United States

Safety-aligned large language models (LLMs) often generate refusal responses to harmless queries due to the over-refusal problem. However, existing methods for mitigating over-refusal cannot maintain a low refusal ratio for harmless queries while keeping a high refusal ratio for malicious ones. In this paper, we analyze how system prompts with varying safety levels affect LLM refusal behaviors when facing over-refusal queries. A key observation is that, when LLMs suffer from the over-refusal issue, non-refusal tokens remain present in the next-token candidate list, but the model systematically fails to select them, despite the generation of refusal tokens. Based on this observation, we propose a training-free and model-agnostic approach, Adaptive Contrastive Decoding (AdaCD), to mitigate over-refusal while maintaining LLM safety. First, AdaCD compares the output distributions of the LLM with or without an extreme safety system prompt to refine the refusal token distribution. Second, we introduce an adaptive contrastive decoding strategy that dynamically incorporates or removes the refusal token distribution, adaptively boosting the probability of selecting refusal or non-refusal tokens. Experimental results on five benchmark datasets show that, on average, AdaCD reduces the refusal ratio for over-refusal queries by 10.35%, yet still increases the refusal ratio for malicious queries by 0.13%. Code is available at https://shorturl.at/Z31Oe.

ACL 2026

Please refuse to answer me! Mitigating Over-Refusal in Large Language Models via Adaptive Contrastive Decoding

## 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." 

*Message from the General Chair, Dirk Hovy*

EMNLP 2025

Celebrate 30 Years of EMNLP! 
EMNLP 2025 will be held in Suzhou, China from November 5th to November 9th, 2025.

technical paper

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

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

Existing LLM hallucination mitigation methods, including prompt engineering and model optimization, either hardly alter models' internal knowledge or have poor cross-domain generalization. Contrastive decoding mitigates hallucinations by using layer-wise differences in LLMs. However, prior studies only explore transformer-based models (e.g., GPT), ignoring other effective frameworks like mixture-of-experts (MoE) models. Since MoE alters the traditional transformer architecture, we conduct empirical studies to investigate whether similar layer-wise differences exist in MoEs. Our results show that they do not exist in MoE with shared experts; nevertheless, across different MoEs, higher layers exhibit distinct expert activation patterns between factual and non-factual outputs. Building on these, we propose EAACD, an expert-aware adaptive contrast decoding that uses expert differences in MoE's higher layers to mitigate hallucinations on QA tasks. EAACD splits high‑layer experts into a higher-reliability group and several lower-reliability groups based on their confidence and consistency. It contrasts the higher-reliability group's prediction with each lower-reliability group's prediction to calibrate the model's original predictions. To strengthen this contrast, EAACD amplifies hallucinations from lower-reliability experts via attention and masking to provide stronger negative references. EAACD outperforms all baselines on four datasets

Knowledge Injection Exists in MoE? Exploring Expert-Aware Contrast Decoding in MoE for Mitigating LLMs’ Hallucinations

Large language models (LLMs) are largely static and often redo reasoning or repeat mistakes. Prior experience reuse typically relies on external retrieval, which is similarity-based, can introduce noise, and adds latency. We introduce \textbf{SEAM} (\textbf{S}tructured \textbf{E}xperience \textbf{A}dapter \textbf{M}odule), a lightweight, executor-specific plug-in that stores experience in its parameters and generates a structured, instance-tailored experience entry in a single forward pass to guide a frozen LLM executor. SEAM is trained for utility via executor rollouts and GRPO while keeping the executor frozen, and can be further improved with logged-success SFT after deployment. Experiments on mathematical reasoning benchmarks show consistent accuracy gains across executors with low overhead. Extensive ablation and analysis further elucidate the mechanisms underlying SEAM’s effectiveness and robustness.\footnote{We release our code at \url{https://anonymous.4open.science/r/SEAM}.}

Beyond Experience Retrieval: Learning to Generate Utility-Optimized Structured Experience for Frozen LLMs

This paper primarily demonstrates a method to quantitatively assess the alignment between multi-step, structured reasoning in large language models and human preferences. We introduce the Alignment Score, a semantic-level metric that compares a model-produced chain of thought traces with a human-preferred reference by constructing semantic-entropy-based matrices over intermediate steps and measuring their divergence. Our analysis shows that Alignment Score tracks task accuracy across models and hop depths, and peaks at 2-hop reasoning. Empirical results further indicate that misalignment at greater reasoning depths is driven mainly by alignment errors such as thematic shift and redundant reasoning. Viewing chain sampling as drawing from a distribution over reasoning paths, we empirically demonstrate a strong and consistent correlation between Alignment Score and accuracy, readability, and coherence, supporting its use as a diagnostic signal. The code is available.

Chain-of-Thought as a Lens: Evaluating Structured Reasoning Alignment between Human Preferences and Large Language Models

Membership inference attack (MIA) has emerged as a promising tool for auditing the training data of LLMs, supporting data privacy and copyright protection. 
Most existing MIA methods rely on the assumption that LLMs assign higher confidence scores to training samples than to non-training ones.
However, since LLMs generate text by sampling high-confidence tokens, they naturally produce AI-generated texts (AIGTs) that also satisfy this assumption.
In this work, we empirically confirm that such AIGTs, regardless of whether they are generated by the target LLM, can lead existing MIAs to assign even higher membership likelihoods than those of true training samples, thereby significantly undermining their reliability.
To address this challenge, we propose a robust membership inference framework for reliably identifying training data.
Our method adopts a mixture-of-experts formulation to jointly model interactions across complementary features derived from multiple MIA methods and AIGT detectors, which can remain robust against adversarially generated samples.
Furthermore, by leveraging expert components, our method provides explainable insights into the characteristics of member data.
Experiments on various datasets and LLMs show that adversarial samples substantially degrade the performance of baselines, whereas our method preserves performance close to that of the unattacked setting.
Codes and datasets are released at https://github.com/kong-hyh/MoMIA.

Robust Membership Inference for Large Language Models under Adversarial Generative Corruption

Reward models (RMs) are the surrogate objectives in reinforcement learning from human feedback (RLHF), and their scores directly steer policy optimization. We show that standard RM training is vulnerable in data subsets where response quality depends only weakly on the context: such instances encourage the RM to ignore the context, leading to context neglect and degraded accuracy. To address this failure mode, we propose Distribution-Aware Reward Modeling (DARM), which augments the RM objective with a conditional mutual information regularizer that maximizes context and the predicted reward conditioned on the response. By explicitly preserving the sensitivity of reward signals to the prompting context, DARM reduces over-reliance on response-only features and improves robustness to contextual variation. Extensive experiments across in-distribution and out-of-distribution settings show that DARM trained RMs deliver more accurate and consistent scoring than strong baselines. We further evaluate its downstream impact in RLHF, where DARM produce better aligned policies. We also demonstrate the necessity of each DARM design component and the impact of key parameters on performance through ablation experiments.

DARM: Distribution-Aware Reward Modeling by Alleviating Biases from Low Preference-Context Dependency Data

Figurative language recognition poses significant challenges in NLP, particularly when distinguishing between fine-grained rhetorical categories such as metaphor, metonymy, and simile. This paper formulates the problem as a four-way sentence-level classification task and proposes FL-MSCL, a unified framework integrating prompt-based knowledge injection with supervised contrastive learning. Experiments across both unified and single-class benchmarks demonstrate that FL-MSCL achieves competitive performance compared to State-of-the-Art (SOTA) methods, indicating consistent advantages in cross-category generalization and category-specific detection.

FL-MSCL: A Unified Figurative Language Detection Model Driven by Multi-Type Signals and Contrastive Learning

Vision-Language-Action (VLA) models are a promising paradigm for generalist robotic manipulation by grounding high-level semantic instructions into executable physical actions. However, prevailing approaches typically adopt a monolithic generation paradigm, directly mapping visual-linguistic features to high-frequency motor commands in a flat, non-hierarchical fashion. This strategy overlooks the inherent hierarchy of robotic manipulation, where complex actions can be naturally modeled in a Hybrid Action Space, decomposing into discrete macro-directional reaching and continuous micro-pose alignment, severely widening the semantic-actuation gap and imposing a heavy representational burden on grounding high-level semantics to continuous actions. To address this, we introduce Libra-VLA, a novel Coarse-to-Fine Dual-System VLA architecture. We explicitly decouple the learning complexity into a coarse-to-fine hierarchy to strike a training equilibrium, while simultaneously leveraging this structural modularity to implement an asynchronous execution strategy. The Semantic Planner predicts discrete action tokens capturing macro-directional intent, while the Action Refiner conditions on coarse intent to generate high-frequency continuous actions for precise alignment. Crucially, our empirical analysis reveals that performance follows an inverted-U curve relative to action decomposition granularity, peaking exactly when the learning difficulty is balanced between the two sub-systems. With the asynchronous design, our approach offers a scalable, robust, and responsive solution for open-world manipulation.

Libra-VLA: Achieving Learning Equilibrium via Asynchronous Coarse-to-Fine Dual-System

Large Language Models (LLMs) exhibit powerful capabilities but inevitably memorize sensitive information, raising privacy, copyright, and safety concerns. Existing LLM unlearning methods typically rely on updating model parameters. While effective, they are often limited in real-world scenarios: fine-tuning large-scale models is costly, may introduce potential irreversible risks, and depends on both forget and retain datasets, which are often difficult to obtain in full. To address these challenges, an ideal solution is to achieve unlearning at inference time. To this end, we propose SEGUE, a training-free, plug-and-play inference-time unlearning strategy. SEGUE employs a probe to detect queries involving forgettable concepts and applies entropy-guided decoding to suppress target knowledge, enabling controllable non-factual generation while preserving overall model capabilities. Experiments on the MUSE, RWKU, and WMDP datasets, covering copyright, entity, and potential-risk knowledge, show that SEGUE effectively balances sensitive knowledge suppression and generation quality, outperforming existing most inference-time unlearning methods.

Decoding-Unlearning: Fact Forgetting via Entropy-Guided Inference

Microscaling Floating-Point (MXFP) has emerged as a promising low-precision format for large language models (LLMs). Despite various post-training quantization (PTQ) algorithms being proposed, they mostly focus on integer quantization, while their applicability and behavior under MXFP formats remain largely unexplored. To address this gap, this work conducts a systematic investigation of PTQ under MXFP formats, encompassing over 7 PTQ algorithms, 15 evaluation benchmarks, and 3 LLM families. The key findings include: 1) MXFP8 consistently achieves near-lossless performance, while MXFP4 introduces substantial accuracy degradation and remains challenging; 2) PTQ effectiveness under MXFP depends strongly on format compatibility, with some algorithmic paradigms being consistently more effective than others; 3) PTQ performance exhibits highly consistent trends across model families and modalities, in particular, quantization sensitivity is dominated by the language model rather than the vision encoder in multimodal LLMs; 4) The scaling factor of quantization is a critical error source in MXFP4, and a simple pre-scale optimization strategy can significantly mitigate its impact. Together, these results provide practical guidance on adapting existing PTQ methods to MXFP quantization.

Benchmarking Post-Training Quantization of Large Language Models under Microscaling Floating Point Formats

Lifelong model editing (LME) aims to sequentially rectify outdated or inaccurate knowledge in deployed LLMs while minimizing side effects on unrelated inputs. However, existing approaches typically apply parameter perturbations to a static and dense set of LLM layers for all editing instances. This practice is counter-intuitive, as we hypothesize that different pieces of knowledge are stored in distinct layers of the model. Neglecting this layer-wise specificity can impede adaptability in integrating new knowledge and result in catastrophic forgetting for both general and previously edited knowledge. To address this, we propose HiEdit, a hierarchical reinforcement learning framework that adaptively identifies the most knowledge-relevant layers for each editing instance. By enabling dynamic, instance-aware layer selection and incorporating an intrinsic reward for sparsity, HiEdit achieves precise, localized updates. Experiments on various LLMs show that HiEdit boosts the performance of the competitive RLEdit by an average of 8.48\% with perturbing only half of the layers per edit. Our code is available at: https://github.com/yangfanww/hiedit.

Downloads

Next from ACL 2026

Knowledge Injection Exists in MoE? Exploring Expert-Aware Contrast Decoding in MoE for Mitigating LLMs’ Hallucinations

Similar lecture

Understanding Refusal in Language Models with Sparse Autoencoders

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES