Austria

This study addresses a critical gap in safety tuning practices for Large Language Models (LLMs) by identifying and tackling a refusal position bias within safety tuning data, which compromises the models&#39; ability to appropriately refuse generating unsafe content. We introduce a novel approach, \textbf{De}coupled \textbf{R}efusal \textbf{T}r\textbf{a}ining (DeRTa), designed to empower LLMs to refuse compliance to harmful prompts at any response position, significantly enhancing their safety capabilities. DeRTa incorporates two novel components: (1) Maximum Likelihood Estimation (MLE) with Harmful Response Prefix, which trains models to recognize and avoid unsafe content by appending a segment of harmful response to the beginning of a safe response, and (2) Reinforced Transition Optimization (RTO), which equips models with the ability to transition from potential harm to safety refusal consistently throughout the harmful response sequence. Our empirical evaluation, conducted using LLaMA3 and Mistral model families across six attack scenarios, demonstrates that our method not only improves model safety without compromising performance but also surpasses baseline methods in defending against attacks.

ACL 2025

Refuse Whenever You Feel Unsafe: Improving Safety in LLMs via Decoupled Refusal Training

safety alignment

jailbreak

It is my great pleasure and honor to welcome you to the 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026), held in San Diego, California, USA, July 2–7, 2026. ACL 2026 continues our fields tradition of excellence in scholarship, innovation, and inclusivity, and I am deeply grateful to the many volunteers who have worked tirelessly to bring this event to life.

It is exciting to see the many developments in the field reflected at our conference, such as the advancement of large language models (born out of our community) into many application directions. This is not without challenges, partly due to unsolved problems, and partly due to misuse of the technology. All this is reflected well at ACL 2026. The conference also reflects that the field is not only about advancing technology, but also the science of fundamentally understanding language and how it is use by many different speaker communities. 
[Full Message from General Chair Philipp Koehn](https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing) 
<html><button style="display: inline-flex; align-items: center; justify-content: center; white-space: nowrap; border-radius: 9999px; font-weight: bold; background: #7c3aed; color: white; font-family: 'Space Grotesk', sans-serif; height: 40px; font-size: 16px; padding: 0 20px; border: none; cursor: pointer" onclick="window.open('https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing')">Conference Handbook</button></html>

To access this event page, you need to log in with the **email address you registered with**. Access credentials will be sent to your email from Underline - subject line "Welcome to ACL 2026". Please be sure to check your spam email folder if you do not see an email confirmation right away.

To access this event page, you are required to register.
Please complete your registration to continue.

We recommend reading [**the registration information**](https://2026.aclweb.org/registration/) first.

**Online Registration Form**: https://acl.swoogo.com/acl2026

Registration Required

ACL 2026

The 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026) will take place in San Diego, California, from July 2 - 7, 2026.

This study addresses a critical gap in safety tuning practices for Large Language Models (LLMs) by identifying and tackling a refusal position bias within safety tuning data, which compromises the models' ability to appropriately refuse generating unsafe content. We introduce a novel approach, \textbf{De}coupled \textbf{R}efusal \textbf{T}r\textbf{a}ining (DeRTa), designed to empower LLMs to refuse compliance to harmful prompts at any response position, significantly enhancing their safety capabilities. DeRTa incorporates two novel components: (1) Maximum Likelihood Estimation (MLE) with Harmful Response Prefix, which trains models to recognize and avoid unsafe content by appending a segment of harmful response to the beginning of a safe response, and (2) Reinforced Transition Optimization (RTO), which equips models with the ability to transition from potential harm to safety refusal consistently throughout the harmful response sequence. Our empirical evaluation, conducted using LLaMA3 and Mistral model families across six attack scenarios, demonstrates that our method not only improves model safety without compromising performance but also surpasses baseline methods in defending against attacks.

poster

### Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

Message from the General Chair: 
*It is my great pleasure and honor to welcome you to the 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025), held in beautiful Vienna, Austria, from July 27 to August 1, 2025. ACL2025continues our field’s tradition of excellence in scholarship, innovation, and inclusivity, and I am deeply grateful to the many volunteers who have worked tirelessly to bring this event to life.* 
[Read more](https://drive.google.com/file/d/1GI_hvOpjswAuYdUTromfeDiPpCcqidwg/view?usp=sharing)

To access this event page, you need to log in with the **email address you registered with**. Access credentials will be sent to your email from Underline - subject line "Welcome to ACL 2025". Please be sure to check your spam email folder if you do not see an email confirmation right away.

Please log in

To access this event page, you need to log in with the **email address you registered with**. 

Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

Peer review is a cornerstone of quality control in scientific publishing. With the increasing workload, the unintended use of `quick' heuristics, referred to as \emph{lazy thinking}, has emerged as a recurring issue compromising review quality. Automated methods to detect such heuristics can help improve the peer-reviewing process. However, there is limited NLP research on this issue, and no real-world dataset exists to support the development of detection tools. This work introduces $\textsc{LazyReview}$, a dataset of peer-review sentences annotated with fine-grained \emph{lazy thinking} categories. Our analysis reveals that Large Language Models (LLMs) struggle to detect these instances in a zero-shot setting. However, instruction-based fine-tuning on our dataset significantly boosts performance by 10-20 performance points, highlighting the importance of high-quality training data. Furthermore, a controlled experiment demonstrates that reviews revised with \emph{lazy thinking} feedback are more comprehensive and actionable than those written without such feedback. We will release our dataset and the enhanced guidelines that can be used to train junior reviewers in the community.

LazyReview A Dataset for Uncovering Lazy Thinking in NLP Peer Reviews

LLMs are aligned to follow input instructions by learning which of two responses users prefer for a prompt. However, such preference data do not convey *why* users prefer responses that are chosen or rejected, so LLMs trained on these datasets cannot tailor responses to varied user needs. To surface these parameters of personalization, we apply *abductive reasoning* to preference data, inferring needs and interests of users, i.e., personas, that may prefer either response. We test this idea in two steps: **Persona Inference (PI)**—abductively inferring personas of users who prefer chosen or rejected outputs—and **Persona Tailoring (PT)**—training models to tailor outputs to personas from PI. We show: 1) LLMs infer personas accurately explaining why different users may prefer *both* chosen or rejected outputs; 2) Training on preference data augmented with PI personas via PT boosts personalization and generalizes to supporting user-written personas; and 3) Rejected response personas form harder personalization evaluations, showing PT better aids users with uncommon preferences versus typical alignment methods. We argue for an abductive view of preferences for personalization, asking not only which response is better but when, why, and for whom.

Whose Boat Does it Float? Improving Personalization in Preference Tuning via Inferred User Personas

The rise of social media has amplified the spread of fake news, now further complicated by large language models (LLMs) like ChatGPT, which ease the generation of highly convincing, error-free misinformation, making it increasingly challenging for the public to discern truth from falsehood. Traditional fake news detection methods relying on linguistic cues have also become less effective. Moreover, current detectors primarily focus on binary classification and English texts, often overlooking the distinction between machine-generated true vs. fake news and the detection in low-resource languages. To this end, we updated the detection schema to include machine-generated news focusing on Urdu. We further propose a conjoint detection strategy to improve the accuracy and robustness. Experiments show its effectiveness across four datasets in various settings.

Detection of Human and Machine-Authored Fake News in Urdu

Text-attributed graphs (TAGs) are prevalent in various real-world applications, including academic networks, e-commerce platforms, and social networks. Effective learning on TAGs requires leveraging both textual node features and structural graph information. While language models (LMs) excel at processing text and graph neural networks (GNNs) effectively capture relational structures, their direct integration is computationally prohibitive due to the high cost of text and graph representation learning. Existing approaches address this challenge by adopting a two-step pipeline where LMs generate fixed node embeddings, which are then used for GNN training. However, this method neglects the interaction between textual and structural information, leading to suboptimal learning outcomes. To overcome these limitations, we propose SKETCH (Semantic Knowledge and Structure Enrichment), a novel framework that decouples node aggregation from graph convolution and integrates it into the text representation learning process. SKETCH enhances TAG learning by incorporating two key aggregation mechanisms: (1) Semantic aggregation, which retrieves semantically relevant node texts for contextual enrichment, and (2) Structural aggregation, which propagates textual features beyond immediate neighbors to capture broader graph relationships. Extensive experiments demonstrate that SKETCH outperforms state-of-the-art TAG learning methods while requiring fewer computational resources. By enabling a more efficient and effective fusion of textual and structural information, SKETCH provides new insights into TAG problems and offers a practical solution for real applications.

Taming Language Models for Text-attributed Graph Learning with Decoupled Aggregation

Large vision-language models (LVLMs) have made substantial progress in integrating large language models (LLMs) with visual inputs, enabling advanced multimodal reasoning. Despite their success, a persistent challenge is hallucination—where generated text fails to accurately reflect visual content—undermining both accuracy and reliability. Existing methods focus on alignment training or decoding refinements but primarily address symptoms at the generation stage without probing the underlying causes. In this work, we investigate the internal mechanisms driving hallucination in LVLMs, with an emphasis on the multi-head attention module. Specifically, we introduce Vision-aware Head Divergence (VHD), a metric that quantifies the sensitivity of attention head outputs to visual context. Based on this, our findings reveal the presence of vision-aware attention heads that are more attuned to visual information; however, the model's overreliance on its prior language patterns is closely related to hallucinations. Building on these insights, we propose Vision-aware Head Reinforcement (VHR), a training-free approach to mitigate hallucination by enhancing the role of vision-aware attention heads. Extensive experiments demonstrate that our method achieves superior performance compared to state-of-the-art approaches in mitigating hallucinations, while maintaining high efficiency with negligible additional time overhead. The code is available at https://github.com/jinghan1he/VHR.

Cracking the Code of Hallucination in LVLMs with Vision-aware Head Divergence

Retrieval-augmented generation (RAG) has emerged as a pivotal technology in natural language processing, owing to its efficacy in generating factual content. However, its informative inputs and complex paradigms often lead to a greater variety of errors. Consequently, achieving automated on-policy assessment and error-oriented correction remain unresolved issues. In this paper, we propose RAG-Critic, a novel framework that leverages a critic-guided agentic workflow to improve RAG capabilities autonomously. Specifically, we initially design a data-driven error mining pipeline to establish a hierarchical RAG error system. Based on this system, we progressively align an error-critic model using a coarse-to-fine training objective, which automatically provides fine-grained error feedback. Finally, we design a critic-guided agentic RAG workflow that customizes executor-based solution flows based on the error-critic model's feedback, facilitating an error-driven self-correction process. Experimental results across seven RAG-related datasets confirm the effectiveness of RAG-Critic, while qualitative analysis offers practical insights for achieving reliable RAG systems. Our dataset and code are available at https://github.com/RUC-NLPIR/RAG-Critic.

RAG-Critic: Leveraging Automated Critic-Guided Agentic Workflow for Retrieval Augmented Generation

In visual question answering (VQA) context, users often pose ambiguous questions to visual language models (VLMs) due to varying expression habits. Existing research addresses such ambiguities primarily by rephrasing questions. These approaches neglect the inherently interactive nature of user interactions with VLMs, where ambiguities can be clarified through user feedback. However, research on interactive clarification faces two major challenges: (1) Benchmarks are absent to assess VLMs' capacity for resolving ambiguities through interaction; (2) VLMs are trained to prefer answering rather than asking, preventing them from seeking clarification. To overcome these challenges, we introduce ClearVQA benchmark, which targets three common categories of ambiguity in VQA context, and encompasses various VQA scenarios. Furthermore, we propose an automated pipeline to generate ambiguity-clarification question pairs, enabling VLMs to ask reasonable clarification questions and generate more accurate and specific answers based on user feedback, as demonstrated by experimental results.

Teaching Vision-Language Models to Ask: Resolving Ambiguity in Visual Questions

This paper introduces LongBench v2, a benchmark designed to assess the ability of LLMs to handle long-context problems requiring deep understanding and reasoning across real-world multitasks. LongBench v2 consists of 503 challenging multiple-choice questions, with contexts ranging from 8k to 2M words, across six major task categories: single-document QA, multi-document QA, long in-context learning, long-dialogue history understanding, code repository understanding, and long structured data understanding. To ensure the breadth and the practicality, we collect data from nearly 100 highly educated individuals with diverse professional backgrounds. We employ both automated and manual review processes to maintain high quality and difficulty, resulting in human experts achieving only 53.7% accuracy under a 15-minute time constraint. Our evaluation reveals that the best-performing model, when directly answers the questions, achieves only 50.1% accuracy. In contrast, the o1-preview model, which includes longer reasoning, achieves 57.7%, surpassing the human baseline by 4%. These results highlight the importance of enhanced reasoning ability and scaling inference-time compute to tackle the long-context challenges in LongBench v2.

LongBench v2: Towards Deeper Understanding and Reasoning on Realistic Long-context Multitasks

The development of large language models (LLMs) depends on **trustworthy evaluation**. However, most current evaluations rely on public benchmarks, which are prone to data contamination issues that significantly compromise fairness. Previous researches have focused on constructing dynamic benchmarks to address contamination. However, continuously building new benchmarks is costly and cyclical.
In this work, we aim to tackle contamination by analyzing the mechanisms of contaminated models themselves. Through our experiments, we discover that the overestimation of contaminated models is likely due to parameters acquiring shortcut solutions in training. We further propose a novel method for identifying shortcut neurons through **comparative and causal analysis**.
Building on this, we introduce an evaluation method called **shortcut neuron patching** to suppress shortcut neurons. Experiments validate the effectiveness of our approach in mitigating contamination. Additionally, our evaluation results exhibit a strong linear correlation with MixEval, a recently released trustworthy benchmark, achieving a Spearman coefficient ($\rho$) exceeding 0.95. This high correlation indicates that our method closely reveals true capabilities of the models and is trustworthy. We conduct further experiments to demonstrate the generalizability of our method across various benchmarks and hyperparameter settings. **Code**: https://github.com/GaryStack/Trustworthy-Evaluation.

Establishing Trustworthy LLM Evaluation via Shortcut Neuron Analysis

Speculative sampling has emerged as an important technique for accelerating the auto-regressive generation process of large language models (LLMs) by utilizing a draft-then-verify mechanism to produce multiple tokens per forward pass. While state-of-the-art speculative sampling methods use only a single layer and a language modeling (LM) head as the draft model to achieve impressive layer compression, their efficiency gains are substantially reduced for large-vocabulary LLMs, such as Llama-3-8B with a vocabulary of 128k tokens. To address this, we present FR-Spec, a frequency-ranked speculative sampling framework that optimizes draft candidate selection through vocabulary space compression. By constraining the draft search to a frequency-prioritized token subset, our method reduces LM Head computation overhead by 75% while ensuring the equivalence of the final output distribution. Experiments across multiple datasets demonstrate an average of 1.12$\times$ speedup over the state-of-the-art speculative sampling method EAGLE-2. Code is available
at https://github.com/thunlp/FR-Spec.

Downloads

Next from ACL 2025

LazyReview A Dataset for Uncovering Lazy Thinking in NLP Peer Reviews

Similar lecture

Deactivating Refusal Triggers: Understanding and Mitigating Overrefusal in Safety Alignment

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES