Austria

LLMs demonstrate remarkable utility but remain vulnerable to jailbreak attacks that aim to elicit harmful responses. Existing defenses, including post-training alignment and prompt engineering, rely on training on safety-annotated datasets and safe prompt templates, struggling with adaptability to out-of-distribution (OOD) attacks. Steering internal representations of LLMs provides real-time adjustments to defend against OOD attacks. However, it struggles with maintaining model utility, since modifying the representation disrupts the forward pass of inference. It barely considers the competitive objectives of helpfulness and harmlessness in LLMs. We argue that adversarial game-based approaches promise a solution for conflicts between the two objectives. In this paper, we propose **A**dversarial **G**ame **D**efense (AGD), an adversarial game-based defense method that dynamically adjusts LLMs&#39; internal representations to achieve a balanced trade-off between helpfulness and harmlessness. AGD first proposes an interquartile range (IQR) method to detect abnormal attention weights and correct the abnormal weights via adversarial training. AGD adopts a bi-level optimization to play a two-player variable-sum game to approach Nash Equilibrium (NE), where the two players adversarially refine head activations for helpfulness and harmlessness respectively. Furthermore, AGD applies an expert model to next-token sampling to generate safer responses. Experiments show that AGD significantly improves LLMs&#39; safety over all baselines.

ACL 2025

AGD: Adversarial Game Defense Against Jailbreak Attacks in Large Language Models

safety-utility trade-off

adversarial game theory

jailbreak attacks

bi-level optimization

large language models

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.

LLMs demonstrate remarkable utility but remain vulnerable to jailbreak attacks that aim to elicit harmful responses. Existing defenses, including post-training alignment and prompt engineering, rely on training on safety-annotated datasets and safe prompt templates, struggling with adaptability to out-of-distribution (OOD) attacks. Steering internal representations of LLMs provides real-time adjustments to defend against OOD attacks. However, it struggles with maintaining model utility, since modifying the representation disrupts the forward pass of inference. It barely considers the competitive objectives of helpfulness and harmlessness in LLMs. We argue that adversarial game-based approaches promise a solution for conflicts between the two objectives. In this paper, we propose **A**dversarial **G**ame **D**efense (AGD), an adversarial game-based defense method that dynamically adjusts LLMs' internal representations to achieve a balanced trade-off between helpfulness and harmlessness. AGD first proposes an interquartile range (IQR) method to detect abnormal attention weights and correct the abnormal weights via adversarial training. AGD adopts a bi-level optimization to play a two-player variable-sum game to approach Nash Equilibrium (NE), where the two players adversarially refine head activations for helpfulness and harmlessness respectively. Furthermore, AGD applies an expert model to next-token sampling to generate safer responses. Experiments show that AGD significantly improves LLMs' safety over all baselines.

technical paper

### 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)

Most written natural languages are composed of sequences of words and sentences. Similar to humans, large language models (LLMs) exhibit flexibility in handling textual positions - a phenomenon we term Position Generalization. They can understand texts with position perturbations and generalize to longer texts than those encountered during training with the latest techniques. These phenomena suggest that LLMs handle positions in a tolerant manner, but how LLMs computationally process positional relevance remains largely unexplored. In this work, we show how LLMs enforce certain computational mechanisms to allow for the aforementioned tolerance in position perturbations. Despite the complex design of the self-attention mechanism, in this work, LLMs are revealed to learn a counterintuitive disentanglement of attention logits, where their values show a 0.959 linear correlation with an approximation of the arithmetic sum of positional relevance and semantic importance. Furthermore, we identify a prevalent pattern in intermediate features that enables this effect, suggesting that it is a learned behavior rather than a natural result of the model architecture. Based on these findings, we provide computational explanations and criteria for the aforementioned position flexibilities observed in LLMs.

Computation Mechanism Behind LLM Position Generalization

Large language models (LLMs) have shown remarkable performance in reasoning tasks but face limitations in mathematical and complex logical reasoning. Existing methods to improve LLMs' logical capabilities either involve traceable or verifiable logical sequences that generate more reliable responses by constructing logical structures yet increase computational costs, or introduces rigid logic template rules, reducing flexibility. In this paper, we propose Reversal of Thought (RoT), a plug-and-play and cost-effective reasoning framework designed to enhance the logical reasoning abilities of LLMs during the warm-up phase prior to batch inference. RoT utilizes a Preference-Guided Reverse Reasoning warm-up strategy, which integrates logical symbols for pseudocode planning through meta-cognitive mechanisms and pairwise preference self-evaluation to generate task-specific prompts solely through demonstrations, aligning with LLMs’ cognitive preferences shaped by RLHF. Through reverse reasoning, we utilize a Cognitive Preference Manager to assess knowledge boundaries and further expand LLMs’ reasoning capabilities by aggregating solution logic for known tasks and stylistic templates for unknown tasks. Experiments across various tasks demonstrate that RoT surpasses existing baselines in both reasoning accuracy and efficiency.

Reversal of Thought: Enhancing Large Language Models with Preference-Guided Reverse Reasoning Warm-up

Large Language Models (LLMs) offer a transparent brain with accessible parameters that encode extensive knowledge, which can be analyzed, located and transferred. Consequently, a key research challenge is to transcend traditional knowledge transfer paradigms rooted in symbolic language and achieve genuine Parametric Knowledge Transfer (PKT). Significantly, exploring effective methods for transferring knowledge across LLMs of different scales through parameters presents an intriguing and valuable research direction. In this paper, we first demonstrate $\textbf{Alignment}$ in parametric space is the fundamental prerequisite to achieve successful cross-scale PKT. We redefine the previously explored knowledge transfer as Post-Align PKT (PostPKT), which utilizes extracted parameters for LoRA initialization and requires subsequent fine-tune for alignment. Hence, to reduce cost for further fine-tuning, we introduce a novel Pre-Align PKT (PrePKT) paradigm and propose a solution called $\textbf{LaTen}$ ($\textbf{L}$oc$\textbf{a}$te-$\textbf{T}$h$\textbf{e}$n-Alig$\textbf{n}$) that aligns the parametric spaces of LLMs across scales only using several training steps without following training. Comprehensive experiments on four benchmarks demonstrate that both PostPKT and PrePKT face challenges in achieving consistently stable transfer. Through in-depth analysis, we identify $\textbf{Neural Incompatibility}$ as the ethological and parametric structural differences between LLMs of varying scales, presenting fundamental challenges to achieving effective PKT. These findings provide fresh insights into the parametric architectures of LLMs and highlight promising directions for future research on efficient PKT. Our code is available at https://github.com/Trae1ounG/Neural_Incompatibility.

Neural Incompatibility: The Unbridgeable Gap of Cross-Scale Parametric Knowledge Transfer in Large Language Models

Large Language Models (LLMs) have demonstrated remarkable capabilities across a variety of tasks in different domains. However, they sometimes generate responses that are logically coherent but factually incorrect or misleading, which is known as LLM hallucinations. Data-driven supervised methods train hallucination detectors by leveraging the internal states of LLMs, but detectors trained on specific domains often struggle to generalize well to other domains. In this paper, we aim to enhance the cross-domain performance of supervised detectors with only in-domain data. We propose a novel framework, prompt-guided internal states for hallucination detection of LLMs, namely PRISM. By utilizing appropriate prompts to guide changes to the structure related to text truthfulness in LLMs' internal states, we make this structure more salient and consistent across texts from different domains. We integrated our framework with existing hallucination detection methods and conducted experiments on datasets from different domains. The experimental results indicate that our framework significantly enhances the cross-domain generalization of existing hallucination detection methods.

Prompt-Guided Internal States for Hallucination Detection of Large Language Models

While Large Language Models (LLMs) support long contexts, they struggle with performance degradation within the context window. Current solutions incur prohibitive training costs, leaving statistical behaviors and cost-effective approaches underexplored. From the decoding perspective, we identify the Posterior Salience Attenuation (PSA) phenomenon, where the salience ratio correlates with long-text performance degradation. Notably, despite the attenuation, gold tokens still occupy high-ranking positions in the decoding space. Motivated by it, we propose the training-free Positional Contrastive Decoding (PCD) that contrasts the logits derived from long-aware attention with those from designed local-aware attention, enabling the model to focus on the gains introduced by large-scale short-to-long training. Through the analysis of long-term decay simulation, we demonstrate that PCD effectively alleviates attention score degradation. Experimental results show that PCD achieves state-of-the-art performance on long-context benchmarks.

Mitigating Posterior Salience Attenuation in Long-Context LLMs with Positional Contrastive Decoding

Alignment techniques have become central to ensuring that Large Language Models (LLMs) generate outputs consistent with human values. However, existing alignment paradigms often model an averaged or monolithic preference, failing to account for the diversity of perspectives across cultures, demographics, and communities. This limitation is particularly critical in health-related scenarios, where plurality is essential due to the influence of culture, religion, personal values, and conflicting opinions. Despite progress in pluralistic alignment, no prior work has focused on health, likely due to the unavailability of publicly available datasets. To address this gap, we introduce VITAL, a new benchmark dataset comprising 13.1K value-laden situations and 5.4K multiple-choice questions focused on health, designed to assess and benchmark pluralistic alignment methodologies. Through extensive evaluation of eight LLMs of varying sizes, we demonstrate that existing pluralistic alignment techniques fall short in effectively accommodating diverse healthcare beliefs, underscoring the need for tailored AI alignment in specific domains. This work highlights the limitations of current approaches and lays the groundwork for developing health-specific alignment solutions.

VITAL: A New Dataset for Benchmarking Pluralistic Alignment in Healthcare

We introduce \textbf{Dynamic Retrieval-Augmented Expert Networks (DRAE)}, a groundbreaking architecture that addresses the challenges of lifelong learning, catastrophic forgetting, and task adaptation by combining the dynamic routing capabilities of Mixture-of-Experts (MoE); leveraging the knowledge-enhancement power of Retrieval-Augmented Generation (RAG); incorporating a novel hierarchical reinforcement learning (RL) framework; and coordinating through ReflexNet-SchemaPlanner-HyperOptima (RSHO).DRAE dynamically routes expert models via a sparse MoE gating mechanism, enabling efficient resource allocation while leveraging external knowledge through parametric retrieval (P-RAG) to augment the learning process. We propose a new RL framework with ReflexNet for low-level task execution, SchemaPlanner for symbolic reasoning, and HyperOptima for long-term context modeling, ensuring continuous adaptation and memory retention. Experimental results show that DRAE significantly outperforms baseline approaches in long-term task retention and knowledge reuse, achieving an average task success rate of 82.5\% across a set of dynamic robotic manipulation tasks, compared to 74.2\% for traditional MoE models. Furthermore, DRAE maintains an extremely low forgetting rate, outperforming state-of-the-art methods in catastrophic forgetting mitigation. These results demonstrate the effectiveness of our approach in enabling flexible, scalable, and efficient lifelong learning for robotics.

DRAE: Dynamic Retrieval-Augmented Expert Networks for Lifelong Learning and Task Adaptation in Robotics

Recently, a number of repository-level code generation benchmarks--such as CoderEval, DevEval, RepoEval, RepoBench, and LongCode-Arena--have emerged to evaluate the capabilities of large language models (LLMs) beyond standalone benchmarks like HumanEval and MBPP. Thus, a natural question is, would LLMs have similar performance in real world coding tasks as their performance in these benchmarks? Unfortunately, one cannot answer this question, since these benchmarks consist of short completions, synthetic examples, or focus on limited scale repositories, failing to represent real-world coding tasks.

To address these challenges, we create RepoCod, a Python code-generation benchmark containing complex tasks with realistic dependencies in real-world large projects and appropriate metrics for evaluating source code. It includes 980 whole-function generation tasks from 11 popular projects, 50.8\% of which require repository-level context. RepoCod includes 314 developer-written test cases per instance for better evaluation. We evaluate ten LLMs on RepoCod and find that none achieves more than 30\% pass@1 on RepoCod, indicating the necessity of building stronger LLMs that can help developers in real-world software development. In addition, we found that retrieval-augmented generation achieves better results than using target function dependencies as context.

Can Language Models Replace Programmers? REPOCOD Says ‘Not Yet’

Web development involves turning UI designs into functional webpages, which can be difficult for both beginners and experienced developers due to the complexity of HTML's hierarchical structures and styles. While Large Language Models (LLMs) have shown promise in generating source code, two major challenges persist in UI-to-HTML code generation: (1) effectively representing HTML's hierarchical structure for LLMs, and (2) bridging the gap between the visual nature of UI designs and the text-based format of HTML code. To tackle these challenges, we introduce Waffle, a new fine-tuning strategy that uses a structure-aware attention mechanism to improve LLMs' understanding of HTML's structure and a contrastive fine-tuning approach to align LLMs' understanding of UI images and HTML code. Models fine-tuned with Waffle show up to 9.00 pp (percentage point) higher HTML match, 0.0982 higher CW-SSIM, 32.99 higher CLIP, and 27.12 pp higher LLEM on our new benchmark WebSight-Test and an existing benchmark Design2Code, outperforming current fine-tuning methods.

WAFFLE: Fine-tuning Multi-Modal Model for Automated Front-End Development

Multimodal summarization (MS) combines text and visuals to generate summaries. Recently, many-to-many multimodal summarization (M3S) garnered interest as it enables a unified model for multilingual and cross-lingual MS. Existing methods have made progress by facilitating the transfer of common multimodal summarization knowledge. While, prior M3S models that fully share parameters neglect the language-specific knowledge learning, where potential interference between languages may limit the flexible adaptation of MS modes across different language combinations and hinder further collaborative improvements in joint M3S training. Based on this observation, we propose Language Constrained Multimodal Hyper Adapter (LCMHA) for M3S. LCMHA integrates language-specific multimodal adapters into multilingual pre-trained backbones via a language constrained hypernetwork, enabling relaxed parameter sharing that enhances language-specific learning while preserving shared MS knowledge learning. In addition, a language-regularized hypernetwork is designed to balance intra- and inter-language learning, generating language-specific adaptation weights and enhancing the retention of distinct language features through the regularization of generated parameters. Experimental results on the M3Sum benchmark show LCMHA's effectiveness and scalability across multiple multilingual pre-trained backbones.

Downloads

Next from ACL 2025

Computation Mechanism Behind LLM Position Generalization

Similar lecture

Context Misleads LLMs: The Role of Context Filtering in Maintaining Safe Alignment of LLMs

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES