China

The use of large language models (LLMs) as judges, particularly in preference comparisons, has become widespread, but this reveals a notable bias towards longer responses, undermining the reliability of such evaluations. To better understand such bias, we propose to decompose the preference evaluation metric, specifically the win rate, into two key components: desirability and information mass, where the former is length-independent and related to trustworthiness such as correctness, toxicity, and consistency, and the latter is length-dependent and represents the amount of information in the response. We empirically demonstrated the decomposition through controlled experiments and found that response length impacts evaluations by influencing information mass. To derive a reliable evaluation metric that assesses content quality without being confounded by response length, we propose AdapAlpaca, a simple yet effective adjustment to win rate measurement. Specifically, AdapAlpaca ensures a fair comparison of response quality by aligning the lengths of reference and test model responses under equivalent length intervals.

EMNLP 2025

Explaining Length Bias in LLM-Based Preference Evaluations

preference evaluation

llm evaluation

length bias

## 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.

poster

Large language models (LLMs) exhibit exceptional capabilities across various tasks but also pose risks by generating harmful content. Existing safety mechanisms, while improving model safety, often lead to overly cautious behavior and fail to fully leverage LLMs’ internal cognitive processes. Inspired by humans' reflective thinking capability, we first show that LLMs can similarly perform internal assessments about safety in their internal states. Building on this insight, we propose **SafeSwitch**, a dynamic framework that regulates unsafe outputs by utilizing the prober-based internal state monitor that actively detects harmful intentions, and activates a safety head that leads to safer and more conservative responses only when necessary. SafeSwitch reduces harmful outputs by approximately 80% on harmful queries while maintaining strong utility, reaching a Pareto optimal among several methods. Our method is also advantageous over traditional methods in offering more informative, context-aware refusals, and achieves these benefits while only tuning less than 6% of the original parameters. SafeSwitch demonstrates large language models' capacity for self-awareness and reflection regarding safety, offering a promising approach to more nuanced and effective safety controls.

SafeSwitch: Steering Unsafe LLM Behavior via Internal Activation Signals

Hit identification is a central challenge in early drug discovery, traditionally requiring substantial experimental resources. Recent advances in artificial intelligence, particularly large language models (LLMs), have enabled virtual screening methods that reduce costs and improve efficiency. However, the growing complexity of these tools has limited their accessibility to wet-lab researchers. Multi-agent systems offer a promising solution by combining the interpretability of LLMs with the precision of specialized models and tools. In this work, we present MADD, a multi-agent system that builds and executes customized hit identification pipelines from natural language queries. MADD employs four coordinated agents to handle key subtasks in de novo compound generation and screening. We evaluate MADD across seven drug discovery cases and demonstrate its superior performance compared to existing LLM-based solutions. Using MADD, we pioneer application of AI-first drug design to five biological targets and release the identified hit molecules. Finally, we introduce a new benchmark of query-molecule pairs and docking scores for over three million compounds to contribute to the agentic future of drug design.

MADD: Multi-Agent Drug Discovery Orchestra

Financial documents such as 10-K filings pose significant retrieval challenges due to their length, formal structure, and domain-specific language—features often underutilized by standard retrieval-augmented generation (RAG) models. We present FinGEAR (Financial Mapping-Guided Enhanced Answer Retrieval, a retrieval framework tailored for financial document analysis. FinGEAR introduces a modular architecture that combines lexicon-guided filtering, dual-hierarchy indexing (via a Summary Tree and Question Tree), and cross-encoder reranking. This structure-aware design enables fine-grained retrieval aligned with financial discourse. Extensive evaluations show that FinGEAR significantly outperforms state-of-the-art RAG baselines across multiple retrieval metrics. By explicitly modeling document semantics and structure, FinGEAR improves retrieval fidelity and enhances downstream task performance, offering a principled solution for high-stakes financial information access.

FinGEAR: Financial Mapping-Guided Enhanced Answer Retrieval

With the continuous development of language models and the widespread availability of various types of accessible interfaces, large language models (LLMs) have been applied to an increasing number of fields. However, due to the vast amounts of data and computational resources required for model development, protecting the model's parameters and training data has become an urgent and crucial concern. Due to the revolutionary training and application paradigms of LLMs, many new attacks on language models have emerged in recent years. In this paper, we define these attacks as ``reverse engineering'' (RE) techniques on LMs and aim to provide an in-depth analysis of reverse engineering of language models. We illustrate various methods of reverse engineering applied to different aspects of a model, while also providing an introduction to existing protective strategies. On the one hand, it demonstrates the vulnerabilities of even black box models to different types of attacks; on the other hand, it offers a more holistic perspective for the development of new protective strategies for models.

Towards Reverse Engineering of Language Models: A Survey

Multi-digit addition is a clear probe of the computational power of large language models. To dissect the internal arithmetic processes in LLaMA-3-8B-Instruct, we combine linear probing with logit-lens inspection. Inspired by the step-by-step manner in which humans perform addition, we propose and analyze a coherent four-stage trajectory in the forward pass: Formula-structure representations become linearly decodable first, while the answer token is still far down the candidate list. Core computational features then emerge prominently. At deeper activation layers, numerical abstractions of the result become clearer, enabling near-perfect detection and decoding of the individual digits in the sum. Near the output, the model organizes and generates the final content, with the correct token reliably occupying the top rank. This trajectory suggests a hierarchical process that favors internal computation over rote memorization. We release our code and data to facilitate reproducibility.

Addition in Four Movements: Mapping Layer-wise Information Trajectories in LLMs

Recent studies have highlighted the remarkable knowledge retention capabilities of Large Language Models (LLMs) like GPT-4, while simultaneously revealing critical limitations in maintaining knowledge currency and accuracy. Existing knowledge editing methodologies, designed to update specific factual information without compromising general model performance, often encounter two fundamental challenges: parameter conflict during knowledge overwriting and excessive computational overhead. In this paper, we introduce ForGet (Forget for Get), a novel approach grounded in the principle of "forgetting before learning". By pinpointing the location within the LLM that corresponds to the target knowledge, we first erase the outdated knowledge and then insert the new knowledge at this precise spot. ForGet is the first work to leverage a two-phase gradient-based process for knowledge editing, offering a lightweight solution that also delivers superior results. Experimental findings show that our method achieves more effective knowledge editing at a lower cost compared to previous techniques across various base models.

Forget for Get: A Lightweight Two-phase Gradient Method for Knowledge Editing in Large Language Models

Large Language Models (LLMs) are trained on diverse and often conflicting knowledge spanning multiple domains and time periods. Some of this knowledge is only valid within specific temporal contexts, such as answering the question, ``Who is the President of the United States in 2022?'' Ensuring LLMs generate time-appropriate responses is crucial for maintaining relevance and accuracy. In this work we explore activation engineering as a method for temporally aligning LLMs to improve factual recall without any training. Activation engineering has predominantly been used to steer subjective and qualitative outcomes such as toxicity or behavior. Our research is one of few that uncovers the bounds of activation engineering on objective outcomes. We explore an activation engineering technique to anchor LLaMA 2, LLaMA 3.1, Qwen 2 and Gemma 2 to specific points in time and examine the effects of varying injection layers and prompting strategies. Our experiments demonstrate up to a 44\% and 16\% improvement in relative and explicit prompting respectively, achieving comparable performance to the fine-tuning method proposed by Zhao et al. (2024). Notably, for LLaMA 2 and LLaMA 3.1 our approach achieves similar results to the fine-tuning baseline while being significantly more computationally efficient and requiring no pre-aligned datasets. Both datasets and code will be released publicly upon paper acceptance.

Temporal Alignment of Time Sensitive Facts with Activation Engineering

In this paper, we propose a novel benchmark, ChronoBias, for evaluating group bias in retrieval-augmented language models over time. Our benchmark is built upon a template-based semi-automated generation method, which effectively balances the quality and quantity tradeoff of existing benchmark curation methods. We show that for the type of knowledge that changes over time, group bias must be measured by jointly considering the time-varying nature of the knowledge for each group. Specifically, even a majority group of high parametric knowledge is susceptible to severe performance degradation given incorrect information, especially when the knowledge itself is volatile. Additionally, we show that high-capacity LLMs (e.g., ChatGPT-4o) show knowledge conflict-like patterns even after their knowledge cutoff due to their forecasting capability, which we call forecasting conflict. The forecasting conflict is more prominent in the majority groups with high knowledge volatility. We then propose a fairness metric conditioned on the time interval, and show that time-dependent notion of group fairness is necessary to maintain the group fairness over all time steps. We conclude that an adaptive retrieval for each group and time interval is necessary, to both enhance the group fairness and prevent unnecessary performance degradation.

ChronoBias: A Benchmark for Evaluating Temporal Group Bias in the Time-sensitive Knowledge of Large Language Models

Graphical user interface (GUI) agents powered by multimodal large language models (MLLMs) have shown greater promise for human-interaction. However, due to the high tuning cost, users often rely on open-source GUI agents or APIs offered by AI providers, which introduces a critical but underexplored supply chain threat: backdoor attacks. In this work, we first unveil that MLLM-powered GUI agents naturally expose multiple interaction-level triggers, such as historical steps, environment states, and task progress. Based on this observation, we introduce AgentGhost, an effective and stealthy framework for red-teaming backdoor attacks. Specifically, we first construct composite triggers by combining goal and interaction levels, allowing GUI agents to unintentionally activate backdoors while ensuring task utility. Then, we formulate backdoor injection as a Min-Max optimization problem that uses supervised contrastive learning to maximize the feature difference across sample classes at the representation space, improving flexibility of the backdoor. Meanwhile, it adopts supervised fine-tuning to minimize the discrepancy between backdoor and clean behavior generation, enhancing effectiveness and utility. Extensive evaluations of various agent models in two established mobile benchmarks show that AgentGhost is effective and generic, with attack accuracy that reaches 99.7% on three attack objectives, and shows stealthiness with only 1% utility degradation. Furthermore, we tailor a defense method against AgentGhost that reduces the attack accuracy to 22.1%. Our code is available at anonymous.

Hidden Ghost Hand: Unveiling Backdoor Vulnerabilities in MLLM-Powered Mobile GUI Agents

Recent efforts in LLM alignment have focused on constructing large-scale preference datasets via human or Artificial Intelligence(AI) annotators. However, such approaches rely on instance-wise supervision, incurring substantial annotation cost and limited interpretability. In this paper, we propose **ZEBRA**—a model behavior-wise zero-annotation framework that constructs preference data by leveraging model behavior knowledge derived from benchmark performances. ZEBRA binarizes response pairs by evaluating the quality and similarity of their origin models, entirely bypassing instance-level annotation. This allows scalable, controllable, and cost-effective alignment data generation. Empirical results show that ZEBRA achieves alignment performance comparable to instance-supervised methods, despite requiring no manual or model-based labeling.

Downloads

Next from EMNLP 2025

SafeSwitch: Steering Unsafe LLM Behavior via Internal Activation Signals

Similar lecture

Following Length Constraints in Instructions