasoning chain analysis of the answers generated from consecutive why-questions.

Debating over conflicting issues is a necessary first step towards resolving conflicts. However, intrinsic perspectives of an arguer are difficult to overcome by persuasive argumentation skills. Proceeding from a debate to a deliberative process, where we can identify actionable options for resolving a conflict requires a deeper analysis of arguments and the perspectives they are grounded in - as it is only from there that one can derive mutually agreeable resolution steps. In this work we develop a framework for a deliberative analysis of arguments in a computational argumentation setup. We conduct a fine-grained analysis of perspectivized stances expressed in the arguments of different arguers or stakeholders on a given issue, aiming not only to identify their opposing views, but also shared perspectives arising from their attitudes, values or needs. We formalize this analysis in Perspectivized Stance Vectors that characterize the individual perspectivized stances of all arguers on a given issue. We construct these vectors by determining issue- and argument-specific concepts, and predict an arguer’s stance relative to each of them. The vectors allow us to measure a modulated (dis)agreement between arguers, structured by perspectives, which allows us to identify actionable points for conflict resolution, as a first step towards deliberation.

Tool-calling has changed Large Language Model (LLM) applications by integrating external tools, significantly enhancing their functionality across diverse tasks. However, this integration also introduces new security vulnerabilities, particularly in the tool scheduling mechanisms of LLM, which have not been extensively studied. To fill this gap, we present ToolCommander, a novel framework designed to exploit vulnerabilities in LLM tool-calling systems through adversarial tool injection. Our framework employs a well-designed two-stage attack strategy. Firstly, it injects malicious tools to collect user queries, then dynamically updates the injected tools based on the stolen information to enhance subsequent attacks. These stages enable ToolCommander to execute privacy theft, launch denial-of-service attacks, and even manipulate business competition by triggering unscheduled tool-calling. Notably, the ASR reaches 91.67% for privacy theft and hits 100% for denial-of-service and unscheduled tool calling in certain cases. Our work demonstrates that these vulnerabilities can lead to severe consequences beyond simple misuse of tool-calling systems, underscoring the urgent need for robust defensive strategies to secure LLM Tool-calling systems.

Chain-of-thought (CoT) and subsequent methods adopted a deductive paradigm that decomposes the reasoning process, demonstrating remarkable performances across NLP tasks. However, such a paradigm faces the challenge of getting bogged down in low-level semantic details, hindering large language models (LLMs) from correctly understanding, selecting, and compositing conditions. In this work, we present Overarching Prompting (OaP), a simple prompting method that elicits the high-level thinking of LLMs. Specifically, OaP first abstracts the whole problem into a simplified archetype and formulates strategies grounded in concepts and principles, establishing an overarching perspective for guiding reasoning. We conducted experiments with SoTA models, including ChatGPT, InstructGPT, and Llama3-70B-instruct, and received promising performances across tasks including Knowledge QA, Mathematical, and Open-Domain Reasoning. For instance, OaP improved ChatGPT and CoT by 19.0% and 3.1% on MMLU’s College Physics, 8.8% and 2.3% on GSM8k, and 10.3% and 2.5% on StrategyQA, respectively.

Modern Language Models (LMs) are capable of following long and complex instructions that enable a large and diverse set of user requests. While Information Retrieval (IR) models use these LMs as the backbone of their architectures, virtually none of them allow users to provide detailed instructions alongside queries, thus limiting their ability to satisfy complex information needs. In this work, we study the use of instructions in IR systems. First, we introduce our dataset FollowIR, which contains a rigorous instruction evaluation benchmark as well as a training set for helping IR models learn to better follow real-world instructions. FollowIR repurposes detailed instructions – also known as narratives – developed for professional assessors to evaluate retrieval systems. In particular, we build our benchmark from three collections curated for shared tasks at the Text REtrieval Conference (TREC). These collections contains hundreds to thousands of labeled documents per query, making them suitable for our exploration. Through this process, we can measure how well IR models follow instructions, through a new pairwise evaluation framework. Our results indicate that existing retrieval models fail to correctly use instructions, using them for basic keywords and struggling to understand long-form information. However, we show that it is possible for IR models to learn to follow complex instructions: our new FollowIR-7B model has significant improvements after fine-tuning on our training set.

Vision and language navigation (VLN) is a challenging task towards the creation of embodied agents that requires spatial and temporal reasoning over the instructions provided in natural language and aligning them with the visual perception of an environment. Although a number of methods and approaches have been developed, none achieves human level performance in outdoor settings (by up to 75 percent). The contributions of visual and language modalities to the success of VLN have been studied, however here we focus on an overlooked property of routes and show that navigational instructions can be represented as patterns of actions that also describe trajectory shapes. Through carefully crafted experiments, we show that agents generalization to unseen environments depends not only on visual and linguistic features, but also on the shape of trajectories presented to the model during the fine-tuning. Our experiments show that the diversity of patterns of actions during training is a key contributor to high success rates for agents. Last, we propose a solution based on data augmentation that fills the gap in missing patterns of training data. Our findings will guide researchers towards improved practices in the development and evaluation of VLN datasets and agents.

The rapid proliferation of large language models (LLMs) in natural language processing (NLP) has created a critical need for techniques that enable efficient deployment on memory-constrained devices without compromising performance. We present a method to prune LLMs that selectively prunes model blocks based on an importance score and replaces them with a low-parameter replacement strategy. Specifically, we propose a principled metric to replace each pruned block using a weight-sharing mechanism that leverages unpruned counterparts from the model and block-specific low-rank adapters. Furthermore, we facilitate the learning of these replacement blocks with output feature normalization and an adapter initialization scheme built on low-rank SVD reconstructions. Empirical evaluations demonstrate substantial performance gains over existing methods, achieving state-of-the-art performance on 5/6 benchmarks for a compression rate of 30% and 6/6 benchmarks for a compression rate of 40%. We also demonstrate that our approach can extend smaller models, boosting performance on 6/6 benchmarks using only ~0.3% tokens of extended training with minimal additional parameter costs.

Since the release of ChatGPT, large language models (LLMs) have demonstrated remarkable capabilities across various domains. A key challenge in developing these general capabilities is efficiently sourcing diverse, high-quality data. This becomes especially critical in reasoning-related tasks with sandbox checkers, such as math or code, where the goal is to generate correct solutions to specific problems with higher probability. In this work, we introduce Flaming-hot Initiation with Regular Execution (FIRE) sampling, a simple yet highly effective method to efficiently find good responses. Our empirical findings show that FIRE sampling enhances inference-time generation quality and also benefits training in the alignment stage. Furthermore, we explore how FIRE sampling improves performance by promoting diversity and analyze the impact of employing FIRE at different positions within a response.

Fingerspelling poses challenges for sign language processing due to its high-frequency motion and use for open-vocabulary terms. While prior work has studied fingerspelling recognition, there has been little attention to evaluating how well sign language translation models understand fingerspelling in the context of entire sentences—and improving this capability. We manually annotate instances of fingerspelling within FLEURS-ASL and use them to evaluate the effect of two simple measures to improve fingerspelling recognition within American Sign Language to English translation: 1) use a model family (ByT5) with character- rather than subword-level tokenization, and 2) mix fingerspelling recognition data into the translation training mixture. We find that 1) substantially improves understanding of fingerspelling (and translation quality overall), but the effect of 2) is mixed.

We introduce FAINA, the first dataset for fallacy detection that embraces multiple plausible answers and natural disagreement. FAINA includes over 11K span-level annotations with overlaps across 20 fallacy types on social media posts in Italian about migration, climate change, and public health given by two expert annotators. Through an extensive annotation study that allowed discussion over multiple rounds, we minimize annotation errors whilst keeping signals of human label variation. Moreover, we devise a framework that goes beyond “single ground truth” evaluation and simultaneously accounts for multiple (equally reliable) test sets and the peculiarities of the task, i.e., partial span matches, overlaps, and the varying severity of labeling errors. Our experiments across four fallacy detection setups show that multi-task and multi-label transformer-based approaches are strong baselines across all settings. We release our data, code, and annotation guidelines to foster research on fallacy detection and human label variation more broadly.