China

Large Language Models (LLMs) excel in general language tasks, motivating their adaptation to specialized domains such as healthcare. Effective domain adaptation typically involves supervised fine-tuning (SFT) on carefully selected instruction-tuning data. Current data selection methods adopt a data-centric approach, relying on external annotations and heuristics to identify external defined high-quality and challenging data. Our exploratory experiments highlight this approach fails to improve model&#39;s domain performance, due to misalignment between selected data and the model’s knowledge distribution. To tackle this, we propose Decomposed Difficulty-based Data Selection (3DS), a two-stage model-centric data selection framework that aligns data selection with the model’s distribution. 3DS employs a Prompt-Driven Data Selection to filter out noisy data based on the model&#39;s knowledge via explicit alignment in Stage#1, then adopts a Decomposed Difficulty-based Data Selection to guide selection via three novel data difficulty metrics, including Instruction Understanding, Response Confidence, and Response Correctness in Stage#2. These metrics are enhanced by an attention-based importance weighting mechanism for accurate calibration. Extensive experiments in the healthcare domain show 3DS outperforms existing methods by over 2.97% accuracy, with additional validation in the law domain confirming its generalization ability. Our dataset and code are open-sourced at https://anonymous.4open.science/r/3DS-E67F.

EMNLP 2025

3DS: Medical Domain Adaptation of LLMs via Decomposed Difficulty-based Data Selection

Large Language Models (LLMs) excel in general language tasks, motivating their adaptation to specialized domains such as healthcare. Effective domain adaptation typically involves supervised fine-tuning (SFT) on carefully selected instruction-tuning data. Current data selection methods adopt a data-centric approach, relying on external annotations and heuristics to identify external defined high-quality and challenging data. Our exploratory experiments highlight this approach fails to improve model's domain performance, due to misalignment between selected data and the model’s knowledge distribution. To tackle this, we propose Decomposed Difficulty-based Data Selection (3DS), a two-stage model-centric data selection framework that aligns data selection with the model’s distribution. 3DS employs a Prompt-Driven Data Selection to filter out noisy data based on the model's knowledge via explicit alignment in Stage#1, then adopts a Decomposed Difficulty-based Data Selection to guide selection via three novel data difficulty metrics, including Instruction Understanding, Response Confidence, and Response Correctness in Stage#2. These metrics are enhanced by an attention-based importance weighting mechanism for accurate calibration. Extensive experiments in the healthcare domain show 3DS outperforms existing methods by over 2.97% accuracy, with additional validation in the law domain confirming its generalization ability. Our dataset and code are open-sourced at https://anonymous.4open.science/r/3DS-E67F.

poster

## 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."<br>

*Message from the General Chair, Dirk Hovy*

[**Link to Conference Handbook**](https://drive.google.com/file/d/1johU5QqVVYO4RfH7QcIORr7qrVBdzdwC/view?usp=sharing)





<br>

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

Modeling complex subjective tasks in Natural Language Processing, such as recognizing emotion and morality, is considerably challenging due to significant variation in human annotations. This variation often reflects reasonable differences in semantic interpretations rather than mere noise, necessitating methods to distinguish between legitimate subjectivity and error. We address this challenge by exploring label verification in these contexts using Large Language Models (LLMs). First, we propose a simple In-Context Learning binary filtering baseline that estimates the reasonableness of a document-label pair. We then introduce the Label-in-a-Haystack setting: the query and its label(s) are included in the demonstrations shown to LLMs, which are prompted to predict the label(s) again, while receiving task-specific instructions (e.g., emotion recognition) rather than label copying. We show how the failure to copy the label(s) to the output of the LLM are task-relevant and informative. Building on this, we propose the Label-in-a-Haystack Rectification (LiaHR) framework for subjective label correction: when the model outputs diverge from the reference gold labels, we assign the generated labels to the example instead of discarding it. This approach can be integrated into annotation pipelines to enhance signal-to-noise ratios. Comprehensive analyses, human evaluations, and ecological validity studies verify the utility of LiaHR for label correction.

Humans Hallucinate Too: Language Models Identify and Correct Subjective Annotation Errors With Label-in-a-Haystack Prompts

Large vision-language models (LVLMs) have achieved impressive results in vision-language tasks. However, Therefore, we propose LACING, designed to address such bias with MuunderlinetextbfLtimodal DuunderlinetextbfAl-attention MeunderlinetextbfChanunderlinetextbfIsm (MDA) aunderlinetextbfNd Soft-Image underlinetextbfGuidance (SIG). Specifically, MDA adopts a textbfparallel dual-attention mechanism that constructs separate attention for visual and text inputs to enhance integration of visual inputs across model. SIG uses a textbflearnable soft visual prompt during training and inference to replace visual inputs, designed to compel LVLMs to prioritize text inputs during inference. Experiments across different model architectures and scales demonstrate that LACING effectively debiases LVLMs from their language bias, enhancing visual comprehension and reducing hallucinations without additional resources.

Looking Beyond Text: Reducing Language Bias in Large Vision-Language Models via Multimodal Dual-Attention and Soft-Image Guidance

Foundation multi-modal models are often designed by stitching of multiple existing pretrained uni-modal models: for example, an image classifier with an text model. This stitching process is performed by training a connector module that aims to align the representation spaces of these uni-modal models towards a multi-modal objective. However, given the complexity of training such connectors on large scale web-based datasets coupled with the ever-increasing number of available pretrained uni-modal models, the task of uni-modal models selection and subsequent connector module training becomes computationally demanding. To address this under-studied critical problem, we propose Hypernetwork Model Alignment (Hyma), a novel all-in-one solution for optimal uni-modal model selection and connector training by leveraging hypernetworks. Specifically, our framework utilizes the parameter prediction capability of a hypernetwork to obtain jointly trained connector modules for N times M combinations of uni-modal models. In our experiments, Hyma reduces the cost of searching for the best performing uni-modal model pair by 10times, while matching the ranking and trained connector performance obtained via grid search across a suite of diverse multi-modal benchmarks.

(Almost) Free Modality Stitching of Foundation Models

The quality of supervised fine-tuning (SFT) data is crucial for the performance of large multimodal models (LMMs), yet current data enhancement methods often suffer from factual errors and hallucinations due to inadequate visual perception. To address this challenge, we propose VERITAS, a pipeline that systematically integrates vision priors and multiple state-of-the-art LMMs with statistical methods to enhance SFT data quality. VERITAS leverages visual recognition models (RAM++) and OCR systems (PP-OCRv4) to extract structured vision priors, which are combined with images, questions, and answers. Three LMMs (GPT-4o, Gemini-2.5-Pro, Doubao-1.5-pro) evaluate the original answers, providing critique rationales and scores that are statistically fused into a high-confidence consensus score serving as ground truth. Using this consensus, we train a lightweight critic model via Group Relative Policy Optimization (GRPO), enhancing reasoning capabilities efficiently. Each LMM then refines the original answers based on the critiques, generating new candidate answers; we select the highest-scoring one as the final refined answer. Experiments across six multimodal benchmarks demonstrate that models fine-tuned with data processed by VERITAS consistently outperform those using raw data, particularly in text-rich and fine-grained reasoning tasks. Our critic model exhibits enhanced capability comparable to state-of-the-art LMMs while being significantly more efficient. We release our pipeline, datasets, and model checkpoints to advance research in multimodal data optimization.

VERITAS: Leveraging Vision Priors and Expert Fusion to Improve Multimodal Data

The paradigm of using Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs) as evaluative judges has emerged as an effective approach in RLHF and inference-time scaling. In this work, we propose Multimodal Reasoner as a Judge (MR. Judge), a paradigm for empowering general-purpose MLLMs judges with strong reasoning capabilities. Instead of directly assigning scores for each response, we formulate the judgement process as a reasoning-inspired multiple-choice problem. Specifically, the judge model first conducts deliberate reasoning covering different aspects of the responses and eventually selects the best response from them. This reasoning process not only improves the interpretibility of the judgement, but also greatly enhances the performance of MLLM judges. To cope with the lack of questions with scored responses, we propose the following strategy to achieve automatic annotation: 1) Reverse Response Candidates Synthesis: starting from a supervised fine-tuning (SFT) dataset, we treat the original response as the best candidate and prompt the MLLM to generate plausible but flawed negative candidates. 2) Text-based reasoning distillation: we carefully design a data synthesis pipeline for distilling the reasoning capability from a text-based reasoning model, which is adopted to enable the MLLM judges to regain complex reasoning ability via warm up supervised fine-tuning. Experiments demonstrate that our MR. Judge is effective across a wide range of tasks. Specifically, our MR. Judge-7B surpasses GPT-4o by 9.9% on VL-RewardBench, and improves performance on MM-Vet during inference-time scaling by up to 7.7%.

MR. Judge: Multimodal Reasoner as a Judge

This study presents a novel approach to assessing French text readability for adults with low literacy skills, addressing both global (full-text) and local (segment-level) difficulty. We introduce a dataset of 461 texts annotated using a difficulty scale developed specifically for this population. Using this corpus, we conducted a systematic comparison of key readability modeling approaches, including machine learning techniques based on linguistic variables, fine-tuning of CamemBERT, a hybrid approach combining BERT with linguistic variables, and the use of generative language models (LLMs) to carry out readability assessment at both global and local levels.

Assessing French Readability for Adults with Low Literacy: A Global and Local Perspective

While diffusion models have revolutionized text-to-image generation with their ability to synthesize realistic and diverse scenes, they continue to struggle with generating consistent and legible text within images. This shortcoming is commonly attributed to the locality bias inherent in diffusion-based generation, which limits their capacity to model long-range spatial dependencies. In this paper, we introduce \textbf{STRICT}, a benchmark designed to systematically stress-test the ability of diffusion models to render coherent and instruction-aligned text in images. Our benchmark evaluates models across multiple dimensions: (1) the maximum length of readable text that can be generated and (2) the correctness and legibility of the generated text. We assess several state-of-the-art models, including proprietary and open-source variants, and reveal persistent limitations in long-range consistency and instruction-following capabilities. Our findings provide insights into architectural bottlenecks and motivate future research directions in multimodal generative modeling.

STRICT: Stress-Test of Rendering Image Containing Text

Large Vision-Language Models (LVLMs) achieve remarkable performance in multimodal tasks but suffer from high computational costs due to the large number of visual tokens. Existing pruning methods either apply after visual tokens enter the LLM or perform pre-pruning based solely on visual attention. Both fail to balance efficiency and semantic alignment, as post-pruning incurs redundant computation, while visual-only pre-pruning overlooks multimodal relevance. To address this limitation, we propose Dual-Cue Pruning (DCP), a novel cross-modal pruning framework that jointly considers textual semantics and visual self-attention. DCP consists of a text-aware computation module, which employs a gradient-weighted attention mechanism to enhance text-visual alignment, and an image-aware computation module, which utilizes deep-layer self-attention distributions to retain essential structural information. By integrating both cues, DCP adaptively selects the most informative visual tokens, achieving efficient inference acceleration while maintaining strong task performance. Experimental results show that DCP can retain only 25\% of the visual tokens, with a minimal performance degradation of only 0.063\% on LLaVA-1.5-13B, demonstrating its effectiveness in balancing efficiency and accuracy.

DCP: Dual-Cue Pruning for Efficient Large Vision-Language Models

Large language models (LLMs) often struggle with context fidelity, producing inconsistent answers when responding to questions based on provided information. Existing approaches either rely on expensive supervised fine-tuning to generate evidence post-answer or train models to perform web searches without necessarily improving utilization of the given context. We propose CARE, a novel native retrieval-augmented reasoning framework that teaches LLMs to explicitly integrate in-context evidence within their reasoning process with the model's own retrieval capabilities. Our method requires minimal labeled evidence data while significantly enhancing both retrieval accuracy and answer generation performance through strategically retrieved in-context tokens in the reasoning chain. Extensive experiments on multiple real-world and counterfactual QA benchmarks demonstrate that our approach substantially outperforms supervised fine-tuning, traditional retrieval-augmented generation methods, and external retrieval solutions. This work represents a fundamental advancement in making LLMs more accurate, reliable, and efficient for knowledge-intensive tasks.

Improving Context Fidelity via Native Retrieval-Augmented Reasoning

Large Vision-Language Models (LVLMs) generate contextually relevant responses by jointly interpreting visual and textual inputs. However, our finding reveals they often mistakenly perceive text inputs lacking visual evidence as being part of the image, leading to erroneous responses. In light of this finding, we probe whether LVLMs possess an internal capability to determine if textual concepts are grounded in the image, and discover a specific subset of Feed-Forward Network (FFN) neurons, termed Visual Absence-aware (VA) neurons, that consistently signal the visual absence through a distinctive activation pattern. Leveraging these patterns, we develop a detection module that systematically classifies whether an input token is visually grounded. Guided by its prediction, we propose a method to refine the outputs by reinterpreting question prompts or replacing the detected absent tokens during generation. Extensive experiments show that our method effectively mitigates the models' tendency to falsely presume the visual presence of text input and its generality across various LVLMs.

Downloads

Next from EMNLP 2025

Humans Hallucinate Too: Language Models Identify and Correct Subjective Annotation Errors With Label-in-a-Haystack Prompts

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES