China

Recent advances in Large Language Models (LLMs) have enabled them to process increasingly longer sequences, ranging from 2K to 2M tokens and even beyond. However, simply extending the input sequence length does not necessarily lead to effective long-context understanding. In this study, we integrate Chain-of-Thought (CoT) reasoning into LLMs in a supervised manner to facilitate effective long-context understanding. To achieve this, we introduce LongFinanceQA, a synthetic dataset in the financial domain designed to improve long-context reasoning. Unlike existing long-context synthetic data, LongFinanceQA includes intermediate CoT reasoning before the final conclusion, which encourages LLMs to perform explicit reasoning, improving accuracy and interpretability in long-context understanding. To generate synthetic CoT reasoning, we propose Property-based Agentic Inference (PAI), an agentic framework that simulates human-like reasoning steps, including property extraction, retrieval, and summarization. We evaluate PAI&#39;s reasoning capabilities by assessing GPT-4o-mini w/ PAI on the Loong benchmark, outperforming standard GPT-4o-mini by 20.0%. Furthermore, we fine-tune LLaMA-3.1-8B-Instruct on LongFinanceQA, achieving a 28.0% gain on Loong&#39;s financial subset.

EMNLP 2025

Facilitating Long Context Understanding via Supervised Chain-of-Thought Reasoning

supervised chain-of-thought reasoning

long-context synthetic data

long context understanding

## 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*





<br>

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

Recent advances in Large Language Models (LLMs) have enabled them to process increasingly longer sequences, ranging from 2K to 2M tokens and even beyond. However, simply extending the input sequence length does not necessarily lead to effective long-context understanding. In this study, we integrate Chain-of-Thought (CoT) reasoning into LLMs in a supervised manner to facilitate effective long-context understanding. To achieve this, we introduce LongFinanceQA, a synthetic dataset in the financial domain designed to improve long-context reasoning. Unlike existing long-context synthetic data, LongFinanceQA includes intermediate CoT reasoning before the final conclusion, which encourages LLMs to perform explicit reasoning, improving accuracy and interpretability in long-context understanding. To generate synthetic CoT reasoning, we propose Property-based Agentic Inference (PAI), an agentic framework that simulates human-like reasoning steps, including property extraction, retrieval, and summarization. We evaluate PAI's reasoning capabilities by assessing GPT-4o-mini w/ PAI on the Loong benchmark, outperforming standard GPT-4o-mini by 20.0%. Furthermore, we fine-tune LLaMA-3.1-8B-Instruct on LongFinanceQA, achieving a 28.0% gain on Loong's financial subset.

poster

Sycophancy causes models to produce answers that cater to user expectations rather than providing truthful responses. Sycophantic behavior in models can erode user trust by creating a perception of dishonesty or bias. This lack of authenticity may lead users to question the reliability and objectivity of the system's responses. Although Reinforcement Learning from Human Feedback (RLHF) is effective in aligning models with human preferences, previous studies have observed that it can simultaneously amplify sycophantic behavior. However, these studies primarily focused on proprietary models and employed indirect analysis to demonstrate the influence of human feedback. Our study focuses on sycophancy in open-source models, which are more reproducible and transparent for research. We investigated the impact of human feedback on sycophancy by directly comparing models aligned with human feedback to those not aligned. To address sycophancy, we proposed assessing the user's expected answer rather than ignoring it. Consequently, we developed the Sycophancy Answer Assessment (SAA) dataset and introduced Self-Augmented Preference Alignment, demonstrating that these methods effectively enhance the model's assessment ability and significantly reduce sycophancy across tasks.

Self-Augmented Preference Alignment for Sycophancy Reduction in LLMs

Large language models (LLMs) are now ubiquitous in user-facing applications, yet they still generate undesirable toxic outputs, including profanity, vulgarity, and derogatory remarks. Although numerous detoxification methods exist, most apply broad, surface-level fixes and can therefore easily be circumvented by jailbreak attacks. In this paper we leverage sparse autoencoders (SAEs) to identify toxicity-related directions in the residual stream of models and perform targeted activation steering using the corresponding decoder vectors. We introduce three tiers of steering aggressiveness and evaluate them on GPT-2 Small and Gemma-2-2B, revealing trade-offs between toxicity reduction and language fluency. At stronger steering strengths, these causal interventions surpass competitive baselines in reducing toxicity by up to 20%, though fluency can degrade noticeably on GPT-2 Small depending on the aggressiveness. Crucially, standard NLP benchmark scores upon steering remain stable, indicating that the model’s knowledge and general abilities are preserved. We further show that feature-splitting in wider SAEs hampers safety interventions, underscoring the importance of disentangled feature learning. Our findings highlight both the promise and the current limitations of SAE-based causal interventions for LLM detoxification, further suggesting practical guidelines for safer language-model deployment.

Breaking Bad Tokens: Detoxification of LLMs Using Sparse Autoencoders

Large language models (LLMs) are widely adapted for downstream applications through fine-tuning, a process named customization. However, recent studies have identified a vulnerability during this process, where malicious samples can compromise the robustness of LLMs and amplify harmful behaviors. To address this challenge, we propose an adaptive data curation approach allowing any text to be curated to enhance its effectiveness in counteracting harmful samples during customization. To avoid the need for additional defensive modules, we further introduce a comprehensive mitigation framework spanning the lifecycle of the customization process: before customization to immunize LLMs against future compromise attempts, during customization to neutralize risks, and after customization to restore compromised models. Experimental results demonstrate a significant reduction in compromising effects, achieving up to a 100% success rate in generating safe responses. By combining adaptive data curation with lifecycle-based mitigation strategies, this work represents a solid step forward in mitigating compromising risks and ensuring the secure adaptation of LLMs.

Data to Defense: The Role of Curation in Aligning Large Language Models Against Safety Compromise

Large language models (LLMs) have shown significant potential to improve diagnostic performance for healthcare professionals. Existing multi-agent paradigms rely mainly on prompt engineering, suffering from improper agent selection and insufficient knowledge integration. In this work, we propose a novel framework KACR (Knowledge-Aware Co-Reasoning) that integrates structured knowledge reasoning into multidisciplinary collaboration from two aspects: (1) a reinforcement learning-optimized agent that uses clinical knowledge graphs to guide dynamic discipline determination; (2) a multidisciplinary collaboration strategy that enables robust consensus through integration of domain-specific expertise and interdisciplinary persuasion mechanism. Extensive experiments conducted on both academic and real-world datasets demonstrate the effectiveness of our method.

Knowledge-Aware Co-Reasoning for Multidisciplinary Collaboration

Benchmark data contamination (BDC) silently inflate the evaluation performance of large language models (LLMs), yet current work on BDC has centered on direct token overlap (data/label level), leaving the subtler and equally harmful semantic level BDC largely unexplored. This gap is critical in fake news detection task, where prior exposure to semantic BDC lets a model “remember” the answer instead of reasoning. We (1) are the first to formally defined semantic contamination for this task and (2) introduced the Semantic Sensitivity Amplifier (SSA)—a lightweight, model-agnostic framework that detect BDC risks across semantic to label level via an entity shift perturbation and a comprehensive interpretable metric, the SSA Factor. Evaluating 45 variants of nine LLMs (0.5B–72B parameters) across four BDC levels, we find LIAR2 accuracy climbs monotonically with injected contamination, while the SSA Factor escalates in near-perfect lock-step (rgeq.97, for models geq3B, p<.05; ρ geq.9 overall, p<.05). These results show that SSA provides a sensitive, scalable audit of comprehensive BDC risk and paves the way for more integrity evaluation of LLM-driven fake news detection task.

SSA: Semantic Contamination of LLM-Driven Fake News Detection

With the increasing use of Retrieval-Augmented Generation (RAG), strong retrieval models have become more important than ever. In healthcare, multimodal retrieval models that combine information from both text and images offer major advantages for many downstream tasks such as question answering, cross-modal retrieval, and multimodal summarization, since medical data often includes both formats. However, there is currently no standard benchmark to evaluate how well these models perform in medical settings. To address this gap, we introduce M3Retrieve, a Multimodal Medical Retrieval Benchmark. M3Retrieve spans 5 domains,16 medical fields, and 4 distinct tasks, with over 1.2 Million text documents and 164K multimodal queries, all collected under approved licenses. We evaluate leading multimodal retrieval models on this benchmark to explore the challenges specific to different medical specialities and to understand their impact on retrieval performance. By releasing M3Retrieve, we aim to enable systematic evaluation, foster model innovation, and accelerate research toward building more capable and reliable multimodal retrieval systems for medical applications.

M3Retrieve: Benchmarking Multimodal Retrieval for Medicine

We introduce CoDe-KG, an open-source, end-to-end pipeline for extracting sentence-level knowledge graphs by combining robust coreference resolution with syntactic sentence decomposition. Using our model, we contribute a dataset of over 150 000 knowledge triples, which is open source. We also contribute a training corpus of 7248 rows for sentence complexity, 200 rows of gold human annotations for coreference resolution using lung-cancer abstracts from PubMed, 900 rows of gold human annotations for sentence conversion policies from sentences in the abstract, and 398 triples of gold human annotations. We systematically select optimal prompt-model pairs across five complexity categories, showing that hybrid chain-of-thought and few-shot prompting yields up to 99.8% exact-match accuracy on sentence simplification. On relation extraction (RE), our pipeline achieves 65.8% macro-F1 on REBEL, an 8-point gain over the prior state of the art, and 75.7% micro-F1 on WebNLG2, while matching or exceeding performance on Wiki-NRE and CaRB. Ablation studies demonstrate that integrating coreference and decomposition increases recall on rare relations by over 20%

Automated Knowledge Graph Construction using Large Language Models and Sentence Complexity Modelling

This paper presents a comprehensive analysis of the linguistic diversity of LLM safety research, highlighting the English-centric nature of the field. Through a systematic review of nearly 300 publications from 2020--2024 across major NLP conferences and workshops at ACL, we identify a significant and growing language gap in LLM safety research, with even high-resource non-English languages receiving minimal attention. We further observe that non-English languages are rarely studied as a standalone language and that English safety research exhibits poor language documentation practice. To motivate future research into multilingual safety, we make several recommendations based on our survey, and we then pose three concrete future directions on safety evaluation, training data generation, and crosslingual safety generalization. Based on our survey and proposed directions, the field can develop more robust, inclusive AI safety practices for diverse global populations.

The State of Multilingual LLM Safety Research: From Measuring The Language Gap To Mitigating It

Large Language Models (LLMs) have been explored for automating or enhancing penetration testing tasks, but their effectiveness and reliability across diverse attack phases remain open questions. This study presents a comprehensive evaluation of multiple LLM-based agents—from singular to modular—across realistic penetration testing scenarios, analyzing their empirical performance and recurring failure patterns. We further investigate the impact of core functional capabilities on agent success, operationalized through five targeted augmentations: Global Context Memory (GCM), Inter-Agent Messaging (IAM), Context-Conditioned Invocation (CCI), Adaptive Planning (AP), and Real-Time Monitoring (RTM). These interventions respectively support the capabilities of Context Coherence & Retention, Inter-Component Coordination & State Management, Tool Usage Accuracy & Selective Execution, Multi-Step Strategic Planning & Error Detection & Recovery, and Real-Time Dynamic Responsiveness. Our findings reveal that while some architectures natively exhibit select properties, targeted augmentations significantly enhance modular agent performance—particularly in complex, multi-step, and real-time penetration testing scenarios.

From Capabilities to Performance: Evaluating Key Functional Properties of LLM Architectures in Penetration Testing

Large Language Models (LLMs) increasingly power mental-health chatbots, yet the field still lacks a scalable, theory-grounded way to decide which model is more effective to deploy. We present ESC-Judge, the first end-to-end evaluation framework that (i) grounds head-to-head comparison of Emotional-Support LLMs (ES-LLMs) in an established psychological theory—Clara Hill’s Exploration–Insight–Action (E-I-A) counselling model—thereby delivering a structured, interpretable lens on performance, and (ii) fully automates the pipeline at scale. ESC-Judge proceeds in three stages: (1) it synthesizes realistic help-seeker roles by sampling empirically salient attributes (stressors, personality, life history); (2) it has two candidate ES-Agents conduct separate sessions with the same role, isolating model-specific strategies; and (3) it asks a specialised judge LLM to issue pairwise preferences across rubric-anchored skills that exhaustively cover the E-I-A spectrum. In our empirical study, ESC-Judge matches PhD-level annotators in 85% of Exploration, 83% of Insight, and 86% of Action decisions, demonstrating human-level reliability at a fraction of the cost. We release all code, prompts, synthetic roles, transcripts, and judgment scripts to catalyze transparent progress in emotionally supportive AI

Downloads

Next from EMNLP 2025

Self-Augmented Preference Alignment for Sycophancy Reduction in LLMs

Similar lecture

Long Chain-of-Thought Fine-tuning via Understanding-to-Reasoning Transition

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES