United States

The emergence of large language models (LLMs) has significantly influenced numerous fields, including healthcare, by enhancing the capabilities of automated systems to process and generate human-like text. However, despite their advancements, the reliability and accuracy of LLMs in medical contexts remain critical concerns. Current evaluation methods often lack robustness and fail to provide a comprehensive assessment of LLM performance, leading to potential risks in clinical settings. In this work, we propose Med-CoDE, a specifically designed evaluation framework for medical LLMs to address these challenges. The framework leverages a critique-based approach to quantitatively measure the degree of disagreement between model-generated responses and established medical ground truths. This framework captures both accuracy and reliability in medical settings. The proposed evaluation framework aims to fill the existing gap in LLM assessment by offering a systematic method to evaluate the quality and trustworthiness of medical LLMs. Through extensive experiments and case studies, we illustrate the practicality of our framework in providing a comprehensive and reliable evaluation of medical LLMs.

NAACL 2025

Med-CoDE: Medical Critique based Disagreement Evaluation Framework

medical evaluation

large language models

automatic evaluation

## Welcome!
"I am excited to welcome you to this year’s edition of the Conference on Empirical Methods in Natural Language Processing! Importantly, it marks the 30th edition of EMNLP. With over 8,000 submissions, more than 3,000 accepted papers, and thousands of attendees, we have come a long way from that first
workshop, which had 14 accepted papers. As the field looks ahead, Suzhou is the fitting location for celebrating this milestone: rooted in a long literary tradition, yet modern and forward-looking, and home to a large share of the NLP community." 

*Message from the General Chair, Dirk Hovy*

EMNLP 2025

Celebrate 30 Years of EMNLP! 
EMNLP 2025 will be held in Suzhou, China from November 5th to November 9th, 2025.

poster

### Welcome to 2025 Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics

Welcome to the 2025 meeting of the Nations of the Americas Chapter of the Association for Computational Linguistics! I am proud to help organize the first NAACL conference to carry the new name of our organization, one that emphasizes inclusion for all of the Americas. I am also pleased to welcome you to Albuquerque, New Mexico, a state whose unique blend of cultural influences will make for an excellent backdrop for NAACL 2025, especially with this year’s special theme on NLP in a Multicultural World. 
**[Continue reading...](https://drive.google.com/file/d/1jX-qGhqVSZZCIrAnJaz798pflu5Irrdn/view?usp=sharing)**

*- Colin Cherry, Google, NAACL 2025 General Chair* 

[![](https://assets.underline.io/markdown_image/1/image/b087f8a4dc5816d6e1a6514e59c59ac3.png)](https://drive.google.com/file/d/1T96GzPqObXrMln2BMByCSXTSizjTg69P/view?usp=sharing)

You need to log in with the email address you registered with. Access credentials have been sent to your email. 

Please be sure to check your spam and other email folders if you do not see an email confirmation right away.

Please log in to explore this event.

To access NAACL 2025 event page you are required to register. Please follow [**this link**](https://2025.naacl.org/registration/registration/) to register. Access will depend on your registration type.

Please register!

Welcome to 2025 Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics

Distinguishing vulnerable code from non-vulnerable code is challenging due to high inter-class similarity. Supervised contrastive learning (SCL) improves embedding separation but struggles with intra-class clustering, especially when variations within the same class are subtle. We propose Cluster-Enhanced Supervised Contrastive Loss (CESCL), an extension of SCL with a distance-based regularization term that tightens intra-class clustering while maintaining inter-class separation. Evaluating on CodeBERT and GraphCodeBERT with Binary Cross Entropy (BCE), BCE + SCL, and BCE + CE-SCL, our method improves F1 score by 1.76% on CodeBERT and 4.1% on GraphCodeBERT, demonstrating its effectiveness in code vulnerability detection and broader applicability to high-similarity classification tasks.

Tighter Clusters, Safer Code? Improving Vulnerability Detection with Enhanced Contrastive Loss

Large Language Models (LLMs) excel in a wide range of tasks across natural language processing, multimodal applications, and other domains but are computationally expensive to train and fine-tune. Model merging is an efficient alternative that requires minimal computational resources, yet existing methods overlook the structural importance of individual weights within an LLM. We first investigate the concept of superweights, analyzing their significance in LLMs and their stability across domains. Based on these insights, we develop a superweight-based model merging strategy that selectively prioritizes critical parameters. Building upon this merging framework, we introduce SNTIES, a superweight-normalized TIES approach that enhances existing merging techniques through targeted normalization. By refining model representations and amplifying the importance of superweights, SNTIES improves performance, achieving a margin over TIES merging and demonstrating its effectiveness as a computationally efficient enhancement to model merging strategies.

Are a Handful of Weights Enough? Superweight-based Merging for Efficient Knowledge Transfer in Large Language Models

Mental manipulation is a subtle yet pervasive form of abuse in interpersonal communication, making its detection critical for safeguarding potential victims. However, due to manipulation's nuanced and context-specific nature, identifying manipulative language in complex, multi-turn, and multi-person conversations remains a significant challenge for large language models (LLMs). To address this gap, we introduce the MultiManip dataset, comprising 220 multi-turn, multi-person dialogues balanced between manipulative and non-manipulative interactions, all drawn from reality shows that mimic real-world scenarios. The dataset includes 11 distinct manipulations depicting real-life scenarios for manipulative interactions. We conduct extensive evaluations of state-of-the-art LLMs, such as GPT-4o and Llama-3.1-8B, employing various prompting strategies. Despite their advanced capabilities, these models often struggle to detect manipulation effectively. To overcome this limitation, we propose SELF-PERCEPT, a novel, two-stage prompting framework inspired by Self-Perception Theory, demonstrating strong performance in detecting multi-person, multi-turn mental manipulation. Our code and data are publicly available at (https://anonymous.4open.science/r/SELF-PERCEPT-B2FF).

SELF-PERCEPT: Detecting Multi-Person Mental Manipulation in Conversations

Retrieval-Augmented Generation (RAG) frameworks leveraging large language models (LLMs) frequently retrieve extraneous or weakly relevant information, leading to factual inaccuracies and hallucinations in generated responses. Existing document-level retrieval approaches lack sufficient granularity to effectively filter non-essential content. This paper introduces ChunkRAG, a retrieval framework that refines information selection through semantic chunking and chunk-level evaluation. ChunkRAG applies a dynamic greedy chunk aggregation strategy to segment documents into semantically coherent, variable-length sections based on cosine similarity. Empirical evaluations on the PopQA dataset indicate that ChunkRAG improves response accuracy over state-of-the-art RAG methods. The analysis further demonstrates that chunk-level filtering reduces redundant and weakly related information, enhancing the factual consistency of responses. By incorporating fine-grained retrieval mechanisms, ChunkRAG provides a scalable and domain-agnostic approach to mitigate hallucinations in knowledge-intensive tasks such as fact-checking and multi-hop reasoning.

ChunkRAG: A Novel LLM-Chunk Filtering Method for RAG Systems

Debate education fosters critical thinking skills but often incurs high human costs. Recent advancements in Large Language Models (LLMs) show promise in automating counter-argument generation. However, it remains unclear how best to guide LLMs to target both implicit and critical premises. In this study, we systematically compare multi-step and one-step generation methods for counter-arguments across 100 debate topics. Our findings reveal that one-step approaches consistently outperform multi-step pipelines, owing to their better grasp of the “motion spirit,” minimized propagation of hallucinations, and avoidance of challenging intermediate tasks. Among premise-targeting methods, a one-step strategy that accounts for both implicit and explicit premises—Generated and Targeted Premise Attack (GTG)—emerges as the strongest performer in expert and automated evaluations. These results highlight the value of direct, integrated prompts for leveraging LLMs in complex argumentation tasks and offer insights for developing more effective automated debate agents.

LLM DEBATE OPPONENT : Counter-argument Generation focusing on Implicit and Critical Premises

Detecting human emotions is a complex process often restricted by predetermined emotion labels, limiting the ability of systems to recognize nuanced or novel emotional expressions. To overcome this constraint, we explore a generative approach that uses large language models (LLMs) context creation and analyze the results with the usage of saliency maps. Instead of relying solely on static sample texts, our method generates dynamic backstories that provide additional context for understanding emotions in text or speech, and these backstories improved our evaluations. This novel framework re-imagines emotion detection as a context-driven process, which in turn enables systems to "see the bigger picture" and make more informed and accurate predictions based on added adjectives, nouns, and punctuation.

Exploring Taxonomic Emotion Detection via Saliency Mapping and Large Language Model Context Generation

Stigma has emerged as one of the major obstacles to effectively diagnosing depression, as it prevents users from open conversations about their struggles. This requires advanced questioning skills to carefully probe the presence of specific symptoms in an unobtrusive manner. While recent efforts have been made on depression-diagnosis-oriented dialogue systems, they largely ignore this problem, ultimately hampering their practical utility. To this end, we propose a novel and effective method, UPSD4, developing a series of strategies to promote a sense of unobtrusiveness within the dialogue system and assessing depression disorder by probing symptoms. We experimentally show that UPSD4 demonstrates a significant improvement over current baselines, including unobtrusiveness evaluation of dialogue content and diagnostic accuracy. We believe our work contributes to developing more accessible and user-friendly tools for addressing the widespread need for depression diagnosis.

Breaking the Stigma! Unobtrusively Probe Symptoms in Depression Disorder Diagnosis Dialogue

Recent large language models (LLMs) advancements sparked a growing research interest in tool assisted LLMs solving real-world challenges, which calls for comprehensive evaluation of tool-use capabilities. While previous works focused on either evaluating over stateless web services (RESTful API), based on a single turn user prompt, or an off-policy dialog trajectory, ToolSandbox includes stateful tool execution, implicit state dependencies between tools, a built-in user simulator supporting on-policy conversational evaluation and a dynamic evaluation strategy for intermediate and final milestones over arbitrary trajectory. We show that open source and proprietary models has a significant performance gap, and complex tasks like State Dependency, Canonicalization and Insufficient Information defined in ToolSandbox are challenging even the most capable SOTA LLMs, providing brand-new insights to tool-use LLM capabilities. Datasets and evaluation scripts of ToolSandbox are released at <placeholder>.

ToolSandbox: A Stateful, Conversational, Interactive Evaluation Benchmark for LLM Tool Use Capabilities

Retrieval-Augmented Generation (RAG) has become a powerful paradigm for enhancing large language models (LLMs) through external knowledge retrieval. Despite its widespread attention, existing academic research predominantly focuses on single-turn RAG, leaving a significant gap in addressing the complexities of multi-turn conversations found in real-world applications. To bridge this gap, we introduce CORAL, a large-scale benchmark designed to assess RAG systems in realistic multi-turn conversational settings. CORAL includes diverse information-seeking conversations automatically derived from Wikipedia and tackles key challenges such as open-domain coverage, knowledge intensity, free-form responses, and topic shifts. It supports three core tasks of conversational RAG: passage retrieval, response generation, and citation labeling. We propose a unified framework to standardize various conversational RAG methods and conduct a comprehensive evaluation of these methods on CORAL, demonstrating substantial opportunities for improving existing approaches.

CORAL: Benchmarking Multi-turn Conversational Retrieval-Augmented Generation

High annotation costs from hiring or crowdsourcing complicate the creation of large, high-quality datasets needed for training reliable text classifiers. Recent research suggests using Large Language Models (LLMs) to automate the annotation process, reducing these costs while maintaining data quality. LLMs have shown promising results in annotating downstream tasks like hate speech detection and political framing. Building on the success in these areas, this study investigates whether LLMs are viable for annotating a complex task of media bias detection and whether a downstream media bias classifier can be trained on such data. We create Annolexical, the first large-scale dataset for media bias classification with over 48k synthetically annotated examples. Our classifier fine-tuned on it surpasses all of the annotator LLMs by 5-9% in Mathew’s Correlation Coefficient (MCC) and performs close to or outperforms the model trained on human-labeled data when evaluated on two media bias benchmark datasets (BABE and BASIL). This study demonstrates how our approach significantly reduces the cost of dataset creation in the media bias domain and, by extension - the development of the classifiers, while our subsequent behavioral stress-testing reveals some of its current limitations and trade-offs.

Downloads

Next from NAACL 2025

Tighter Clusters, Safer Code? Improving Vulnerability Detection with Enhanced Contrastive Loss

Similar lecture

LLMEval-Med: A Real-world Clinical Benchmark for Medical LLMs with Physician Validation