United States

Optimizing communication topology is fundamental to the efficiency and effectiveness of Large Language Model (LLM)-based Multi-Agent Systems (MAS). While recent approaches utilize reinforcement learning to dynamically construct task-specific graphs, they typically rely on single-sample policy gradients with absolute rewards (e.g., binary correctness). This paradigm suffers from severe gradient variance and the credit assignment problem: simple queries yield non-informative positive rewards for suboptimal structures, while difficult queries often result in failures that provide no learning signal. To address these challenges, we propose Graph-GRPO, a novel topology optimization framework that integrates Group Relative Policy Optimization. Instead of evaluating a single topology in isolation, Graph-GRPO samples a group of diverse communication graphs for each query and computes the advantage of specific edges based on their relative performance within the group. By normalizing rewards across the sampled group, our method effectively mitigates the noise derived from task difficulty variance and enables fine-grained credit assignment. Extensive experiments on reasoning and code generation benchmarks demonstrate that Graph-GRPO significantly outperforms state-of-the-art baselines, achieving superior training stability and identifying critical communication pathways previously obscured by reward noise.

ACL 2026

Graph-GRPO: Stabilizing Multi-Agent Topology Learning via Group Relative Policy Optimization

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. <br>
[Full Message from General Chair Philipp Koehn](https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing) <br><br><br>
<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>
<br><br><br><br>

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

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

technical paper

With the rapid progress of large language models (LLMs), aligning a general-purpose model with downstream tasks through fine-tuning has become a central research focus. Selecting only high-quality examples for training has been shown to be one of the most effective ways to improve fine-tuning performance. However, prior work concentrates almost exclusively on data preprocessing: filtering and cleaning data before training begins. While the order and composition of training data during training have received little fine-grained attention. To fill this gap, our work proposed Fine-Grained Order Fine-Tuning, a fine-grained scheduling method of data order in epochs. Drawing on curriculum-learning principles, FOT defines data difficulty based on the relevance between the data and the model, and then performs dynamic scheduling of the training order in each epoch according to the difficulty. On both large-scale continued pre-training and small-scale supervised fine-tuning experiments, FOT has achieved an average 2.4% improvement over baselines. Our study offers a new perspective on data governance in the fine-tuning phase.

Fine-Grained Data Ordering Improves Fine-Tuning for Large Language Models

Autoregressive (AR) language modeling remains the dominant paradigm due to its dense supervision signal and highly optimized serving infrastructure, but its strictly causal, token-by-token decoding limits parallelism and non-causal modeling. While masked diffusion offers a promising path toward parallel generation, it faces two critical bottlenecks: training inefficiency stemming from sparse masked objectives, and high latency caused by iterative whole-sequence denoising. We present a systematic study of blockwise discrete diffusion, a pragmatic middle ground that preserves AR-compatible serving while enabling parallel intra-block generation. Our study proceeds in four steps: (i) a \textbf{controlled, compute- and scale-matched comparison} revealing that AR is a more effective backbone for blockwise hybrids than masked diffusion objectives; (ii) a \textbf{scalable conversion recipe, \textsc{SDAR}}, validating that AR models spanning 1.7B to 30B parameters can be adapted into block diffusion models with minimal compute while preserving backbone capabilities; and (iii) a \textbf{systematic characterization of decoding dynamics}, which reveals a virtuous cycle where larger models enable more aggressive parallel decoding, achieving {theoretical speedups over 5$\times$} and {wall-clock speedups of 2.3$\times$} on H200 GPUs in latency-critical regimes; and (iv) an \textbf{investigation of local non-causal modeling capabilities}, showing that SDAR's local bidirectional attention overcomes causal bottlenecks in scientific domains (e.g., chemistry) and enables robust test-time scaling. We release the full model suite, the training framework, and our inference engines for further innovation in non-autoregressive generative paradigms.

SDAR: A Synergistic Diffusion-AutoRegression Paradigm for Scalable Sequence Generation

Machine Unlearning (MU) enables Large Language Models (LLMs) to remove unsafe or outdated information. However, existing work assumes that all facts are equally forgettable and largely ignores whether the forgotten knowledge originates from pretraining or supervised fine-tuning (SFT). In this paper, we introduce DUAL (Dual Unlearning Evaluation across Training Stages), a benchmark of 28.6k Wikidata-derived triplets annotated with fact popularity using Wikipedia link counts and LLM-based salience scores. Our experiments show that pretrained and SFT models respond differently to unlearning. An SFT step on the forget data yields smoother forgetting, more stable tuning, and 10–50% higher retention, while direct unlearning on pretrained models remains unstable and prone to relearning or catastrophic forgetting.

Anatomy of Unlearning: The Dual Impact of Fact Salience and Model Fine-Tuning

While Large Language Models (LLMs) are widely used, they remain susceptible to jailbreak prompts that can elicit harmful or inappropriate responses. This paper introduces STAR-Teaming, a novel black-box framework for automated red teaming that 
effectively generates such prompts. STAR-Teaming integrates a Multi-Agent System (MAS) with a Strategy-Response Multiplex 
Network and employs network-driven optimization to sample effective attack strategies. This network-based approach 
recasts the intractable high-dimensional embedding space into a tractable structure, yielding two key advantages: it enhances 
the interpretability of the LLM's strategic vulnerabilities, and it streamlines the search for effective strategies by 
organizing the search space into semantic communities, thereby preventing redundant exploration. Empirical results demonstrate 
that STAR-Teaming significantly surpasses existing methods, achieving a higher attack success rate (ASR) at a lower 
computational cost. Extensive experiments validate the effectiveness and explainability of the Multiplex Network. The code is available at https://github.com/selectstar-ai/STAR-Teaming-paper.

STAR-Teaming: A Strategy-Response Multiplex Network Approach to Automated LLM Red Teaming

As Large Language Models (LLMs) are deployed globally, ensuring their safety and alignment across multiple languages becomes paramount. However, safety behaviors often vary unpredictably between languages, posing significant challenges for consistent and ethical AI. In this work, we systematically investigate the dynamics of multilingual alignment, exploring whether single-language alignment transfers cross-lingually, how language consistency is preserved during training, and the resulting trade-offs with general knowledge capabilities. We introduce RefusEU a novel refusal alignment dataset covering 12 European languages, including a dedicated test set for evaluating current state-of-the-art models. Our controlled Direct Preference Optimization (DPO) experiments provide two key insights: aligning models exclusively in English is insufficient to ensure cross-lingual safety, even for the same harm categories, whereas training on multilingual datasets can improve safety without degrading general performance, as measured by the Global MMLU benchmark.

Multilingual Refusal Alignment for Safer Large Language Models

Large language models (LLMs) have achieved good performance in multiple reasoning tasks. However, they are limited to adapt the rapid knowledge updates in the real-world scenario without retraining the entire LLM or modifying the model weights. Excluding these consuming methods, knowledge graphs (KGs) are used as external memory under knowledge updating because of their structural knowledge and efficient updating ability, which is yet limited by the gap between structural KG and LLM, and the deficient entity-independent semantics. To this end, we propose an LLM reasoning framework with hierarchical relational retrieval for large-scale knowledge updating, named G-HiRel. To integrate the structural edited KG into continuous LLMs, G-HiRel generates hierarchical instructions based on natural language questions. In order to handle the knowledge inconsistency between the KG and LLM and obtain the entity independence, G-HiRel utilizes a designed hierarchical relational retrieval for relational path candidates, which are selected by a designed semantics-based strategy. Finally, top entity-independent relational paths are instantiated and integrated into LLMs to generate the answer, in order to verify the reasoning performance under knowledge edits. Extensive experiments of G-HiRel on three benchmarks show that G-HiRel achieves superiority in terms of accuracy and interpretability. The code of G-HiRel is available at the link: https://github.com/HJJ-designed/G-HiRel.

G-HiRel: Enhancing the Adaption to Knowledge Updating for Large Language Model Reasoning

Best-of-$n$ sampling improves the accuracy of large language models (LLMs) and large reasoning models (LRMs) by generating multiple candidate solutions and selecting the one with the highest reward. The key challenge for reasoning tasks is designing a scoring function that can identify correct reasoning chains without access to ground-truth answers. We propose Probabilistic Confidence Selection and Ranking for Reasoning Chains (PiCSAR): a simple, training-free method that scores each candidate generation using the joint log-likelihood of the reasoning and final answer. This method utilises both the scores of the reasoning path (*reasoning confidence*) and the final answer (*answer confidence*). PiCSAR achieves substantial gains across several benchmarks ($+11.7$ on AIME2024, $+9.81$ on AIME2025), outperforming baselines with at least 2x fewer samples in 20 out of 25 comparisons. Our analysis reveals that correct reasoning chains exhibit higher reasoning and answer confidence, justifying the effectiveness of PiCSAR.

PiCSAR: Probabilistic Confidence Selection and Ranking for Reasoning Chains

Model editing-based jailbreak backdoor attacks against LLMs have gained attention for being lightweight, enabling vulnerability discovery in LLMs. Existing methods are implemented by binding backdoors to predefined phrases as first few output tokens, inducing the LLM’s next-token prediction to produce continuous responses. However, their effectiveness is heavily dependent on the number of bound phrases, with attack costs rising as this number increases. In this work, we propose JEST, which achieves jailbreak backdoor attacks by hijacking LLM representations into a acceptance domain rather than binding to a few output tokens. Specifically, we propose a representation transition-guided model editing to inject jailbreak backdoors into LLMs. The activated backdoor transitions the LLM from rejection domain to acceptance domain, causing it to accept and generate jailbreak behavior. To clearly distinguish between rejection and acceptance domains within LLMs, we also design a domain modeling strategy for JEST that models these two opposing domains within the representation space. Additionally, JEST-hijacked LLMs exhibit greater vulnerability to direct prompt attacks. Experimental results show that JEST outperforms existing model editing methods, demonstrating stronger jailbreak capabilities across various LLMs and datasets. We also provide analysis to explore the safety boundary of LLM.

Rejection-to-Acceptance Transition: Model Editing-Based Jailbreak Backdoor Injection Not Limited to Few Output Tokens

Common ground plays a critical role in situated spoken dialogues, where interlocutors must establish and maintain shared references to entities, events, and relations to sustain coherent interaction. For dialog systems, the ability to correctly ground conversational content in order to refer back to it later is particularly important. Prior studies have demonstrated that LLMs are capable of performing grounding acts such as requesting clarification or producing acknowledgments, yet relatively little work has investigated how common ground can be explicitly represented and stored for later use. Without such mechanisms, it remains unclear whether acknowledgment or clarification behaviors truly reflect a grounded understanding. In this work, we evaluate a model’s ability to establish and exploit common ground through relational references to entities within the shared context in a situational dialogue. We test multiple methods for representing common ground in situated dialogues and further propose approaches to improve both the establishment of common ground and its subsequent use in the conversation.

Frame of Reference: Addressing the Challenges of Common Ground Representation in Situational Dialogs

Diversity has been gaining interest in the NLP community in recent years. At the same time, state-of-the-art transformer models such as ModernBERT use very large pre-training datasets, which are driven by size rather than by diversity. This summons to investigate the
impact of diversity on pre-training. We do so in this study, with the express intent of reducing pre-training dataset size, while retaining at
least comparable performance. We compare diversity-driven sampling algorithms, and we use the best one to pre-train several ModernBERT models on French with a fixed compute budget. We fine-tune and evaluate them on a variety of French benchmarks. We compare them with models pre-trained on randomly sampled data of commensurate size, with the same compute budget. We find that both random and diversity-driven sampling may reduce the pre-training dataset by up to 94% and the pre-training time by up to 73% while maintaining performance. Moreover, in some tasks, the inherent quality of models, estimated via head-only fine-tuning, is up to 10 points higher with diversity sampling than with random sampling.

Downloads

Next from ACL 2026

Fine-Grained Data Ordering Improves Fine-Tuning for Large Language Models

Similar lecture

Enhancing LLM-based Search Agents via Contribution Weighted Group Relative Policy Optimization