China

Recent studies have shown that Large Language Models (LLMs) augmented with chain-of-thought (CoT) reasoning demonstrate impressive problem-solving abilities. However, in this work, we identify a recurring issue where these models occasionally generate overly short reasoning, leading to degraded performance on even simple mathematical problems. Specifically, we investigate how reasoning length is embedded in the hidden representations of reasoning models and its impact on accuracy. Our analysis reveals that reasoning length is governed by a linear direction in the representation space, allowing us to induce overly short reasoning by steering the model along this direction. Building on this insight, we introduce ThinkEdit, a simple yet effective weight-editing approach to mitigate the issue of overly short reasoning. We first identify a small subset of attention heads (approximately 4%) that predominantly drive short reasoning behavior. We then edit the output projection weights of these heads to remove the short reasoning direction. With changes to only 0.2% of the model&#39;s parameters, ThinkEdit effectively reduces overly short reasoning and yields notable accuracy gains for short reasoning outputs (+6.39%), along with an overall improvement across multiple math benchmarks (+3.34%). Our findings provide new mechanistic insights into how reasoning length is controlled within LLMs and highlight the potential of fine-grained model interventions to improve reasoning quality.

EMNLP 2025

ThinkEdit: Interpretable Weight Editing to Mitigate Overly Short Thinking in Reasoning Models

Recent studies have shown that Large Language Models (LLMs) augmented with chain-of-thought (CoT) reasoning demonstrate impressive problem-solving abilities. However, in this work, we identify a recurring issue where these models occasionally generate overly short reasoning, leading to degraded performance on even simple mathematical problems. Specifically, we investigate how reasoning length is embedded in the hidden representations of reasoning models and its impact on accuracy. Our analysis reveals that reasoning length is governed by a linear direction in the representation space, allowing us to induce overly short reasoning by steering the model along this direction. Building on this insight, we introduce ThinkEdit, a simple yet effective weight-editing approach to mitigate the issue of overly short reasoning. We first identify a small subset of attention heads (approximately 4%) that predominantly drive short reasoning behavior. We then edit the output projection weights of these heads to remove the short reasoning direction. With changes to only 0.2% of the model's parameters, ThinkEdit effectively reduces overly short reasoning and yields notable accuracy gains for short reasoning outputs (+6.39%), along with an overall improvement across multiple math benchmarks (+3.34%). Our findings provide new mechanistic insights into how reasoning length is controlled within LLMs and highlight the potential of fine-grained model interventions to improve reasoning quality.

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.

Accurately evaluating the word sense disambiguation (WSD) capabilities of large language models (LLMs) remains challenging, as existing studies primarily rely on single-task evaluations and classification-based metrics that overlook the fundamental differences between generative LLMs and traditional classification models. To bridge this gap, we propose \textbf{RoDEval}, the first comprehensive evaluation framework specifically tailored for assessing LLM-based WSD methods. RoDEval introduces four novel metrics: Disambiguation Scope, Disambiguation Robustness, Disambiguation Reliability, and Definition Generation Quality Score, enabling a multifaceted evaluation of LLMs’ WSD capabilities. Experimental results using RoDEval across five mainstream LLMs uncover significant limitations in their WSD performance. Specifically, incorrect definition selections in multiple-choice WSD tasks stem not from simple neglect or forget of correct options, but rather from incomplete acquisition of the all senses for polysemous words. Instead, disambiguation reliability is often compromised by the models' persistent overconfidence. In addition, inherent biases continue to affect performance, and scaling up model parameters alone fails to meaningfully enhance their ability to generate accurate sense definitions. These findings provide actionable insights for enhancing LLMs' WSD capabilities. The source code and evaluation scripts are open-sourced at https://github.com/DayDream405/RoDEval.

RoDEval: A Robust Word Sense Disambiguation Evaluation Framework for Large Language Models

During sudden disaster events, accurately predicting public panic sentiment on social media is crucial for proactive governance and crisis management. Current efforts on this problem face three main challenges: lack of finely annotated data hinders emotion prediction studies, unmodeled risk perception causes prediction inaccuracies, and insufficient interpretability of panic formation mechanisms. We address these issues by proposing a Psychology-driven generative Agent framework (PsychoAgent) for explainable panic prediction based on emotion arousal theory. Specifically, we first construct a fine-grained open panic emotion dataset (namely COPE) via human-large language models (LLMs) collaboration to mitigate semantic bias. Then, we develop a framework integrating cross-domain heterogeneous data grounded in psychological mechanisms to model risk perception and cognitive differences in emotion generation. To enhance interpretability, we design an LLM-based role-playing agent that simulates individual psychological chains through dedicatedly designed prompts. Experimental results on our annotated dataset show that PsychoAgent improves panic emotion prediction performance by 12.6% to 21.7% compared to baseline models. Furthermore, the explainability and generalization of our approach is validated. Crucially, this represents a paradigm shift from opaque "data-driven fitting" to transparent "role-based simulation with mechanistic interpretation" for panic emotion prediction during emergencies. Our implementation is publicly available at: https://anonymous.4open.science/r/PsychoAgent-19DD.

PychoAgent: A Psychology-driven LLM Agents for Explainable Panic Prediction on Social Media during Sudden Disaster Events

Conversational aspect-based sentiment quadruple analysis (DiaASQ) is a newly-emergent task aiming to extract quadruples of target-aspect-opinion-sentiment from a conversation text. Existing studies struggle to capture complete dialogue semantics, largely due to inadequate inter-utterance modeling and the underutilization of dialogue structure. To address these issues, we propose an Inter-sentence Context Modeling and Structure-aware Representation Enhancement model (ICMSR) to extract dialogue aspect sentiment quadruple. We design the Dialog Inter-sentence Contextual Enhancer (DICE) module after the sentence-by-sentence encoding phase to enhance inter-sentence interactions and mitigate contextual fragmentation caused by traditional sequential encoding. Moreover, to fully exploit structural information within dialogues, we propose the Dialog Feature Amplifier (DFA), which consists of two submodules: STREAM and SMM. The STREAM module integrates diverse structural dialogue information to generate structure-aware sentence representations, effectively improving the modeling of intra-dialogue structural relations. Furthermore, the Structural Multi-scale Mechanism (SMM) employs a multi-scale modeling approach, simulating varying extents of contextual awareness, thereby enhancing the model’s ability to capture cross-sentence structural dependencies. We extensively evaluate our method on benchmark datasets, and the empirical results consistently confirm its effectiveness.

Inter-sentence Context Modeling and Structure-aware Representation Enhancement for Conversational Sentiment Quadruple Extraction

Recent advances in causal interpretability have extended from language models to vision-language models (VLMs), seeking to reveal their internal mechanisms through input interventions. While textual interventions often target semantics, visual interventions typically rely on coarse pixel-level perturbations, limiting semantic insights on multimodal integration. In this study, we introduce V-SEAM, a novel framework that combines Visual Semantic Editing and Attention Modulating for causal interpretation of VLMs. V-SEAM enables concept-level visual manipulations and identifies attention heads with positive or negative contributions to predictions across three semantic levels: objects, attributes, and relationships. We observe that positive heads are often shared within the same semantic level but vary across levels, while negative heads tend to generalize broadly. Finally, we introduce an automatic method to modulate key head embeddings, demonstrating enhanced performance for both LLAVA and InstructBLIP across three diverse VQA benchmarks.

V-SEAM: Visual Semantic Editing and Attention Modulating for Causal Interpretability of Vision-Language Models

Recent advances in test-time scaling have enabled Large Language Models (LLMs) to display sophisticated reasoning abilities via extended Chain-of-Thought (CoT) generation. Despite their impressive reasoning abilities, Reasoning LLMs (RLMs) frequently display unstable behaviors, e.g., hallucinating unsupported premises, overthinking simple tasks, and displaying higher sensitivity to prompt variations. This raises a deeper research question: *How can we represent the reasoning process of RLMs to map their minds?* To address this, we propose a unified graph-based analytical framework for fine-grained modeling and quantitative analysis of RLM reasoning dynamics. Our method first clusters long, verbose CoT outputs into semantically coherent reasoning steps, then constructs directed reasoning graphs to capture contextual and logical dependencies among these steps. Through a comprehensive analysis of derived reasoning graphs, we also reveal that key structural properties, such as exploration density, branching, and convergence ratios, strongly correlate with models' performance. The proposed framework enables quantitative evaluation of internal reasoning structure and quality beyond conventional metrics and also provides practical insights for prompt engineering and cognitive analysis of LLMs. Code and resources will be released to facilitate future research in this direction.

Mapping the Minds of LLMs: A Graph-Based Analysis of Reasoning LLMs

Reasoning-focused large language models (LLMs) are rapidly evolving across various domains, yet their capabilities in handling complex legal problems remains underexplored. In this paper, we introduce Unilaw-R1, a large language model tailored for legal reasoning. With a lightweight 7-billion parameter scale, Unilaw-R1 significantly reduces deployment cost while effectively tackling three core challenges in the legal domain: insufficient legal knowledge, unreliable reasoning logic, and weak business generalization. To enhance its reasoning capability, we first construct Unilaw-R1-Data, a high-quality dataset containing sim17K distilled and screened chain-of-thought (CoT) samples. Based on this, we adopt a two-stage training strategy combining Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL), which significantly boosts the model’s performance on complex legal reasoning tasks and supports interpretable decision-making in legal AI applications. To assess legal reasoning ability, we also introduce Unilaw-R1-Eval, a dedicated benchmark designed to evaluate models across single- and multi-choice legal tasks. Unilaw-R1 demonstrates strong results on authoritative benchmarks, outperforming all models of similar scale and achieving performance on par with the much larger DeepSeek-R1-Distill-Qwen-32B (54.9). Following domain-specific training, it also showed significant gains on LawBench and LexEval, exceeding Qwen-2.5-7B-Instruct (46.6) by an average margin of 6.6 points.

Unilaw-R1: A Large Language Model for Legal Reasoning with Reinforcement Learning and Iterative Inference

Tokenization is a crucial step that bridges human-readable text with model-readable discrete tokens. However, recent studies have revealed that tokenizers can be exploited to elicit unwanted model behaviors. In this work, we investigate incomplete tokens, i.e., undecodable tokens with stray bytes resulting from byte-level byte-pair encoding (BPE) tokenization. We hypothesize that such tokens are heavily reliant on their adjacent tokens and are fragile when paired with unfamiliar tokens. To demonstrate this vulnerability, we introduce improbable bigrams: out-of-distribution combinations of incomplete tokens designed to exploit their dependency. Our experiments show that improbable bigrams are significantly prone to hallucinatory behaviors. Surprisingly, the same phrases have drastically lower rates of hallucination (90% reduction in Llama3.1) when an alternative tokenization is used. We caution against the potential vulnerabilities introduced by byte-level BPE tokenizers, which may introduce blind spots to language models.

Improbable Bigrams Expose Vulnerabilities of Incomplete Tokens in Byte-Level Tokenizers

Concept editing aims to control specific concepts in large language models (LLMs) and is an emerging subfield of model editing. Despite the emergence of various editing methods in recent years, there remains a lack of rigorous theoretical analysis and a unified perspective to systematically understand and compare these methods. To address this gap, we propose a unified paradigm for concept editing methods, in which all forms of conceptual injection are aligned at the neuron level. We study four representative concept editing methods: Neuron Editing (NE), Supervised Fine-tuning (SFT), Sparse Autoencoder (SAE), and Steering Vector (SV). Then we categorize them into two classes based on their mode of conceptual information injection: indirect (NE, SFT) and direct (SAE, SV). We evaluate above methods along four dimensions: editing reliability, output generalization, neuron level consistency, and mathematical formalization. Experiments show that SAE achieves the best editing reliability. In output generalization, SAE captures features closer to human-understood concepts, while NE tends to locate text patterns rather than true semantics. Neuron-level analysis reveals that direct methods share high neuron overlap, as do indirect methods, indicating methodological commonality within each category. Our unified paradigm offers a clear framework and valuable insights for advancing interpretability and controlled generation in LLMs.

Towards a Unified Paradigm of Concept Editing in Large Language Models

The quality of data plays a vital role in the development of Large-scale Generative Models. Understanding how important a data point is for a generative model is essential for explaining its behavior and improving the performance. The influence function provides a framework for quantifying the impact of individual training data on model predictions. However, the high computational cost has hindered their applicability in large-scale applications. In this work, we present RRInf, a novel and principled method for estimating influence function in large-scale generative AI models. We show that influence function estimation can be transformed into a ridge regression problem. Based on this insight, we develop an algorithm that is efficient and scalable to large models. Experiments on noisy data detection and influential data identification tasks demonstrate that RRInf outperforms existing methods in terms of both efficiency and effectiveness for three commonly used large models: RoBERTa-large, Llama-2-13B-chat, and stable-diffusion-v1.5.

RRInf: Efficient Influence Function Estimation via Ridge Regression for Large Language Models and Text-to-Image Diffusion Models

Sparse AutoEncoders (SAEs) have recently been employed as a promising unsupervised approach for understanding the representations of layers of Large Language Models (LLMs). However, with the growth in model size and complexity, training SAEs is computationally intensive, as typically one SAE is trained for each model layer. To address such limitation, we propose \textit{Group-SAE}, a novel strategy to train SAEs. Our method considers the similarity of the residual stream representations between contiguous layers to group similar layers and train a single SAE per group. To balance the trade-off between efficiency and performance, we further introduce \textit{AMAD} (Average Maximum Angular Distance), an empirical metric that guides the selection of an optimal number of groups based on representational similarity across layers. Experiments on models from the Pythia family show that our approach significantly accelerates training with minimal impact on reconstruction quality and comparable downstream task performance and interpretability over baseline SAEs trained layer by layer. This method provides an efficient and scalable strategy for training SAEs in modern LLMs.

Downloads

Next from EMNLP 2025

RoDEval: A Robust Word Sense Disambiguation Evaluation Framework for Large Language Models

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES