NAACL 2025 论文列表

Dysarthric speech recognition often suffers from performance degradation due to the intrinsic diversity of dysarthric severity and extrinsic disparity from normal speech. To bridge these gaps, we propose a Dynamic Phoneme-level Contrastive Learning (DyPCL) method, which leads to obtaining invariant representations across diverse speakers. We decompose the speech utterance into phoneme segments for phoneme-level contrastive learning, leveraging dynamic connectionist temporal classification alignment. Unlike prior studies focusing on utterance-level embeddings, our granular learning allows discrimination of subtle parts of speech. In addition, we introduce dynamic curriculum learning, which progressively transitions from easy negative samples to difficult-to-distinguishable negative samples based on phonetic similarity of phoneme. Our approach to training by difficulty levels alleviates the inherent variability of speakers, better identifying challenging speeches. Evaluated on the UASpeech dataset, DyPCL outperforms baseline models, achieving an average 22.10% relative reduction in word error rate (WER) across the overall dysarthria group.

Despite their widespread success, Text-to-Image models (T2I) still struggle to produce images that are both aesthetically pleasing and faithful to the user’s input text. We introduce DreamSync, a simple yet effective training algorithm that improves T2I models to be faithful to the text input. DreamSync utilizes large vision-language models (VLMs) to effectively identify the fine-grained discrepancies between generated images and the text inputs and enable T2I models to self-improve without labeled data. First, it prompts the model to generate several candidate images for a given input text. Then, it uses two VLMs to select the best generation: a Visual Question Answering model that measures the alignment of generated images to the text, and another that measures the generation’s aesthetic quality. After selection, we use LoRA to iteratively finetune the T2I model to guide its generation towards the selected best generations. DreamSync does not need any additional human annotation, model architecture changes, or reinforcement learning. Despite its simplicity, DreamSync improves both the semantic alignment and aesthetic appeal of two diffusion-based T2I models, evidenced by multiple benchmarks (+1.7% on TIFA, +2.9% on DSG1K, +3.4% on VILA aesthetic) and human evaluation shows that DreamSync improves text rendering compared to SDXL by 18.5% on DSG1K benchmark.

In real-world settings, vision language models (VLMs) should robustly handle naturalistic, noisy visual content as well as domain-specific language and concepts. For example, K-12 educators using digital learning platforms may need to examine and provide feedback across many images of students’ math work. To assess the potential of VLMs to support educators in settings like this one, we introduce DrawEduMath, an English-language dataset of 2,030 images of students’ handwritten responses to K-12 math problems. Teachers provided detailed annotations, including free-form descriptions of each image and 11,661 question-answer (QA) pairs. These annotations capture a wealth of pedagogical insights, ranging from students’ problem-solving strategies to the composition of their drawings, diagrams, and writing. We evaluate VLMs on teachers’ QA pairs, as well as 44,362 synthetic QA pairs derived from teachers’ descriptions using language models (LMs). We show that even state-of-the-art VLMs leave much room for improvement on DrawEduMath questions. We also find that synthetic QAs, though imperfect, can yield similar model rankings as teacher-written QAs. We release DrawEduMath to support the evaluation of VLMs’ abilities to reason mathematically over images gathered with educational contexts in mind.

The common practice of preprocessing text before feeding it into NLP models introduces many decision points which have unintended consequences on model performance. In this opinion piece, we focus on the handling of diacritics in texts originating in many languages and scripts. We demonstrate, through several case studies, the adverse effects of inconsistent encoding of diacritized characters and of removing diacritics altogether. We call on the community to adopt simple but necessary steps across all models and toolkits in order to improve handling of diacritized text and, by extension, increase equity in multilingual NLP.

Most research on abstractive summarization focuses on single-domain applications, often neglecting how domain shifts between documents affect performance and the generalization ability of summarization models. To address this issue, we introduce DomainSum, a hierarchical benchmark designed to capture fine-grained domain shifts in abstractive summarization. We categorize these shifts into three levels: genre, style, and topic, and demonstrate through comprehensive benchmark analysis that they follow a hierarchical structure. Furthermore, we evaluate the domain generalization capabilities of commonly used pre-trained language models (PLMs) and large language models (LLMs) in both in-domain and cross-domain settings. Our benchmark and source code are released at https://github.com/hpzhang94/DomainSum.

Self-attention has revolutionized natural language processing by capturing long-range dependencies and improving context understanding. However, it comes with high computational costs and struggles with sequential data’s inherent directionality. This paper investigates and presents a simplified approach called “shared weight self-attention,” where a single weight matrix is used for Keys, Queries, and Values instead of separate matrices for each. This approach cuts training parameters by more than half and significantly reduces training time. Our method not only improves efficiency but also achieves strong performance on tasks from the GLUE benchmark, even outperforming the standard BERT baseline in handling noisy and out-of-domain data. Experimental results show a 66.53% reduction in parameter size within the attention block and competitive accuracy improvements of 3.55% and 0.89% over symmetric and pairwise attention-based BERT models, respectively.

Recent studies have highlighted the presence of cultural biases in Large Language Models (LLMs), yet often lack a robust methodology to dissect these phenomena comprehensively. Our work aims to bridge this gap by delving into the Food domain—a universally relevant yet culturally diverse aspect of human life. We introduce FmLAMA, a multilingual dataset centered on food-related cultural facts and variations in food practices. We analyze LLMs across various architectures and configurations, evaluating their performance in both monolingual and multilingual settings. By leveraging templates in six different languages, we investigate how LLMs interact with language-specific and cultural knowledge. Our findings reveal that (1) LLMs demonstrate a pronounced bias towards food knowledge prevalent in the United States; (2) Incorporating relevant cultural context significantly improves LLMs’ ability to access cultural knowledge; (3) The efficacy of LLMs in capturing cultural nuances is highly dependent on the interplay between the probing language, the specific model architecture, and the cultural context in question. This research underscores the complexity of integrating cultural understanding into LLMs and emphasizes the importance of culturally diverse datasets to mitigate biases and enhance model performance across different cultural domains.

Despite their wide adoption, the biases and unintended behaviors of language models remain poorly understood. In this paper, we identify and characterize a phenomenon never discussed before, which we call semantic leakage, where models leak irrelevant information from the prompt into the generation in unexpected ways. We propose an evaluation setting to detect semantic leakage both by humans and automatically, curate a diverse test suite for diagnosing this behavior, and measure significant semantic leakage in 13 flagship models. We also show that models exhibit semantic leakage in languages besides English and across different settings and generation scenarios. This discovery highlights yet another type of bias in language models that affects their generation patterns and behavior.

Nigeria is a multilingual country with 500+ languages. Naija is a Nigerian Pidgin spoken by approximately 120M speakers and it is a mixed language (e.g., English, Portuguese, Yoruba, Hausa and Igbo). Although it has mainly been a spoken language until recently, there are some online platforms (e.g., Wikipedia), publishing in written Naija as well. West African Pidgin English (WAPE) is also spoken in Nigeria and it is used by BBC to broadcast news on the internet to a wider audience not only in Nigeria but also in other West African countries (e.g., Cameroon and Ghana). Through statistical analyses and Machine Translation experiments, our paper shows that these two pidgin varieties do not represent each other (i.e., there are linguistic differences in word order and vocabulary) and Generative AI operates only based on WAPE. In other words, Naija is underrepresented in Generative AI, and it is hard to teach LLMs with few examples. In addition to the statistical analyses, we also provide historical information on both pidgins as well as insights from the interviews conducted with volunteer Wikipedia contributors in Naija.

Large language models (LLMs) have demonstrated great performance across various benchmarks, showing potential as general-purpose task solvers. However, as LLMs are typically trained on vast amounts of data, a significant concern in their evaluation is data contamination, where overlap between training data and evaluation datasets inflates performance assessments. Multiple approaches have been developed to identify data contamination. These approaches rely on specific assumptions that may not hold universally across different settings. To bridge this gap, we systematically review 50 papers on data contamination detection, categorize the underlying assumptions, and assess whether they have been rigorously validated. We identify and analyze eight categories of assumptions and test three of them as case studies. Our case studies focus on detecting direct, instance-level data contamination, which is also referred to as Membership Inference Attacks (MIA). Our analysis reveals that MIA approaches based on these three assumptions can have similar performance to random guessing, on datasets used in LLM pretraining, suggesting that current LLMs might learn data distributions rather than memorizing individual instances. Meanwhile, MIA can easily fail when there are data distribution shifts between the seen and unseen instances.

Evaluating retrieval-augmented generation (RAG) systems remains challenging, particularly for open-ended questions that lack definitive answers and require coverage of multiple sub-topics. In this paper, we introduce a novel evaluation framework based on sub-question coverage, which measures how well a RAG system addresses different facets of a question. We propose decomposing questions into sub-questions and classifying them into three types—core, background, and follow-up—to reflect their roles and importance. Using this categorization, we introduce a fine-grained evaluation protocol that provides insights into the retrieval and generation characteristics of RAG systems, including three commercial generative answer engines: You.com, Perplexity AI, and Bing Chat. Interestingly, we find that while all answer engines cover core sub-questions more often than background or follow-up ones, they still miss around 50% of core sub-questions, revealing clear opportunities for improvement. Further, sub-question coverage metrics prove effective for ranking responses, achieving 82% accuracy compared to human preference annotations. Lastly, we also demonstrate that leveraging core sub-questions enhances both retrieval and answer generation in a RAG system, resulting in a 74% win rate over the baseline that lacks sub-questions.

Uncertainty estimation (UE) of generative large language models (LLMs) is crucial for evaluating the reliability of generated sequences. A significant subset of UE methods utilize token probabilities to assess uncertainty, aggregating multiple token probabilities into a single UE score using a scoring function. Existing scoring functions for probability-based UE, such as length-normalized scoring and semantic contribution-based weighting, are designed to solve certain aspects of the problem but exhibit limitations, including the inability to handle biased probabilities and complex semantic dependencies between tokens. To address these issues, in this work, we propose Learnable Response Scoring (LARS) function, a novel scoring function that leverages supervised data to capture complex dependencies between tokens and probabilities, thereby producing more reliable and calibrated response scores in computing the uncertainty of LLM generations. Our comprehensive experiments across question-answering and arithmetical reasoning tasks with various datasets demonstrate that LARS significantly outperforms existing scoring functions, achieving improvements of up to 16% AUROC score.

The rapid evolution of Large Language Models (LLMs) presents a promising solution to the global shortage of mental health professionals. However, their alignment with essential counseling competencies remains underexplored. We introduce CounselingBench, a novel NCMHCE-based benchmark evaluating 22 general-purpose and medical-finetuned LLMs across five key competencies. While frontier models surpass minimum aptitude thresholds, they fall short of expert-level performance, excelling in Intake, Assessment & Diagnosis but struggling with Core Counseling Attributes and Professional Practice & Ethics. Surprisingly, medical LLMs do not outperform generalist models in accuracy, though they provide slightly better justifications while making more context-related errors. These findings highlight the challenges of developing AI for mental health counseling, particularly in competencies requiring empathy and nuanced reasoning. Our results underscore the need for specialized, fine-tuned models aligned with core mental health counseling competencies and supported by human oversight before real-world deployment. Code and data associated with this manuscript can be found at: https://github.com/cuongnguyenx/CounselingBench

Recent advancements in Large Language Models (LLMs) have led to their adaptation in various domains as conversational agents. We wonder: can personality tests be applied to these agents to analyze their behavior, similar to humans? We introduce TRAIT, a new benchmark consisting of 8K multi-choice questions designed to assess the personality of LLMs. TRAIT is built on two psychometrically validated small human questionnaires, Big Five Inventory (BFI) and Short Dark Triad (SD-3), enhanced with the ATOMIC-10X knowledge graph to a variety of real-world scenarios. TRAIT also outperforms existing personality tests for LLMs in terms of reliability and validity, achieving the highest scores across four key metrics: Content Validity, Internal Validity, Refusal Rate, and Reliability. Using TRAIT, we reveal two notable insights into personalities of LLMs: 1) LLMs exhibit distinct and consistent personality, which is highly influenced by their training data (e.g., data used for alignment tuning), and 2) current prompting techniques have limited effectiveness in eliciting certain traits, such as high psychopathy or low conscientiousness, suggesting the need for further research in this direction.

Open-vocabulary audio language models (ALMs), like Contrastive Language Audio Pretraining (CLAP), represent a promising new paradigm for audio-text retrieval using natural language queries. In this paper, for the first time, we perform controlled experiments on various benchmarks to show that existing ALMs struggle to generalize to linguistic variations in textual queries. To address this issue, we propose RobustCLAP, a novel and compute-efficient technique to learn audio-language representations agnostic to linguistic variations. Specifically, we reformulate the contrastive loss used in CLAP architectures by introducing a multi-view contrastive learning objective, where paraphrases are treated as different views of the same audio scene and use this for training. Our proposed approach improves the text-to-audio retrieval performance of CLAP by 0.8%-13% across benchmarks and enhances robustness to linguistic variation. We make our code publicly available

We have uncovered a powerful jailbreak technique that leverages large language models’ ability to diverge from prior context, enabling them to bypass safety constraints and generate harmful outputs. By simply instructing the LLM to deviate and obfuscate previous attacks, our method dramatically outperforms existing approaches, achieving up to a 62.83% higher success rate in compromising ten leading chatbots, including GPT-4, Gemini, and Llama, while using only 12.9% of the queries. This revelation exposes a critical flaw in current LLM safety training, suggesting that existing methods may merely mask vulnerabilities rather than eliminate them. Our findings sound an urgent alarm for the need to revolutionize testing methodologies to ensure robust and reliable LLM security.

The retrieval augmented generation (RAG) framework addresses an ambiguity in user queries in QA systems by retrieving passages that cover all plausible interpretations and generating comprehensive responses based on the passages. However, our preliminary studies reveal that a single retrieval process often suffers from low-quality results, as the retrieved passages frequently fail to capture all plausible interpretations. Although the iterative RAG approach has been proposed to address this problem, it comes at the cost of significantly reduced efficiency. To address these issues, we propose the diversify-verify-adapt (DIVA) framework. DIVA first diversifies the retrieved passages to encompass diverse interpretations. Subsequently, DIVA verifies the quality of the passages and adapts the most suitable approach tailored to their quality. This approach improves the QA systems’ accuracy and robustness by handling low quality retrieval issue in ambiguous questions, while enhancing efficiency.

LLMs are increasingly used as seq2seq translators from natural language utterances to structured programs, a process called semantic interpretation. Unlike atomic labels or token sequences, programs are naturally represented as abstract syntax trees (ASTs). Such structured representation raises novel issues related to the design and selection of in-context examples (ICEs) presented to the LLM. We focus on decomposing the pool of available ICE trees into fragments, some of which may be better suited to solving the test instance. Next, we propose how to use (additional invocations of) an LLM with prompted syntax constraints to automatically map the fragments to corresponding utterances. Finally, we adapt and extend a recent method for diverse ICE selection to work with whole and fragmented ICE instances. We evaluate our system, SCUD4ICL, on popular diverse semantic parsing benchmarks, showing visible accuracy gains from our proposed decomposed diverse demonstration method. Benefits are particularly notable for smaller LLMs, ICE pools having larger labeled trees, and programs in lower resource languages.

Recently, with the development of tool-calling capabilities in large language models (LLMs), these models have demonstrated significant potential for automating electronic design automation (EDA) flows by interacting with EDA tool APIs via EDA scripts.However, considering the limited understanding of EDA tools, LLMs face challenges in practical scenarios where diverse interfaces of EDA tools exist across different platforms.Additionally, EDA flow automation often involves intricate, long-chain tool-calling processes, increasing the likelihood of errors in intermediate steps.Any errors will lead to the instability and failure of EDA flow automation.To address these challenges, we introduce EDAid, a multi-agent collaboration system where multiple agents harboring divergent thoughts converge towards a common goal, ensuring reliable and successful EDA flow automation. Specifically, each agent is controlled by ChipLlama models, which are expert LLMs fine-tuned for EDA flow automation.Our experiments demonstrate the state-of-the-art (SOTA) performance of our ChipLlama models and validate the effectiveness of our EDAid in the automation of complex EDA flows, showcasing superior performance compared to single-agent systems.