China

LLM-as-Judge has emerged as a scalable alternative to human evaluation, enabling large language models (LLMs) to provide reward signals in trainings. While recent work has explored multi-agent extensions such as multi-agent debate and meta-judging to enhance evaluation quality, the question of how intrinsic biases manifest in these settings remains underexplored. In this study, we conduct a systematic analysis of four diverse bias types: position bias, verbosity bias, chain-of-thought bias, and bandwagon bias. We evaluate these biases across two widely adopted multi-agent LLM-as-Judge frameworks: Multi-Agent-Debate and LLM-as-Meta-Judge. **Our results show that debate framework amplifies biases sharply after the initial debate, and this increased bias is sustained in subsequent rounds, while meta-judge approaches exhibit greater resistance.** We further investigate the incorporation of PINE, a leading single-agent debiasing method, as a bias-free agent within these systems. **The results reveal that this bias-free agent effectively reduces biases in debate settings but provides less benefit in meta-judge scenarios.** Our work provides a comprehensive study of bias behavior in multi-agent LLM-as-Judge systems and highlights the need for targeted bias mitigation strategies in collaborative evaluation settings.

EMNLP 2025

Judging with Many Minds: Do More Perspectives Mean Less Prejudice? On Bias Amplification and Resistance in Multi-Agent Based LLM-as-Judge

llm-as-judge

intrinsic bias

multi-agent system

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.

We critically assess the efficacy of the current SOTA in hallucination detection and find that its performance on the RAGTruth dataset is largely driven by a spurious correlation with data. Controlling for this effect, state-of-the-art performs no better than supervised linear probes, while requiring extensive hyperparameter tuning across datasets. Out-of-distribution generalization is currently out of reach, with all of the analyzed methods performing close to random. We propose a set of guidelines for hallucination detection and its evaluation.

Representation-based Broad Hallucination Detectors Fail to Generalize Out of Distribution

Jailbreak attacks pose persistent threats to large language models (LLMs). Current safety alignment methods have attempted to address these issues, but they experience two significant limitations: insufficient safety alignment depth and unrobust internal defense mechanisms. These limitations make them vulnerable to adversarial attacks such as prefilling and refusal direction manipulation. We introduce DeepRefusal, a robust safety alignment framework that overcomes these issues. DeepRefusal forces the model to dynamically rebuild its refusal mechanisms from jailbreak states. This is achieved by probabilistically ablating the refusal direction across layers and token depths during fine-tuning. Our method not only defends against prefilling and refusal direction attacks but also demonstrates strong resilience against other unseen jailbreak strategies. Extensive evaluations on four open-source LLM families and six representative attacks show that DeepRefusal reduces attack success rates by approximately 95%, while maintaining model capabilities with minimal performance degradation.

Beyond Surface Alignment: Rebuilding LLMs Safety Mechanism via Probabilistically Ablating Refusal Direction

The rapid proliferation of large language models and multimodal generative models has raised concerns about their potential vulnerabilities to a wide range of real-world safety risks. However, a critical gap persists in systematic assessment, alongside the lack of evaluation frameworks to keep pace with the breadth and variability of real-world risk factors. In this paper, we introduce TroGEN, an automated jailbreak prompt generation framework that assesses these vulnerabilities by deriving scenario-driven jailbreak prompts using an adversarial agent. Moving beyond labor-intensive dataset construction, TroGEN features an extensible design that covers broad range of risks, supports plug-and-play jailbreak strategies, and adapts seamlessly to multimodal settings. Experimental results demonstrate that TroGEN effectively uncovers safety weaknesses, revealing susceptibilities to adversarial attacks that conceal malicious intent beneath an apparently benign facade, like a Trojan horse. Furthermore, such stealthy attacks exhibit resilience even against existing jailbreak defense methods.

Beneath the Facade: Probing Safety Vulnerabilities in LLMs via Auto-Generated Jailbreak Prompts

Topic modeling is a key method in text analysis, but existing approaches fail to efficiently scale to large datasets or are limited by assuming one topic per document. Overcoming these limitations, we introduce Semantic Component Analysis (SCA), a topic modeling technique that discovers multiple topics per sample by introducing a decomposition step to the clustering-based topic modeling framework. We evaluate SCA on Twitter datasets in English, Hausa and Chinese. There, it achieves competetive coherence and diversity compared to BERTopic, while uncovering at least double the topics and maintaining a noise rate close to zero. We also find that SCA outperforms the LLM-based TopicGPT in scenarios with similar compute budgets. SCA thus provides an effective and efficient approach for topic modeling of large datasets.

Semantic Component Analysis: Introducing Multi-Topic Distributions to Clustering-Based Topic Modeling

Large Language Models (LLMs) exhibit significant potential in complex software engineering tasks, however, their fault localization capabilities within repository are constrained by inherent limitations in max context length. Although Test-Time Scaling (TTS) can generate multiple candidate solutions, traditional selection strategies often fail to identify the optimal one. To solve this problem, we introduces Hierarchical Localization Reward Model (HiLoRM), which specifically designed to evaluate and select the most accurate fault localization candidates (at file, function, and line levels) from the multiple sampled outputs of LLMs, thereby enhancing localization accuracy. Furthermore, we constructed the HiFL-44k dataset, comprising approximately 44,000 fault localization instances, to train HiLoRM. Experimental results demonstrate that on the SWE-Bench-Lite dataset, HiLoRM improves the final line-level localization recall by 12% compared to a baseline model that does not use a reward model. Concurrently, HiLoRM exhibits a strong capability to evaluate predictions from larger LLMs (e.g., 32B parameters) and demonstrates transferability and generalization potential when applied to other fault localization methods. This work provides an effective methodology and an accessible model to significantly improve the accuracy and reliability of LLMs for repository-level fault localization.

Hierarchical Reward Modeling for Fault Localization in Large Code Repositories

The flexibility of natural language significantly expands the action space in task-oriented dialogue systems, causing inefficient exploration and slow convergence in deep reinforcement learning (DRL)-based policy optimization. Pre-trained large language models (LLMs), with world knowledge and semantic understanding, offer promising solutions. To this end, we propose LLM-Guided DRL via Semantic-Aware Action Pruning (LLMSAP), a novel framework that synergizes pretrained LLMs with DRL. LLMSAP leverages the world knowledge and contextual understanding of LLMs to guide decision-making via an action feasibility assessment. Instead of requiring LLMs to directly generate optimal actions due to their limited precision in sequential decision tasks, LLMSAP employs a lightweight action pruning mechanism. Specifically, LLMs act as action filters, rapidly eliminating semantically implausible or low-potential actions from multi-turn dialogue context, allowing the DRL agent to focus exploration on a refined candidate subset. This two-stage framework ("prune-then-optimize") avoids extensive LLM fine-tuning while preserving the decision-making precision of DRL. Experiments on multiple benchmarks verify the effectiveness of LLMSAP.

Semantic-Aware Action Space Compression via LLM-DRL Synergy for Efficient Task-oriented Dialogue Policy Exploration

Language models can be steered by modifying their internal representations to control concepts such as emotion, style, or truthfulness in generation. However, the conditions for an effective intervention remain unclear and are often validated through heuristics and trial-and-error. To fill this gap, we demonstrate that intervention efficacy, measured by linear steerability (i.e., the ability to adjust output via linear transformations of hidden states), emerges during intermediate stages of training. Moreover, even closely related concepts (e.g., anger and sadness) exhibit steerability emergence at distinct stages of training*. To better interpret the dynamics of steerability during training, we adapt existing intervention techniques into a unified framework, referred to as the “Intervention Detector” (ID), which is designed to reveal how linear steerability evolves over the course of training through hidden state and representation analysis. ID reveals that concepts become increasingly linearly separable in the hidden space as training progresses, which strongly correlates with the emergence of linear steerability. We further introduce ID-based metrics, such as heatmaps, entropy trends, and cosine similarity, to help interpret how linear steerability evolves throughout training. In addition, we apply ID across different model families to ensure the generality of our findings on steerability dynamics.

Linear Steerability in Language Models: When It Emerges and How It Evolves

Large language models (LLMs) have shown remarkable capabilities in general domains, but their application to multi-omics biology remains underexplored. To address this gap, we introduce Biology-Instructions, the first large-scale instruction-tuning dataset for multi-omics biological sequences, including DNA, RNA, proteins, and multi-molecules. This dataset bridges LLMs and complex biological sequences-related tasks, enhancing their versatility and reasoning while maintaining conversational fluency. We also highlight significant limitations of current state-of-the-art LLMs on multi-omics tasks without specialized training. To overcome this, we propose ChatMultiOmics, a strong baseline with a novel three-stage training pipeline, demonstrating superior biological understanding through Biology-Instructions. Both resources are publicly available, paving the way for better integration of LLMs in multi-omics analysis. The Biology-Instructions is publicly available at \href{https://anonymous.4open.science/r/Biology-Instructions-7D86/}{link}.

Biology-Instructions: A Dataset and Benchmark for Multi-Omics Sequence Understanding Capability of Large Language Models

Large language models (LLMs) have advanced the automation of data science workflows. Yet it remains unclear whether they can critically leverage external domain knowledge as human data scientists do in practice. To answer this question, we introduce AssistedDS (Assisted Data Science), a benchmark designed to systematically evaluate how LLMs handle domain knowledge in tabular prediction tasks. AssistedDS features both synthetic datasets with explicitly known generative mechanisms and real-world Kaggle competitions, each accompanied by curated bundles of helpful and adversarial documents. These documents provide domain-specific insights into data cleaning, feature engineering, and model selection. We assess state-of-the-art LLMs on their ability to discern and apply beneficial versus harmful domain knowledge, evaluating submission validity, information recall, and predictive performance. Our results demonstrate three key findings: (1) LLMs frequently exhibit an uncritical adoption of provided information, significantly impairing their predictive performance when adversarial content is introduced, (2) helpful guidance is often insufficient to counteract the negative influence of adversarial information, and (3) in Kaggle datasets, LLMs often make errors in handling time-series data, applying consistent feature engineering across different folds, and interpreting categorical variables correctly. These findings highlight a substantial gap in current models' ability to critically evaluate and leverage expert knowledge, underscoring an essential research direction for developing more robust, knowledge-aware automated data science systems.

AssistedDS: Benchmarking How External Domain Knowledge Assists LLMs in Automated Data Science

Automatic content moderation is crucial to ensuring safety in social media. Language Model-based classifiers are increasingly adopted for this task, but it has been shown that they perpetuate racial and social biases. Even if several resources and benchmark corpora have been developed to challenge this issue, measuring the fairness of models in content moderation remains an open issue. In this work, we present an unsupervised approach that benchmarks models on the basis of their uncertainty in classifying messages annotated by people belonging to vulnerable groups. We use uncertainty, computed by means of the conformal prediction technique, as a proxy to analyze the bias of 11 models (LMs and LLMs) against women and non-white annotators and observe to what extent it diverges from metrics based on performance, such as the F1 score. The results show that some pre-trained models predict with high accuracy the labels coming from minority groups, even if the confidence in their prediction is low. Therefore, by measuring the confidence of models, we are able to see which groups of annotators are better represented in pre-trained models and lead the debiasing process of these models before their effective use.

Downloads

Next from EMNLP 2025

Representation-based Broad Hallucination Detectors Fail to Generalize Out of Distribution