NAACL 2025 论文列表

We examine the representation of African American English (AAE) in large language models (LLMs), exploring (a) the perceptions Black Americans have of how effective these technologies are at producing authentic AAE, and (b) in what contexts Black Americans find this desirable. Through both a survey of Black Americans (n= 104) and annotation of LLM-produced AAE by Black Americans (n= 228), we find that Black Americans favor choice and autonomy in determining when AAE is appropriate in LLM output. They tend to prefer that LLMs default to communicating in Mainstream U.S. English in formal settings, with greater interest in AAE production in less formal settings. When LLMs were appropriately prompted and provided in context examples, our participants found their outputs to have a level of AAE authenticity on par with transcripts of Black American speech. Select code and data for our project can be found here: https://github.com/smelliecat/AAEMime.git

Data scarcity is a major challenge in Few-shot Continual Relation Extraction (FCRE), where models must learn new relations from limited data while retaining past knowledge. Current methods, restricted by minimal data streams, struggle with catastrophic forgetting and overfitting. To overcome this, we introduce a novel *data augmentation strategy* that transforms single input sentences into complex texts by integrating both old and new data. Our approach sharpens model focus, enabling precise identification of word relationships based on specified relation types. By embedding adversarial training effects and leveraging new training perspectives through special objective functions, our method enhances model performance significantly. Additionally, we explore Sharpness-Aware Minimization (SAM) in Few-shot Continual Learning. Our extensive experiments uncover fascinating behaviors of SAM across tasks and offer valuable insights for future research in this dynamic field.

In this work, we propose Mutual Reinforcing Data Synthesis (MRDS) within LLMs to improve few-shot dialogue summarization task. Unlike prior methods that require external knowledge, we mutually reinforce the LLM’s dialogue synthesis and summarization capabilities, allowing them to complement each other during training and enhance overall performances. The dialogue synthesis capability is enhanced by directed preference optimization with preference scoring from summarization capability. The summarization capability is enhanced by the additional high quality dialogue-summary paired data produced by the dialogue synthesis capability. By leveraging the proposed MRDS mechanism, we elicit the internal knowledge of LLM in the format of synthetic data, and use it to augment the few-shot real training dataset. Empirical results demonstrate that our method improves dialogue summarization, achieving a 1.5% increase in ROUGE scores and a 0.3% improvement in BERT scores in few-shot settings. Furthermore, our method attains the highest average scores in human evaluations, surpassing both the pre-trained models and the baselines fine-tuned solely for summarization tasks.

The advent of Music-Language Models has greatly enhanced the automatic music generation capability of AI systems, but they are also limited in their coverage of the musical genres and cultures of the world. We present a study of the datasets and research papers for music generation and quantify the bias and under-representation of genres. We find that only 5.7% of the total hours of existing music datasets come from non-Western genres, which naturally leads to disparate performance of the models across genres.We then investigate the efficacy of Parameter-Efficient Fine-Tuning (PEFT) techniques in mitigating this bias. Our experiments with two popular models – MusicGen and Mustango, for two underrepresented non-Western music traditions – Hindustani Classical and Turkish Makam music, highlight the promises as well as the non-triviality of cross-genre adaptation of music through small datasets, implying the need for more equitable baseline music-language models that are designed for cross-cultural transfer learning.

Multimodal Large Language Models (MLLMs) have shown significant promise in various applications, leading to broad interest from researchers and practitioners alike. However, a comprehensive evaluation of their long-context capabilities remains underexplored. To address these gaps, we introduce the MultiModal Needle-in-a-haystack (MMNeedle) benchmark, specifically designed to assess the long-context capabilities of MLLMs. Besides multi-image input, we employ image stitching to further increase the input context length, and develop a protocol to automatically generate labels for sub-image level retrieval. Essentially, MMNeedle evaluates MLLMs by stress-testing their capability to locate a target sub-image (needle) within a set of images (haystack) based on textual instructions and descriptions of image contents. This setup necessitates an advanced understanding of extensive visual contexts and effective information retrieval within long-context image inputs. With this benchmark, we evaluate state-of-the-art MLLMs, encompassing both API-based and open-source models. The findings reveal that GPT-4o consistently surpasses other models in long-context scenarios, but suffers from hallucination problems in negative samples, i.e., when needles are not in the haystacks. Our comprehensive long-context evaluation of MLLMs also sheds lights on the considerable performance gap between API-based and open-source models. All the code, data, and instructions required to reproduce the main results are available at https://github.com/Wang-ML-Lab/multimodal-needle-in-a-haystack.

Multimodal content generation has become an area of considerable interest. However, existing methods are hindered by limitations related to model constraints and training strategies: (1) Most current approaches rely on training models from scratch, resulting in inefficient training processes when extending these models; (2) There is a lack of constraints on adjacent steps within the models, leading to slow sampling and poor generation stability across various sampling methods. To address the issues, we introduce Multimodal Generation with Consistency Transferring (MGCT). The method introduces two key improvements: (1) A Model Consistency Transferring (MCT) strategy to acquire low-cost prior knowledge, increasing training efficiency and avoiding error accumulation; (2) A Layer Consistency Transferring (LCT) between adjacent steps, enhancing denoising capabilities at each step and improving model stability across various generation methods. These strategies ensure the consistency of jointly generated multimodal content and improving training efficiency. Experiments show that the algorithm enhances the model’s ability to capture actions and depict backgrounds more effectively. In both the AIST++ and Landscape datasets, it improves video generation speed by approximately 40% and quality by about 39.3%, while also achieving a slight 3% improvement in audio quality over the baseline.

Previous research has revealed the potential of large language models (LLMs) to support cognitive reframing therapy; however, their focus was primarily on text-based methods, often overlooking the importance of non-verbal evidence crucial in real-life therapy. To alleviate this gap, we extend the textual cognitive reframing to multimodality, incorporating visual clues. Specifically, we present a new dataset called Multi Modal-Cognitive Support Conversation (M2CoSC), which pairs each GPT-4-generated dialogue with an image that reflects the virtual client’s facial expressions.To better mirror real psychotherapy, where facial expressions lead to interpreting implicit emotional evidence, we propose a multi-hop psychotherapeutic reasoning approach that explicitly identifies and incorporates subtle evidence. Our comprehensive experiments with both LLMs and vision-language models (VLMs) demonstrate that the VLMs’ performance as psychotherapists is significantly improved with the M2CoSC dataset. Furthermore, the multi-hop psychotherapeutic reasoning method enables VLMs to provide more thoughtful and empathetic suggestions, outperforming standard prompting methods.

Although reasoning is innately language-agnostic, the multilingual capacities remains a significant challenge for large language models (LLMs). Their ability to generate structured, step-wise explanations is constantly restricted to dominant languages in pre-training data, making cross-lingual generalisation difficult and hindering broader global adoption. Recent works have introduced eclectic strategies to improve reasoning beyond English; however, these methods remain related to specific language that is not always optimal for reasoning.To improve LLMs’ multilingual reasoning abilities, we propose a modular approach that instructs the models to structure reasoning passages in a different problem space and then self-refine their capabilities to deliver step-wise reasoning passages that lead to the solution. Experiments show that our approach stably achieves significant improvements in the multilingual reasoning of various models and task, with improved reasoning consistency across languages.

While recent large language models (LLMs) demonstrate remarkable abilities in responding to queries in diverse languages, their ability to handle long multilingual contexts is unexplored. As such, a systematic evaluation of the long-context capabilities of LLMs in multilingual settings is crucial, specifically in the context of information retrieval. To address this gap, we introduce the MultiLingual Needle-in-a-Haystack (MLNeedle) test, designed to assess a model’s ability to retrieve relevant information (the needle) from a collection of multilingual distractor texts (the haystack). This test serves as an extension of the multilingual question-answering task, encompassing both monolingual and cross-lingual retrieval. We evaluate four state-of-the-art LLMs on MLNeedle. Our findings reveal that model performance can vary significantly with language and needle position. Specifically, we observe that model performance is the lowest when the needle is (i) in a language outside the English language family, and (ii) located in the middle of the input context. Furthermore, although some models claim a context size of 8k tokens or greater, none demonstrate satisfactory cross-lingual retrieval performance as the context length increases. Our analysis provides key insights into the long-context behavior of LLMs in multilingual settings to guide future evaluation protocols. To our knowledge, this is the first study to investigate the multilingual long-context behavior of LLMs.

Large language models (LLMs) have shown continuously improving multilingual capabilities, and even small-scale open-source models have demonstrated rapid performance enhancement. In this paper, we systematically explore the abilities of open LLMs with less than ten billion parameters to handle multilingual machine translation (MT) tasks. We conduct comprehensive evaluations on six popular LLMs and find that models like Gemma2-9B exhibit impressive multilingual translation capabilities. We then introduce the Parallel-First Monolingual-Second (PFMS) data mixing strategy in the continual pretraining stage to further enhance the MT performance and present GemmaX2-28, a 9B model achieving top-tier multilingual translation performance across 28 languages. Specifically, GemmaX2-28 consistently outperforms the state-of-the-art (SOTA) models such as TowerInstruct and X-ALMA and achieves competitive performance with Google Translate and GPT-4-turbo.

As safety remains a crucial concern throughout the development lifecycle of Large Language Models (LLMs), researchers and industrial practitioners have increasingly focused on safeguarding and aligning LLM behaviors with human preferences and ethical standards. LLMs, trained on extensive multilingual corpora, exhibit powerful generalization abilities across diverse languages and domains. However, current safety alignment practices predominantly focus on single-language scenarios, which leaves their effectiveness in complex multilingual contexts, especially for those complex mixed-language formats, largely unexplored. In this study, we introduce Multilingual Blending, a mixed-language query-response scheme designed to evaluate the safety alignment of various state-of-the-art LLMs (e.g., GPT-4o, GPT 3.5, Llama3) under sophisticated, multilingual conditions. We further investigate language patterns such as language availability, morphology, and language family that could impact the effectiveness of Multilingual Blending in compromising the safeguards of LLMs. Our experimental results show that, without meticulously crafted prompt templates, Multilingual Blending significantly amplifies the detriment of malicious queries, leading to dramatically increased bypass rates in LLM safety alignment (67.23% on GPT-3.5 and 40.34% on GPT-4o), far exceeding those of single-language baselines. Moreover, the performance of Multilingual Blending varies notably based on intrinsic linguistic properties, with languages of different morphology and from diverse families being more prone to evading safety alignments. These findings underscore the necessity of evaluating LLMs and developing corresponding safety alignment strategies in a complex, multilingual context to align with their superior cross-language generalization capabilities.

Multimodal Large Language Models (MLLMs) have demonstrated impressive abilities across various tasks, including visual question answering and chart comprehension, yet existing benchmarks for chart-related tasks fall short in capturing the complexity of real-world multi-chart scenarios. Current benchmarks primarily focus on single-chart tasks, neglecting the multi-hop reasoning required to extract and integrate information from multiple charts, which is essential in practical applications. To fill this gap, we introduce MultiChartQA, a benchmark that evaluates MLLMs’ capabilities in four key areas: direct question answering, parallel question answering, comparative reasoning, and sequential reasoning. Our evaluation of a wide range of MLLMs reveals significant performance gaps compared to humans. These results highlight the challenges in multi-chart comprehension and the potential of MultiChartQA to drive advancements in this field. Our code and data are available at https://github.com/Zivenzhu/Multi-chart-QA.

Successful teamwork requires team members to understand each other and communicate effectively, managing multiple linguistic and paralinguistic tasks at once. Because of the potential for interrelatedness of these tasks, it is important to have the ability to make multiple types of predictions on the same dataset. Here, we introduce Multimodal Communication Annotations for Teams (MultiCAT), a speech- and text-based dataset consisting of audio recordings, automated and hand-corrected transcriptions. MultiCAT builds upon data from teams working collaboratively to save victims in a simulated search and rescue mission, and consists of annotations and benchmark results for the following tasks: (1) dialog act classification, (2) adjacency pair detection, (3) sentiment and emotion recognition, (4) closed-loop communication detection, and (5) vocal (phonetic) entrainment detection. We also present exploratory analyses on the relationship between our annotations and team outcomes. We posit that additional work on these tasks and their intersection will further improve understanding of team communication and its relation to team performance. Code & data: https://doi.org/10.5281/zenodo.14834835

Hate speech moderation on global platforms poses unique challenges due to the multimodal and multilingual nature of content, along with the varying cultural perceptions. How well do current vision-language models (VLMs) navigate these nuances? To investigate this, we create the first multimodal and multilingual parallel hate speech dataset, annotated by a multiculturally diverse set of annotators, called Multi3Hate. It contains 300 parallel meme samples across 5 languages: English, German, Spanish, Hindi, and Mandarin. We demonstrate that cultural background significantly affects multimodal hate speech annotation in our dataset. The average pairwise agreement among countries is just 74%, significantly lower than that of randomly selected annotator groups. Our qualitative analysis indicates that the lowest pairwise label agreement—only 67% between the USA and India—can be attributed to cultural factors. We then conduct experiments with 5 large VLMs in a zero-shot setting, finding that these models align more closely with annotations from the US than with those from other cultures, even when the memes and prompts are presented in the native language of the other culture.

In recent years, large language models (LLMs) have made significant progress in knowledge-intensive applications. However, when adapting them to specific domains, we may encounter a multi-stage continuous learning scenario, especially in cases where domain knowledge evolves rapidly.This issue severely limits traditional fine-tuning approaches for LLMs.To overcome this limitation, we propose a new learning paradigm designed specifically for multi-stage continuous learning. This paradigm includes a preference-based learning bias to identify potential knowledge conflicts, as well as a self-distillation-based data augmentation strategy to expand and enrich the training corpus, thereby improving the integration of knowledge-compatible information.In the experiments, we show that our proposed method achieves a significant improvement in accuracy after 7 stages of fine-tuning compared to previous methods, while also demonstrating excellent performance in preserving general knowledge.We have released our code and dataset at Multi-Stage-Learning.

Utilizing large language models (LLMs) to rank a set of items has become a common approach in recommendation and retrieval systems. Typically, these systems focus on ordering a substantial number of documents in a monotonic order based on a given query. However, real-world scenarios often present a different challenge: ranking a comparatively smaller set of items, but according to a variety of diverse and occasionally conflicting conditions. In this paper, we define and explore the task of multi-conditional ranking by introducing MCRank, a benchmark tailored for assessing multi-conditional ranking across various item types and conditions. Our analysis of LLMs using MCRank indicates a significant decrease in performance as the number and complexity of items and conditions grow. To overcome this limitation, we propose a novel decomposed reasoning method, consisting of EXtracting and Sorting the conditions, and then Iteratively Ranking the items (EXSIR). Our extensive experiments show that this decomposed reasoning method enhances LLMs’ performance significantly, achieving up to a 14.4% improvement over existing LLMs. We also provide a detailed analysis of LLMs performance across various condition categories, and examine the effectiveness of decomposition step. Furthermore, we compare our method with existing approaches such as Chain-of-Thought and existing ranking models, demonstrating the superiority of our approach and complexity of MCR task. We will make our dataset and code publicly available.

Emotion recognition in conversation (ERC) involves identifying emotional labels associated with utterances within a conversation, a task that is essential for developing empathetic robots. Current research emphasizes contextual factors, the speaker’s influence, and extracting complementary information across different modalities. However, it often overlooks the cross-modal noise at the semantic level and the redundant information brought by the features themselves. This study introduces a diffusion-based approach designed to effectively address the challenges posed by redundant information and unexpected noise while robustly capturing shared semantics, thus facilitating the learning of compact and representative features from multimodal data. Specifically, we present the Multi-Condition Guided Diffusion Network (McDiff). McDiff employs a modal prior knowledge extraction strategy to derive the prior distribution for each modality, thereby enhancing the regional attention of each modality and applying the generated prior distribution at each diffusion step. Furthermore, we propose a method to learn the mutual information of each modality through a specific objective constraints approach prior to the forward process, which aims to improve inter-modal interaction and mitigate the effects of noise and redundancy. Comprehensive experiments conducted on two multimodal datasets, IEMOCAP and MELD, demonstrate that McDiff significantly surpasses existing state-of-the-art methodologies, thereby affirming the generalizability and efficacy of the proposed model.

Large Language Models (LLMs) have significantly advanced legal intelligence, but the scarcity of scenario data impedes the progress toward interactive legal scenarios. This paper introduces a Multi-agent Legal Simulation Driver (MASER) to scalably generate synthetic data by simulating interactive legal scenarios. Leveraging real-legal case sources, MASER ensures the consistency of legal attributes between participants and introduces a supervisory mechanism to align participants’ characters and behaviors as well as addressing distractions. A Multi-stage Interactive Legal Evaluation (MILE) benchmark is further constructed to evaluate LLMs’ performance in dynamic legal scenarios. Extensive experiments confirm the effectiveness of our framework.

We introduce MorphNLI, a modular step-by-step approach to natural language inference (NLI). When classifying the premise-hypothesis pairs into entailment, contradiction, neutral, we use a language model to generate the necessary edits to incrementally transform (i.e., morph) the premise into the hypothesis. Then, using an off-the-shelf NLI model we track how the entailment progresses with these atomic changes, aggregating these intermediate labels into a final output. We demonstrate the advantages of our proposed method particularly in realistic cross-domain settings, where our method always outperforms strong baselines with improvements up to 12.6% (relative). Further, our proposed approach is explainable as the atomic edits can be used to understand the overall NLI label.