NAACL 2025 论文列表

Templatic memes, characterized by a semantic structure adaptable to the creator’s intent, represent a significant yet underexplored area within meme processing literature. With the goal of establishing a new direction for computational meme analysis, here we create a knowledge base composed of more than 5,200 meme templates, information about them, and 54,000 examples of template instances (templatic memes). To investigate the semantic signal of meme templates, we show that we can match memes in datasets to base templates contained in our knowledge base with a distance-based lookup. To demonstrate the power of meme templates, we create TSplit, a method to reorganize datasets, where a template or templatic instance can only appear in either the training or test split. Our re-split datasets enhance general meme knowledge and improve sample efficiency, leading to more robust models. Our examination of meme templates results in state-of-the-art performance for every dataset we consider, paving the way for analysis grounded in templateness.

Recently, sharing key-value (KV) cache across layers has been found effective in efficient inference of large language models (LLMs). To systematically investigate different techniques of cross-layer KV sharing, we propose a unified framework that covers several recent methods and their novel variants. We conduct comprehensive experiments on all the configurations of the framework, evaluating their generation throughput and performance in language modeling and downstream tasks. We find that when reducing the size of the KV cache by 2×, most configurations can achieve higher throughput than standard transformers while maintaining competitive performance.When further reducing the size of the KV cache, however, pairing queries of all layers with KVs of upper layers performs better, at the expense of additional training cost and prefilling latency. We hope that this work will help users make more informed choices of cross-layer KV sharing approaches and facilitate future research on efficient LLM inference.

In this work we systematically investigate how specific attributes of preference datasets affect the alignment and downstream performance of LLMs in instruction-following tasks. We use a novel synthetic data generation pipeline to generate 48,000 unique instruction-following prompts with combinations of 23 verifiable constraints that enable fine-grained and automated quality assessments of model responses. With our synthetic prompts, we use rejection sampling (RS) and Monte Carlo Tree Search (MCTS) to obtain preference pairs. Then, we perform experiments investigating the effects of (1) the presence of shared prefixes between the chosen and rejected responses, (2) the contrast and quality of the chosen, rejected responses and (3) the complexity of the training prompts. Our experiments reveal that shared prefixes provide marginal but consistent improvements and greater stability across challenging training configurations. While high-contrast preference pairs generally outperform low-contrast pairs, combining both often yields the best performance. Additionally, training on prompts of moderate difficulty leads to better generalization across different tasks. Our findings provide actionable insights into optimizing preference data curation for instruction-following tasks, offering a scalable and effective framework for enhancing LLM training and alignment.

In-Context Learning (ICL) empowers Large Language Models (LLMs) with the ability to learn from a few examples provided in the prompt, enabling downstream generalization without the requirement for gradient updates. Despite encouragingly empirical success, the underlying mechanism of ICL remains unclear. Existing research remains ambiguous with various viewpoints, utilizing intuition-driven and ad-hoc technical solutions to interpret ICL. In this paper, we leverage a data generation perspective to reinterpret recent efforts from a systematic angle, demonstrating the potential broader usage of these popular technical solutions. For a conceptual definition, we rigorously adopt the terms of skill recognition and skill learning. Skill recognition selects one learned data generation function previously seen during pre-training while skill learning can learn new data generation functions from in-context data. Furthermore, we provide insights into the strengths and weaknesses of both abilities, emphasizing their commonalities through the perspective of data generation. This analysis suggests potential directions for future research. The corresponding paper list can be found here.

Question Under Discussion (QUD), which is originally a linguistic analytic framework, gains increasing attention in the community of natural language processing over the years. Various models have been proposed for implementing QUD for discourse processing. This survey summarizes these models, with a focus on application to written texts, and examines studies that explore the relationship between QUD and mainstream discourse frameworks, including RST, PDTB and SDRT. Some questions that may require further study are suggested.

Despite the increasing effort in including more low-resource languages in NLP/CL development, most of the world’s languages are still absent. In this paper, we take the example of the Sino-Tibetan language family which consists of hundreds of low-resource languages, and we look at the representation of these low-resource languages in papers archived on ACL Anthology. Our findings indicate that while more techniques and discussions on more languages are present in more publication venues over the years, the overall focus on this language family has been minimal. The lack of attention might be owing to the small number of native speakers and governmental support of these languages. The current development of large language models, albeit successful in a few quintessential rich-resource languages, are still trailing when tackling these low-resource languages. Our paper calls for the attention in NLP/CL research on the inclusion of low-resource languages, especially as increasing resources are poured into the development of data-driven language models.

We introduce SAVIS, a sentence-level attention visualization tool that enhances the interpretability of long documents processed by Large Language Models (LLMs). By computing inter-sentence attention (ISA) through token-level attention aggregation, SAVIS reduces the complexity of attention analysis, enabling users to identify meaningful document-level patterns. The tool offers an interactive interface for exploring how sentences relate to each other in model processing. Our comparative analysis with existing visualization tools demonstrates that SAVIS improves task accuracy and reduces error identification time. We demonstrate its effectiveness for text analysis applications through case studies on various analysis tasks. Our open-source tool is available at https://pypi.org/project/savis with a screencast video at https://youtu.be/fTZZPHA55So.

Recent studies have highlighted the potential of exploiting parallel corpora to enhance multilingual large language models, improving performance in both bilingual tasks, e.g., machine translation, and general-purpose tasks, e.g., text classification. Building upon these findings, our comprehensive study aims to identify the most effective strategies for leveraging parallel corpora. We investigate the impact of parallel corpora quality and quantity, training objectives, and model size on the performance of multilingual large language models enhanced with parallel corpora across diverse languages and tasks. Our analysis reveals several key insights: (i) filtering noisy translations is essential for effectively exploiting parallel corpora, while language identification and short sentence filtering have little effect; (ii) even a corpus with just 10K parallel sentences can yield results comparable to those obtained from much larger datasets; (iii) employing only the machine translation objective yields the best results among various training objectives and their combinations; (iv) larger multilingual language models benefit more from parallel corpora than smaller models. Our study offers valuable insights into the optimal utilization of parallel corpora to enhance multilingual large language models, extending the generalizability of previous findings from limited languages and tasks to a broader range of scenarios.

We describe Belief Tree Propagation (BTProp), a probabilistic framework for LLM hallucination detection. To judge the truth of a statement, BTProp generates a belief tree by recursively expanding the initial statement into a set of logically related claims, then reasoning globally about the relationships between these claims. BTProp works by constructing a probabilistic model of the LM itself: it reasons jointly about logical relationships between claims and relationships between claim probabilities and LM factuality judgments via probabilistic inference in a “hidden Markov tree”. This method improves over state-of-the-art baselines by 3%-9% (evaluated by AUROC and AUC-PR) on multiple hallucination detection benchmarks.

Interventions targeting the representation space of language models (LMs) have emerged as an effective means to influence model behavior. Such methods are employed, for example, to eliminate or alter the encoding of demographic information such as gender within the model’s representations and, in so doing, create a counterfactual representation. However, because the intervention operates within the representation space, understanding precisely what aspects of the text it modifies poses a challenge. In this paper, we give a method to convert representation counterfactuals into string counterfactuals. We demonstrate that this approach enables us to analyze the linguistic alterations corresponding to a given representation space intervention and to interpret the features utilized to encode a specific concept. Moreover, the resulting counterfactuals can be used to mitigate bias in classification through data augmentation.

The proliferation of large language models has raised growing concerns about their misuse, particularly in cases where AI-generated text is falsely attributed to human authors. Machine-generated content detectors claim to effectively identify such text under various conditions and from any language model. This paper critically evaluates these claims by assessing several popular detectors (RADAR, Wild, T5Sentinel, Fast-DetectGPT, PHD, LogRank, Binoculars) on a range of domains, datasets, and models that these detectors have not previously encountered. We employ various prompting strategies to simulate practical adversarial attacks, demonstrating that even moderate efforts can significantly evade detection. We emphasize the importance of the true positive rate at a specific false positive rate (TPR@FPR) metric and demonstrate that these detectors perform poorly in certain settings, with TPR@.01 as low as 0%. Our findings suggest that both trained and zero-shot detectors struggle to maintain high sensitivity while achieving a reasonable true positive rate.

At the forefront of state-of-the-art human alignment methods are preference optimization methods (*PO). Prior research has often concentrated on identifying the best-performing method, typically involving a grid search over hyperparameters, which can be impractical for general practitioners. In this paper, we examine the robustness of existing state-of-the-art methods to varying hyperparameters in a realistic out-of-distribution (OOD) scenario that mirrors real-world applications of human alignment. Our goal is to empirically find the method that increases the likelihood of achieving better results through the lens of various metrics, such as KL divergence and response length. We also introduce LN-DPO, a simple length-normalized version of DPO that is more stable across hyperparameters, effectively reduces the average response length, and improves performance. Our analysis of state-of-the-art reference-free (i.e., SimPO) and reference-dependent (i.e., DPO and LN-DPO) methods reveals that they perform similarly at their peak (i.e., best possible scenario). However, we uncover that the pattern of change in performance greatly varies as we move away from the best possible scenario.

Large language models (LLMs), with demonstrated reasoning abilities across multiple domains, have been largely underexplored fortime-series reasoning (TsR), which is ubiquitous in the real world. In this work, wepropose TimerBed, the first comprehensivetestbed for evaluating LLMs’ TsR performance.Specifically, TimerBed includes stratified reasoning patterns with real-world tasks, diversecombinations of LLMs and reasoning strategies, and various supervised models as comparison anchors. We perform extensive experiments with TimerBed, test multiple current beliefs, and observe the initial failuresof LLMs in TsR, as evidenced by the ineffectiveness of zero shot (ZST) and performancedegradation of few shot in-context learning(ICL). Further, we identify one possible rootcause: the numerical modeling of data. Toaddress this, we propose a prompt-based solution VL-Time, with visualization-modeled dataand language-guided reasoning. Experimental results demonstrate that VL-Time enablesmultimodal LLMs to be non-trivial ZST andpowerful ICL reasoners for time series, achieving about 140% average performance improvement and 99% average token costs reduction.TimerBed and VL-Time are available at https://github.com/AdityaLab/DeepTime/.

Coherence is an essential property of well-written texts, that refers to the way textual units relate to one another. In the era of generative AI, coherence assessment is essential for many NLP tasks such as summarization, long-form question-answering, and more.Current NLP approaches for modeling coherence often rely on a proxy task, specifically, sentence reordering. However, such an approach may not capture the full range of factors contributing to coherence.To remedy this, in this work we employ the formal linguistic definition by Reinhart:1980 of what makes a discourse coherent, consisting of three conditions, cohesion, consistency and relevance, and formalize these conditions as respective computational tasks, which are in turn jointly trained. We evaluate this modeling approach on two human-rated coherence benchmarks: one of automatically-generated stories and one of real-world texts.Our experiments show that jointly training on the proposed tasks leads to better performance on each task compared with task-specific models, and to better performance on assessing coherence overall.Our proposed computational framework thus paves the way for a more advanced, broad-coverage coherence assessment.

Chain-of-Thought (CoT) prompting has been shown to be effective in guiding Large Language Models (LLMs) to decompose complex tasks into multiple intermediate steps, and constructing a rational reasoning chain for inferring answers. However, the linear nature of CoT falls short from enabling LLMs to effectively handle graph structures, which are essential for personalized recommendation tasks that rely on user-item interaction graphs. To bridge this gap, we introduce GollaRec, which leverages a Graph-of-Thought (GoT) prompting technique in a Multi-modal LLM, namely LLaVA, to effectively exploit the complex structure of the interaction graphs. GollaRec enhances the recommendation effectiveness by integrating both visual and textual “thoughts” into a graph-structured prompt, using both item images and descriptions to produce richer multi-modal user/item representations. In our proposed approach, GollaRec leverages text-graph alignment and graph instruction tuning to allow the Multi-modal LLM to capture complex graph structures. In addition, GollaRec leverages a graph adaptor to integrate user-item interactions into the resulting user/item embeddings, therefore effectively adapting the model to the recommendation task. Our extensive experiments on 6 benchmark datasets demonstrate the superiority of our proposed GollaRec model over 12 existing state-of-the-art models in various multi-modal recommendation tasks, including general and multi-domain recommendation tasks.

Existing research on news summarization primarily focuses on single-language single-document (SLSD), single-language multi-document (SLMD) or cross-language single-document (CLSD). However, in real-world scenarios, news about an international event often involves multiple documents in different languages, i.e., mixed-language multi-document (MLMD). Therefore, summarizing MLMD news is of great significance. However, the lack of datasets for MLMD news summarization has constrained the development of research in this area. To fill this gap, we construct a mixed-language multi-document news summarization dataset (MLMD-news), which contains four different languages and 10,992 source document cluster and target summary pairs. Additionally, we propose a graph-based extract-generate model and benchmark various methods on the MLMD-news dataset and publicly release our dataset and code, aiming to advance research in summarization within MLMD scenarios.

Despite the remarkable performance of large language models (LLMs), they still struggle with generating logically sound arguments, resulting in potential risks such as spreading misinformation. An important factor contributing to LLMs’ suboptimal performance in generating coherent arguments is their oversight of logical fallacies. To address this issue, we introduce fallacy-informed preference optimization (FIPO) that helps steer LLMs toward generating logically sound arguments. FIPO includes a classification loss to capture the fine-grained information on fallacy types. Our results on argument generation tasks show that FIPO reduces the fallacy errors by up to 17.5%. Furthermore, our human evaluation results reveal that the quality of the arguments generated by our method significantly outperforms the fine-tuned baselines and other preference optimization methods, such as DPO. These findings highlight the importance of ensuring models are aware of logical fallacies for effective argument generation.

Driver fatigue is a significant factor contributing to road accidents, highlighting the need for reliable and accurate detection methods. In this study, we introduce a novel learning-based multi-frame visual feature framework (LMVFF) designed for precise fatigue detection. Our methodology comprises several clear and interpretable steps. Initially, facial landmarks are detected, enabling the calculation of distances between eyes, lips, and the assessment of head rotation angles based on 68 identified landmarks. Subsequently, visual features from the eye region are extracted, and an effective visual model is developed to accurately classify eye openness. Additionally, features characterizing lip movements are analyzed to detect yawning, thereby enriching fatigue detection through continuous monitoring of eye blink frequency, yawning occurrences, and head movements. Compared to conventional single-feature detection approaches, LMVFF significantly reduces instances of fatigue misidentification. Moreover, we employ various quantization and compression techniques for multiple computation stages, substantially reducing the latency of our system and achieving a real-time frame rate of 25-30 FPS for practical applications.

Distinguishing between extremely similar diseases is a critical and challenging aspect of clinical decision-making. Traditional classification, contrastive learning, and Large Language Models (LLMs) based methods fail to detect the subtle clues necessary for differentiation. This task demands complex reasoning and a variety of tools to identify minor differences and make informed decisions. This paper probes a novel framework that leverages LLMs and a multi-agent system to achieve accurate disease diagnosis through a process of repeated debate and reassessment. The approach aims to identify subtle differences between similar disease candidates. We structure patient information and integrate extensive medical knowledge to guide the analysis towards discerning these differences for precise diagnosis. Comprehensive experiments were conducted on two public datasets and two newly introduced datasets, JarvisD2-Chinese and JarvisD2-English, to validate the effectiveness of our method. The results confirm the efficacy of our approach, demonstrating its potential to enhance diagnostic precision in healthcare.