China

Recent advances in large reasoning models (LRMs) have enabled strong multi-step reasoning capabilities. However, existing machine unlearning algorithms are tailored to standard language modeling and fail to address the unique challenges posed by LRMs. In this work, we present the first systematic study of LRM unlearning and reveal that conventional unlearning methods often overlook critical information leakage in reasoning traces, even when final answers are successfully removed. To address this, we propose Reasoning-aware Representation Misdirection for Unlearning (R²MU), a method that suppresses sensitive reasoning traces while preserving the model’s general reasoning ability. Our experiments demonstrate that R²MU significantly reduces reasoning trace leakage and achieves strong performance across both reasoning and safety benchmarks, including WMDP, StrongReject, JBB-Behaviors and WildJailbreak, under state-of-the-art models such as DeepSeek-R1-Distill-LLaMA-8B and DeepSeek-R1-Distill-Qwen-14B. To the best of our knowledge, MU is the first principled approach to both expose and mitigate reasoning trace leakage in LRM unlearning, while preserving reasoning ability.

EMNLP 2025

Reasoning Model Unlearning: Forgetting Traces, Not Just Answers, While Preserving Reasoning Skills

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.

While densely annotated image captions significantly facilitate the learning of robust vision-language alignment, methodologies for systematically optimizing human annotation efforts remain underexplored. We introduce Chain-of-Talkers (CoTalk), an AI-in-the-loop methodology designed to maximize the number of annotated samples and improve their comprehensiveness under fixed budget constraints (e.g., total human annotation time). The framework is built upon two key insights. First, sequential annotation reduces redundant workload compared to conventional parallel annotation, as subsequent annotators only need to annotate the ``residual''---the missing visual information that previous annotations have not covered. Second, humans process textual input faster by reading while outputting annotations with much higher throughput via talking; thus a multimodal interface enables optimized efficiency. We evaluate our framework from two aspects: intrinsic evaluations that assess the comprehensiveness of semantic units, obtained by parsing detailed captions into object-attribute trees and analyzing their effective connections; extrinsic evaluation measures the practical usage of the annotated captions in facilitating vision-language alignment. Experiments with eight participants show our Chain-of-Talkers (CoTalk) improves annotation speed (0.42 vs. 0.30 units/sec) and retrieval performance (41.13% vs. 40.52%) over the parallel method.

Chain-of-Talkers (CoTalk): Fast Human Annotation of Dense Image Captions

The development of Emotional Support Conversation (ESC) systems is critical for delivering mental health support tailored to the needs of help-seekers. Recent advances in large language models (LLMs) have contributed to progress in this domain, while most existing studies focus on generating responses directly and overlook the integration of domain-specific reasoning and expert interaction. Therefore, in this paper, we propose a training-free \underline{Multi}-\underline{Agent} collaboration framework for \underline{ESC} ({MultiAgentESC}). The framework is designed to emulate the human-like process of providing emotional support through three stages: dialogue analysis, strategy deliberation, and response generation. At each stage, a multi-agent system is employed to iteratively enhance information understanding and reasoning, simulating real-world decision-making processes by incorporating diverse interactions among these expert agents. Additionally, we introduce a novel response-centered approach to handle the one-to-many problem on strategy selection, where multiple valid strategies are initially employed to generate diverse responses, followed by the selection of the optimal response through multi-agent collaboration. Experiments on the ESConv dataset reveal that our proposed framework excels at providing emotional support as well as diversifying support strategy selection.

MultiAgentESC: A LLM-based Multi-Agent Collaboration Framework for Emotional Support Conversation

Traditional robot simulators focus on physical process modeling and realistic rendering, often suffering from high computational costs, inefficiencies, and limited adaptability. To handle this issue, we concentrate on behavior simulation in robotics to analyze and validate the logic behind robot behaviors, aiming to achieve preliminary evaluation before deploying resource-intensive simulators and thus enhance simulation efficiency. In this paper, we propose BeSimulator, a modular and novel LLM-powered framework, as an attempt towards behavior simulation in the context of text-based environments. By constructing text-based virtual environments and performing semantic-level simulation, BeSimulator can generalize across scenarios and achieve long-horizon complex simulation. Inspired by human cognition paradigm, it employs a ``consider-decide-capture-transfer'' four-phase simulation process, termed Chain of Behavior Simulation (CBS), which excels at analyzing action feasibility and state transition. Additionally, BeSimulator incorporates code-driven reasoning to enable arithmetic operations and enhance reliability, and reflective feedback to refine simulation. Based on our manually constructed behavior-tree-based simulation benchmark, BTSIMBENCH, our experiments show a significant performance improvement in behavior simulation compared to baselines, ranging from 13.60% to 24.80%. Code and data are available at https://anonymous.4open.science/r/BeSimulator-4A62/.

BeSimulator: A Large Language Model Powered Text-based Behavior Simulator

Long chain-of-thought (CoT) supervision has become a common strategy to enhance reasoning in language models. While effective for large models, we identify a phenomenon we call Long CoT Degradation, in which small language models (SLMs; leq3B parameters) trained on limited long CoT data experience significant performance deterioration. Through extensive experiments on the Qwen2.5, LLaMA3 and Gemma3 families, we demonstrate that this degradation is widespread across SLMs. In some settings, models trained on only 8k long CoT examples lose up to 75% of their original performance before fine-tuning. Strikingly, we further observe that for some particularly small models, even training on 220k long CoT examples fails to recover or surpass their original performance prior to fine-tuning. Our analysis attributes this effect to error accumulation: while longer responses increase the capacity for multi-step reasoning, they also amplify the risk of compounding mistakes. Furthermore, we find that Long CoT Degradation may negatively impacts downstream reinforcement learning (RL), although this can be alleviated by sufficiently scaled supervised fine-tuning (SFT). Our findings challenge common assumptions about the benefits of long CoT training for SLMs and offer practical guidance for building more effective small-scale reasoning models.

Through the Valley: Path to Effective Long CoT Training for Small Language Models

LLMs have shown strong performance on human-centric reasoning tasks. While previous evaluations have explored whether LLMs can infer intentions or detect deception, they often overlook the individualized reasoning styles that influence how people interpret and act in social contexts. Social deduction games (SDGs) provide a natural testbed for evaluating individualized reasoning styles, where different players may adopt diverse but contextually valid reasoning strategies under identical conditions. To address this, we introduce InMind, a cognitively grounded evaluation framework designed to assess whether LLMs can capture and apply personalized reasoning styles in SDGs. InMind enhances structured gameplay data with round-level strategy traces and post-game reflections, collected under both Observer and Participant modes. It supports four cognitively motivated tasks that jointly evaluate both static alignment and dynamic adaptation. As a case study, we apply InMind to the game Avalon, evaluating 11 state-of-the-art LLMs. General-purpose LLMs, even GPT-4o frequently rely on lexical cues, struggling to anchor reflections in temporal gameplay or adapt to evolving strategies. In contrast, reasoning-enhanced LLMs like DeepSeek-R1 exhibit early signs of style-sensitive reasoning. These findings reveal key limitations in current LLMs’ capacity for individualized, adaptive reasoning, and position InMind as a step toward cognitively aligned human–AI interaction.

InMind: Evaluating LLMs in Capturing and Applying Individual Human Reasoning Styles

The social impact of Natural Language Processing (NLP) is increasingly important, with a rising community focus on initiatives related to NLP for Social Good (NLP4SG). Indeed, in recent years, almost 20% of all papers in the ACL Anthology address topics related to social good as defined by the UN Sustainable Development Goals (Adauto et al., 2023). In this study we take an author- and venue-level perspective to map the landscape of NLP4SG, quantifying the proportion of work addressing social good concerns both within and beyond the ACL community, by both core ACL contributors and non-ACL authors. With this approach we discover two surprising facts about the landscape of NLP4SG. First, ACL authors are dramatically more likely to do work addressing social good concerns when publishing in venues outside of ACL. Second, the vast majority greater than 80% – of publications using NLP techniques to address concerns of social good are done by non-ACL authors in venues outside of ACL. We discuss the implications of these findings on agenda-setting considerations for the ACL community related to NLP4SG.

Good Intentions Beyond ACL: Who Does NLP for Social Good, and Where?

**Warning: This paper contains content that may involve potentially harmful behaviours, discussed strictly for research purposes.** Jailbreak attacks can hinder the safety of Large Language Model (LLM) applications, especially chatbots. Studying jailbreak techniques is an important AI red teaming task for improving the safety of these applications. In this paper, we introduce TombRaider, a novel jailbreak technique that exploits the ability to store, retrieve, and use historical knowledge of LLMs. TombRaider employs two agents, the inspector agent to extract relevant historical information and the attacker agent to generate adversarial prompts, enabling effective bypassing of safety filters. We intensively evaluated TombRaider on six popular models. Experimental results showed that TombRaider could outperform state-of-the-art jailbreak techniques, achieving nearly 100% attack success rates (ASRs) on bare models and maintaining over 55.4% ASR against defence mechanisms. Our findings highlight critical vulnerabilities in existing LLM safeguards, underscoring the need for more robust safety defences.

TombRaider: Entering the Vault of History to Jailbreak Large Language Models

Large language models (LLMs) can produce fluent but factually incorrect outputs, and has limited ability to attribute their claims to source material. This undermines their reliability, particularly in multi-hop and high-stakes domains such as medicine. We propose Tree-of-Quote (ToQ) prompting, a structured framework that decomposes complex questions into subquestions, generates quotes to support each step, and selectively advances reasoning based on quote quality. We also introduce FQ-Score, a unified metric that captures answer correctness, attribution fidelity, and reasoning quality. Experiments on StrategyQA, 2WikiMultiHopQA, MuSiQue, MoreHopQA, and MedQA demonstrate that ToQ improves factuality and attribution over standard prompting baselines. To validate FQ-Score as a proxy for human judgment, we conduct two reader studies with clinicians on medical questions, and observe strong correlations. Both clinician scores and FQ-Scores also indicate a preference for ToQ over baselines due to a combination of greater correctness, completeness, and logical flow. Our results suggest ToQ is a promising approach for building more trustworthy and auditable LLM systems. Code to run ToQ, FQ-Score, and prompt templates will be made available on GitHub.

Tree-of-Quote Prompting Improves Factuality and Attribution in Multi-Hop and Medical Reasoning

Misinformation spans various domains, but detection methods trained on specific domains often perform poorly when applied to others. With the rapid development of Large Language Models (LLMs), researchers have begun to utilize LLMs for cross-domain misinformation detection. However, existing LLM-based methods often fail to adequately analyze news in the target domain, limiting their detection capabilities. More importantly, these methods typically rely on manually designed decision rules, which are limited by domain knowledge and expert experience, thus limiting the generalizability of decision rules to different domains. To address these issues, we propose a Multi-Agent Framework for cross-domain misinformation detection with Automated Decision Rule Optimization (MARO). Under this framework, we first employs multiple expert agents to analyze target-domain news. Subsequently, we introduce a question-reflection mechanism that guides expert agents to facilitate higher-quality analysis. Furthermore, we propose a decision rule optimization approach based on carefully designed cross-domain validation tasks to iteratively enhance decision rule effectiveness across domains. Experimental results and analysis on commonly used datasets demonstrate that MARO achieves significant improvements over existing methods.

A Multi-Agent Framework with Automated Decision Rule Optimization for Cross-Domain Misinformation Detection

The rapid advancement of large language models has intensified public concerns about the potential misuse. Therefore, it is important to build trustworthy AI-generated text detection systems. Existing methods neglect stylistic modeling and mostly rely on static thresholds, which greatly limits the detection performance. In this paper, we propose the Mixture of Stylistic Experts (MoSEs) framework that enables stylistics-aware uncertainty quantification through conditional threshold estimation. MoSEs contain three core components, namely, the Stylistics Reference Repository (SRR), the Stylistics-Aware Router (SAR), and the Conditional Threshold Estimator (CTE). For input text, SRR can activate the appropriate reference data in SRR and provide them to CTE. Subsequently, CTE jointly models the linguistic statistical properties and semantic features to dynamically determine the optimal threshold. With a discrimination score, MoSEs yields prediction labels with the corresponding confidence level. Our framework achieves an average improvement 11.34% in detection performance compared to baselines. More inspiringly, MoSEs shows a more evident improvement 39.15% in the low-resource case.

Downloads

Next from EMNLP 2025

Chain-of-Talkers (CoTalk): Fast Human Annotation of Dense Image Captions

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES