NAACL 2025 论文列表

Large language models (LLMs) hold revolutionary potential to digitize and enhance the Health & Public Services (H&PS) industry. Despite their advanced linguistic abilities, concerns about accuracy, stability, and traceability still persist, especially in high-stakes areas such as transportation systems. Moreover, the predominance of English in LLM development raises questions about how they perform in non-English contexts. This study originated from a real world industrial GenAI application, introduces a novel cross-lingual benchmark dataset comprising nearly 99,869 real traffic incident records from Vienna (2013-2023) to assess the robustness of state-of-the-art LLMs (≥ 9) in the spatio vs temporal domain for traffic incident classification. We then explored three hypotheses — sentence indexing, date-to-text conversion, and German-to-English translation — and incorporated Retrieval Augmented Generation (RAG) to further examine the LLM hallucinations in both spatial and temporal domain. Our experiments reveal significant performance disparities in the spatio-temporal domain and demonstrate what types of hallucinations that RAG can mitigate and how it achieves this. We also provide open access to our H&PS traffic incident dataset, with the project demo and code available at Website https://sites.google.com/view/llmhallucination/home

**This paper discusses and contains offensive content.** Language models (LMs) are used in decision-making systems and as interactive assistants. However, how well do these models making judgements align with the diversity of human values, particularly regarding social and moral norms? In this work, we investigate how inclusively LMs perceive norms across demographic groups (e.g., gender, age, and income). We prompt 11 LMs on rules-of-thumb (RoTs) and compare their outputs with the existing responses of 100 human annotators. We introduce the Absolute Distance Alignment Metric (ADA-Met) to quantify alignment on ordinal questions. We find notable disparities in LM responses, with younger, higher-income groups showing closer alignment, raising concerns about the representation of marginalized perspectives. Our findings highlight the importance of further efforts to make LMs more inclusive of diverse human values. The code and prompts are available on GitHub under the CC BY-NC 4.0 license.

Recent research has focused on literary machine translation (MT) as a new challenge in MT. However, the evaluation of literary MT remains an open problem. We contribute to this ongoing discussion by introducing LITEVAL-CORPUS, a paragraph-level parallel corpus containing verified human translations and outputs from 9 MT systems, which totals over 2k translations and 13k evaluated sentences across four language pairs, costing 4.5k€. This corpus enables us to (i) examine the consistency and adequacy of human evaluation schemes with various degrees of complexity, (ii) compare evaluations by students and professionals, assess the effectiveness of (iii) LLM-based metrics and (iv) LLMs themselves. Our findings indicate that the adequacy of human evaluation is controlled by two factors: the complexity of the evaluation scheme (more complex is less adequate) and the expertise of evaluators (higher expertise yields more adequate evaluations). For instance, MQM (Multidimensional Quality Metrics), a complex scheme and the de facto standard for non-literary human MT evaluation, is largely inadequate for literary translation evaluation: with student evaluators, nearly 60% of human translations are misjudged as indistinguishable or inferior to machine translations. In contrast, BWS (BEST-WORST SCALING), a much simpler scheme, identifies human translations at a rate of 80-100%. Automatic metrics fare dramatically worse, with rates of at most 20%. Our overall evaluation indicates that published human translations consistently outperform LLM translations, where even the most recent LLMs tend to produce considerably more literal and less diverse translations compared to humans.

This paper studies the problem of dynamical system modeling, which involves the evolution of multiple interacting objects. Recent data-driven methods often utilize graph neural networks (GNNs) to learn these interactions by optimizing the neural network in an end-to-end fashion. While large language models (LLMs) have shown exceptional zero-shot performance across various applications, their potential for modeling dynamical systems has not been extensively explored. In this work, we design prompting techniques for dynamical system modeling and systematically evaluate the capabilities of LLMs on two tasks, including dynamic forecasting and relational reasoning. An extensive benchmark LLM4DS across nine datasets is built for performance comparison. Our extensive experiments yield several key findings: (1) LLMs demonstrate competitive performance without training compared to state-of-the-art methods in dynamical system modeling. (2) LLMs effectively infer complex interactions among objects to capture system evolution. (3) Prompt engineering plays a crucial role in enabling LLMs to accurately understand and predict the evolution of systems.

Recent advancements in Large Language Models (LLMs) have positioned them as powerful tools for clinical decision-making, with rapidly expanding applications in healthcare. However, concerns about bias remain a significant challenge in the clinical implementation of LLMs, particularly regarding gender and ethnicity. This research investigates the evaluation and mitigation of bias in LLMs applied to complex clinical cases, focusing on gender and ethnicity biases. We introduce a novel Counterfactual Patient Variations (CPV) dataset derived from the JAMA Clinical ChallengeUsing this dataset, we built a framework for bias evaluation, employing both Multiple Choice Questions (MCQs) and corresponding explanations. We explore prompting with eight LLMs and fine-tuning as debiasing methods. Our findings reveal that addressing social biases in LLMs requires a multidimensional approach as mitigating gender bias can occur while introducing ethnicity biases, and that gender bias in LLM embeddings varies significantly across medical specialities. We demonstrate that evaluating both MCQ response and explanation processes is crucial, as correct responses can be based on biased reasoning. We provide a framework for evaluating LLM bias in real-world clinical cases, offer insights into the complex nature of bias in these models, and present strategies for bias mitigation.

The widespread application of Large Language Models (LLMs) across various tasks and fields has necessitated the alignment of these models with human values and preferences. Given various approaches of human value alignment, such as Reinforcement Learning with Human Feedback (RLHF), constitutional learning, and safety fine-tuning etc., there is an urgent need to understand the scope and nature of human values injected into these LLMs before their deployment and adoption. We propose UniVar, a high-dimensional neural representation of symbolic human value distributions in LLMs, orthogonal to model architecture and training data. This is a continuous and scalable representation, self-supervised from the value-relevant output of 8 LLMs and evaluated on 15 open-source and commercial LLMs. Through UniVar, we visualize and explore how LLMs prioritize different values in 25 languages and cultures, shedding light on the complex interplay between human values and language modeling.

Speculative Decoding (SD) utilizes an efficient draft model to generate multiple tokens, which are subsequently verified in parallel by a target model. This approach has shown significant potential for accelerating inference in large language models (LLMs), with performance heavily reliant on the hyperparameter K—the window size. However, previous methods often depend on simple heuristics to select K or dynamically adjust the window size, which may necessitate additional training or careful resource management to avoid competition.To address these challenges, we propose Hierarchical Speculative Decoding with Dynamic Window (HSDDW), a straightforward framework that eliminates the need for additional training. Specifically, we introduce a self-verify mechanism that enables the draft model to autonomously decide when to stop generating tokens. Additionally, by integrating a hierarchical structure that leverages the capabilities of models of different sizes, we significantly enhance the overall speed of the system.HSDDW demonstrates competitive performance across four datasets, achieving notable speedups of 2.91× on MT-Bench and 2.99× on Alpaca, outperforming existing state-of-the-art methods.

Due to the scarcity of agent-oriented pre-training data, LLM-based autonomous agents typically rely on complex prompting or extensive fine-tuning, which often fails to introduce new capabilities while preserving strong generalizability. We introduce Hephaestus-Forge, the first large-scale pre-training corpus designed to enhance the fundamental capabilities of LLM agents in API function calling, intrinsic reasoning and planning, and adapting to environmental feedback. Hephaestus-Forge comprises 103B agent-specific data encompassing 76,537 APIs, including both tool documentation to introduce knowledge of API functions and function calling trajectories to strengthen intrinsic reasoning. To explore effective training protocols, we investigate scaling laws to identify the optimal recipe in data mixing ratios. By continual pre-training on Hephaestus-Forge, Hephaestus outperforms small- to medium-scale open-source LLMs and rivals commercial LLMs on three agent benchmarks, demonstrating the effectiveness of our pre-training corpus in enhancing fundamental agentic capabilities and generalization of LLMs to new tasks or environments.

Open-domain dialogue systems have seen remarkable advancements with the development of large language models (LLMs). Nonetheless, most existing dialogue systems predominantly focus on brief single-session interactions, neglecting the real-world demands for long-term companionship and personalized interactions with chatbots. Crucial to addressing this real-world need are event summary and persona management, which enable reasoning for appropriate long-term dialogue responses. Recent progress in the human-like cognitive and reasoning capabilities of LLMs suggests that LLM-based agents could significantly enhance automated perception, decision-making, and problem-solving. In response to this potential, we introduce a model-agnostic framework, the Long-term Dialogue Agent (LD-Agent), which incorporates three independently tunable modules dedicated to event perception, persona extraction, and response generation. For the event memory module, long and short-term memory banks are employed to separately focus on historical and ongoing sessions, while a topic-based retrieval mechanism is introduced to enhance the accuracy of memory retrieval. Furthermore, the persona module conducts dynamic persona modeling for both users and agents. The integration of retrieved memories and extracted personas is subsequently fed into the generator to induce appropriate responses. The effectiveness, generality, and cross-domain capabilities of LD-Agent are empirically demonstrated across various illustrative benchmarks, models, and tasks. The code is released at https://github.com/leolee99/LD-Agent.

Existing household robots have made significant progress in performing routine tasks, such as cleaning floors or delivering objects. However, a key limitation of these robots is their inability to recognize potential problems or dangers in home environments. For example, a child may pick up and ingest medication that has fallen on the floor, posing a serious risk. We argue that household robots should proactively detect such hazards or anomalies within the home, and propose the task of anomaly scenario generation. To accomplish this task, we leverage foundational models instead of relying on manually labeled data to build simulated environments. Specifically, we introduce a multi-agent brainstorming approach, where agents collaborate and generate diverse scenarios covering household hazards, hygiene management, and child safety. These textual task descriptions are then integrated with designed 3D assets to simulate realistic environments. Within these constructed environments, our LLM-based robotic agent learns the necessary skills to proactively discover and handle the proposed anomalies through task decomposition, optimal learning approach selection. We demonstrate that our generated environment outperforms others in terms of task description and scene diversity, ultimately enabling robotic agents to better address potential household hazards.

With the rise of AI-generated content spewed at scale from large language models (LLMs), genuine concerns about the spread of fake news have intensified. The perceived ability of LLMs to produce convincing fake news at scale poses new challenges for both human and automated fake news detection systems. To address this gap, this paper presents the findings from a university-level competition that aimed to explore how LLMs can be used by humans to create fake news, and to assess the ability of human annotators and AI models to detect it. A total of 110 participants used LLMs to create 252 unique fake news stories, and 84 annotators participated in the detection tasks. Our findings indicate that LLMs are ~68% more effective at detecting real news than humans. However, for fake news detection, the performance of LLMs and humans remains comparable (~60% accuracy). Additionally, we examine the impact of visual elements (e.g., pictures) in news on the accuracy of detecting fake news stories. Finally, we also examine various strategies used by fake news creators to enhance the credibility of their AI-generated content. This work highlights the increasing complexity of detecting AI-generated fake news, particularly in collaborative human-AI settings.

Knowledge editing methods (KEs) can update language models’ obsolete or inaccurate knowledge learned from pre-training. However, KEs can be used for malicious applications, e.g., inserting misinformation and toxic content. Knowing whether a generated output is based on edited knowledge or first-hand knowledge from pre-training can increase users’ trust in generative models and provide more transparency. Driven by this, we propose a novel task: detecting knowledge edits in language models. Given an edited model and a fact retrieved by a prompt from an edited model, the objective is to classify the knowledge as either unedited (based on the pre-training), or edited (based on subsequent editing). We instantiate the task with four KEs, two large language models (LLMs), and two datasets. Additionally, we propose using hidden state representations and probability distributions as features for the detection model. Our results reveal that using these features as inputs to a simple AdaBoost classifier establishes a strong baseline. This baseline classifier requires a small amount of training data and maintains its performance even in cross-domain settings. Our work lays the groundwork for addressing potential malicious model editing, which is a critical challenge associated with the strong generative capabilities of LLMs.

Emergent abilities of large language models (LLMs) have significantly advanced their application in autonomous vehicle (AV) research. Safe integration of LLMs into vehicles, however, necessitates their thorough understanding of dynamic traffic environments. Towards this end, this study introduces a framework leveraging LLMs’ built-in extrapolation capabilities for vehicle trajectory prediction, thereby evaluating their comprehension of the evolution of traffic agents’ behaviors and interactions over time. The framework employs a traffic encoder to extract spatial-level scene features from agents’ observed trajectories to facilitate efficient scene representation. To focus on LLM’s innate capabilities, scene features are then converted into LLM-compatible tokens through a reprogramming adapter and finally decoded into predicted trajectories with a linear decoder. Experimental results quantitatively demonstrate the framework’s efficacy in enabling off-the-shelf, frozen LLMs to achieve competitive trajectory prediction performance, with qualitative analyses revealing their enhanced understanding of complex, multi-agent traffic scenarios.

Language models perform well on emotion datasets but it remains unclear whether these models indeed understand emotions expressed in text or simply exploit supperficial lexical cues (e.g., emotion words). In this paper, we present two novel test evaluation sets sourced from two existing datasets that allow us to evaluate whether language models make real inferential decisions for emotion detection or not. Our human-annotated test sets are created by iteratively rephrasing input texts to gradually remove explicit emotion cues (while preserving the semantic similarity and the emotions) until a strong baseline BERT model yields incorrect predictions. Using our new test sets, we carry out a comprehensive analysis into the capabilities of small and large language models to predict emotions. Our analysis reveals that all models struggle to correctly predict emotions when emotion lexical cues become scarcer and scarcer, but large language models perform better than small pre-trained language models and push the performance by 14% over the 5% BERT baseline. We make our evaluation test sets and code publicly available.

Zero-shot Named Entity Recognition (ZS-NER) aims to recognize entities in unseen domains without specific annotated data. A key challenge is handling missing entities while ensuring accurate type recognition, hindered by: 1) the pre-training assumption that each entity has a single type, overlooking diversity, and 2) insufficient contextual knowledge for type reasoning. To address this, we propose IRRA (Integrated Recall and Retrieval Augmentation), a novel two-stage framework leveraging large language model techniques. In the Recall Augmented Entity Extracting stage, we built a perturbed dataset to induce the model to exhibit missing or erroneous extracted entities. Based on this, we trained an enhanced model to correct these errors. This approach can improve the ZS-NER’s recall rate. In the Retrieval Augmented Type Correcting stage, we employ Retrieval-Augmented Generation techniques to locate entity-related unannotated contexts, with the additional contextual information significantly improving the accuracy of type correcting. Extensive evaluations demonstrate the state-of-the-art performance of our IRRA, with significant improvements in zero-shot cross-domain settings validated through both auto-evaluated metrics and analysis. Our implementation will be open-sourced athttps://github.com/DMIRLAB-Group/IRRA.

Transformer-based large language models (LLM) have been widely used in language processing applications. However, due to the memory constraints of the devices, most of them restrict the context window. Even though recurrent models in previous works can memorize past tokens to enable unlimited context and maintain effectiveness, they have “flat” memory architectures. Such architectures have limitations in selecting and filtering information. Since humans are good at learning and self-adjustment, we believe that imitating brain memory hierarchy is beneficial for model memorization. Thus, we propose the Hierarchical Memory Transformer (HMT), a novel framework that facilitates a model’s long-context processing ability by imitating human memorization behavior. Leveraging memory-augmented segment-level recurrence, we organize the memory hierarchy by preserving tokens from early input segments, passing memory embeddings along the sequence, and recalling relevant information from history. Evaluating general language modeling, question-answering tasks, and the summarization task, we show that HMT consistently improves the long-context processing ability of existing models. Furthermore, HMT achieves a comparable or superior generation quality to long-context LLMs with 2 ∼ 57× fewer parameters and 2.5 ∼ 116× less inference memory, significantly outperforming previous memory-augmented models.

Aligning language models with human values is crucial, especially as they become more integrated into everyday life. While models are often adapted to user preferences, it is equally important to ensure they align with moral norms and behaviours in real-world social situations. Despite significant progress in languages like English and Chinese, French has seen little attention in this area, leaving a gap in understanding how LLMs handle moral reasoning in this language. To address this gap, we introduce HistoiresMorales, a French dataset derived from MoralStories, created through translation and subsequently refined with the assistance of native speakers to guarantee grammatical accuracy and adaptation to the French cultural context. We also rely on annotations of the moral values within the dataset to ensure their alignment with French norms. HistoiresMorales covers a wide range of social situations, including differences in tipping practices, expressions of honesty in relationships, and responsibilities toward animals. To foster future research, we also conduct preliminary experiments on the alignment of multilingual models on French and English data and the robustness of the alignment. We find that while LLMs are generally aligned with human moral norms by default, they can be easily influenced with user-preference optimization for both moral and immoral data.

Quantizing large language models has become a standard way to reduce their memory and computational costs. Typically, existing methods focus on breaking down the problem into individual layer-wise sub-problems, and minimizing per-layer error, measured via various metrics. Yet, this approach currently lacks theoretical justification and the metrics employed may be sub-optimal. In this paper, we present a “linearity theorem” establishing a direct relationship between the layer-wise reconstruction error and the model perplexity increase due to quantization. This insight enables two novel applications: (1) a simple data-free LLM quantization method using Hadamard rotations and MSE-optimal grids, dubbed HIGGS, which outperforms all prior data-free approaches such as the extremely popular NF4 quantized format, and (2) an optimal solution to the problem of finding non-uniform per-layer quantization levels which match a given compression constraint, obtained by reduction to dynamic programming. On the practical side, we demonstrate improved accuracy-compression trade-offs on Llama-family models, advancing both data-free and non-uniform quantization for large language models.

The growth of conversational AI services has increased demand for effective information retrieval from dialogue data. However, existing methods often face challenges in capturing semantic intent or require extensive labeling and fine-tuning. This paper introduces HEISIR (Hierarchical Expansion of Inverted Semantic Indexing for Retrieval), a novel framework that enhances semantic understanding in conversational data retrieval through optimized data ingestion, eliminating the need for resource-intensive labeling or model adaptation.HEISIR implements a two-step process: (1) Hierarchical Triplets Formulation and (2) Adjunct Augmentation, creating semantic indices consisting of Subject-Verb-Object-Adjunct (SVOA) quadruplets. This structured representation effectively captures the underlying semantic information from dialogue content. HEISIR achieves high retrieval performance while maintaining low latency during the actual retrieval process. Our experimental results demonstrate that HEISIR outperforms fine-tuned models across various embedding types and language models. Beyond improving retrieval capabilities, HEISIR also offers opportunities for intent and topic analysis in conversational data, providing a versatile solution for dialogue systems.