United States

As Large Language Models (LLMs) continue to advance, reinforcement learning is increasingly viewed as a promising paradigm for optimizing their general capabilities. Nevertheless, in open-ended tasks that involve complex preferences, incorporating Reward Models (RMs) into the training pipeline remains essential. Despite its importance, RM-based RL faces significant challenges: the inherent imperfections of RMs often lead to issues such as reward hacking and reward overoptimization, undermining learning reliability. In this paper, we propose RRRL (Reward-Robust Reinforcement Learning), a framework designed to improve learning stability and mitigate reward hacking in RM-based RL. Our method introduces Bayesian Reward Model Ensembles, an innovative reward modeling method to capture reward uncertainty synchronously, alongside a trade-off optimization objective to balance performance and robustness in RL, which ensures stable learning despite imperfect RMs. The proposed method effectively mitigates reward hacking and consistently surpasses baselines across diverse benchmarks, delivering improved accuracy and long-term stability. We also provide both experimental results and analysis demonstrating that the proposed method remains effective under stochastic conditions. These findings underscore the method’s potential to improve both the performance and learning stability.

ACL 2026

Reward-Robust Reinforcement Learning in LLMs via Uncertainty Set Modeling

reward hacking

### 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**. 

ACL 2025

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

workshop 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.

Large Language Models are increasingly used as safety infrastructure for detecting harmful online content and moderating social media across multiple languages. Yet their effectiveness remains uneven across linguistic communities. This disparity reflects not only disparities in training data availability but also structural problems in annotation design. We argue that a central source of multilingual safety failure lies in the annotation gap underlying existing hate speech datasets. Most annotation guidelines and safety benchmarks are developed for English and standard language varieties, overlooking dialectal variation and culturally embedded forms of hostility. Using Arabic dialectal discourse as a case study, we show how harmful speech expressed through dialects, sarcasm, code-switching, and culturally specific expressions often remains undetected by current annotation schemes. We introduce the concept of the Multilingual Safety Annotation Gap (MSAG), identifying four sources of bias: language coverage gaps, dialect representation gaps, cultural semantic gaps, and annotation guideline gaps. We discuss implications for LLM safety alignment and outline directions for culturally grounded multilingual annotation. This paper is primarily a conceptual and methodological position paper; rather than introducing a new benchmark or empirical evaluation, we aim to formalize the MSAG as a framework for analyzing systematic weaknesses in multilingual safety annotation pipelines.

Lost in Dialect: The Annotation Gap in Multilingual LLM Safety

Text normalization, the task of converting noisy, informal text into a standardized form - is a fundamental preprocessing step for many NLP applications. Despite the growing need for Kyrgyz language processing tools, to the best of our knowledge, no prior work has addressed automatic text normalization for Kyrgyz, a morphologically rich, low-resource Turkic language. In this paper, we present the first systematic study of Kyrgyz text normalization. We collect a dataset of 1.67 million noisy–clean text pairs sourced from YouTube comments, Instagram posts, and Telegram channels, where users frequently write without punctuation, capitalization, or standard spelling. Pairs were annotated with Gemini 3 Pro; the 1,000-example test set was fully verified by two native Kyrgyz speakers with adjudication, and a random subset of the training data was spot-checked, while the full 1.67M training set was not verified exhaustively. For continual pre-training, we additionally use a 538 MB Kyrgyz corpus compiled from news portals and books. We evaluate five systems: a rule-based baseline, zero-shot mT5, a fine-tuned mT5-small model, a continually pre-trained mT5-small followed by fine-tuning, and zero-shot Gemma 4. Our experiments show that fine-tuned mT5-small achieves a CER of 0.0796, outperforming the rule-based baseline (CER 0.2029), zero-shot mT5 (CER 0.9887), and zero-shot Gemma 4 (CER 0.1620), a roughly 32× larger model in a fine-tuned vs. zero-shot setting. Human evaluation by two native Kyrgyz speakers confirms these results, with fine-tuned mT5-small rated as correct in 99.8% of cases. We further analyze why continual pre-training with span corruption does not improve over direct fine-tuning, finding hallucination in 35/40 of the inspected failure cases (87.5%, 95% Wilson CI [74%, 95%]).

Kyrgyz Text Normalization: A Comparative Study of Neural and Rule-Based Approaches

Instruction-tuned LLMs are able to provide *an* explanation about their output to users by generating self-explanations, without requiring the application of complex interpretability techniques. In this paper, we analyse whether this ability results in a *good* explanation. We evaluate self-explanations in the form of input rationales with respect to their plausibility to humans. We study three text classification tasks: sentiment classification, forced labour detection and claim verification. We include Danish and Italian translations of the sentiment classification task and compare self-explanations to human annotations. For this, we collected human rationale annotations for Climate-Fever, a claim verification dataset. We furthermore evaluate the faithfulness of human and self-explanation rationales with respect to correct model predictions, and extend the study by incorporating post-hoc attribution-based explanations. We analyse four open-weight LLMs and find that alignment between self-explanations and human rationales highly depends on text length and task complexity. Nevertheless, self-explanations yield faithful subsets of token-level rationales, whereas post-hoc attribution methods tend to emphasize structural and formatting tokens, reflecting fundamentally different explanation strategies.

A Systematic Comparison between Extractive Self-Explanations and Human Rationales in Text Classification

As large language models (LLMs) are increasingly deployed in real-world applications, the need to selectively remove unwanted knowledge while preserving model utility has become paramount. Recent work has explored sparse autoencoders (SAEs) to perform precise interventions on monosemantic features. However, most SAE-based methods operate at inference time, which does not create persistent changes in the model's parameters. Such interventions can be bypassed or reversed by malicious actors with parameter access. We introduce CRISP, a parameter-efficient method for persistent concept unlearning using SAEs. CRISP automatically identifies salient SAE features across multiple layers and suppresses their activations. We experiment with two LLMs and show that our method outperforms prior approaches on safety-critical unlearning tasks from the WMDP benchmark, successfully removing harmful knowledge while preserving general and in-domain capabilities. Feature-level analysis reveals that CRISP achieves semantically coherent separation between target and benign concepts, allowing precise suppression of the target features.

CRISP: Persistent Concept Unlearning via Sparse Autoencoders

Controlling undesirable LLM behaviors typically requires costly fine-tuning, while existing inference-time steering methods lack fine-grained adaptivity. We introduce a lightweight, trainable controller network for adaptive inference-time control. The controller observes intermediate LLM activations to predict a global scaling factor and layer-specific weights, which dynamically modulate a pre-computed ``refusal direction'' vector. Trained on harmful and benign prompts, the controller learns to apply nuanced, layer-aware steering selectively. Experiments on Llama and Mistral models show our method significantly increases refusal rates on safety benchmarks like ToxicChat, outperforming existing approaches without altering the original model parameters.

Guiding Giants: Lightweight Controllers for Weighted Activation Steering in LLMs

We present SURGeLLM (Structured Understanding, Retrieval, and Generation with LLMs), a unified transformer framework for multi-task classification across structurally heterogeneous NLP domains. SURGeLLM augments a standard encoder with two lightweight mechanisms: (i) a surgical feature gate that adaptively fuses curated lexical indicators with contextual [CLS] representations, and (ii) task-conditioned prefix tokens that inject quantized feature values and task identity into the input sequence. We provide a theoretical analysis of the gate, including an approximation-error bound in terms of surgical feature alignment and an identifiability result showing that the gate degenerates to the identity map when lexical features carry no task-relevant signal. We evaluate SURGeLLM on four diverse tasks: structured physical reasoning (D1), multi-hop retrieval classification (D2), LLM-prompt attribution (D3), and human/LLM authorship detection (D4), totaling 10,764 examples. The best variant, SURGeLLM-G-RoBERTa, achieves a mean macro-F1 of 0.921, within 0.001 of the strongest baseline, while SURGeLLM variants outperform their backbone-matched baselines on retrieval, with the best D2 score reaching 0.963. Bootstrap confidence intervals (B=2,000), ablations, and feature-significance analyses further clarify when surgical lexical priors are effective and where class imbalance limits their effectiveness.

SURGELLM: Surgical Feature Injection and Task-Conditioned Prefix Learning for Unified Multi-Task NLP

Distilling multi-step reasoning abilities from large language models (LLMs) into compact student models remains challenging due to noisy rationales, hallucinated supervision, and static teacher–student interactions. Existing reasoning distillation methods, including mentor-based approaches, predominantly operate in an open-loop manner, implicitly assuming uniform teacher reliability and consequently propagating erroneous intermediate reasoning. We propose \textbf{GateKD}, a confidence-gated closed-loop distillation framework that enables robust reasoning transfer by treating the teacher as a dynamic gatekeeper rather than a static oracle. GateKD introduces three complementary mechanisms: (i) confidence-gated soft supervision that selectively distills reliable predictive signals, (ii) gated hidden-state evolution that aligns intermediate representations only when teacher confidence is high, and (iii) reliability-filtered attention distillation that preserves stable reasoning structures while suppressing noisy patterns. These components jointly form a closed feedback loop in which teacher confidence continuously modulates the distillation process, reducing hallucination transfer and stabilizing student reasoning. Extensive experiments across commonsense, logical, and symbolic reasoning benchmarks, using T5 and Flan-T5 backbones of varying sizes, demonstrate that GateKD consistently outperforms strong open-loop distillation baselines. Notably, GateKD yields substantial gains in logical and symbolic reasoning, remains robust under low-resource distillation settings, and shows clear performance degradation when any gating component is removed. Our results highlight that confidence-gated closed-loop supervision is critical for building reliable and scalable small reasoning models.

GateKD: Confidence-Gated Closed-Loop Distillation for Robust Reasoning

Modern Large Language Models (LLMs) rely on extensive safety alignment, yet the mechanistic basis of refusal remains opaque. In this work, we investigate whether safety compliance is a deep semantic decision or a manipulable linear feature. We introduce Contrastive Logit Steering (CLS), a zero-optimization framework that isolates the "refusal direction" by contrasting hidden states derived from safe and unrestricted system prompts. Unlike representation engineering methods that intervene on internal activations, CLS operates directly on the output distribution, serving as a diagnostic probe for alignment fragility. When coupled with prefix injection to bypass initial refusal reflexes, this method induces a phase transition where guardrails collapse. Our experiments on 7 model families reveal that safety implementation is architecturally deterministic. While models like Llama-3.1 exhibit a "Late Decision" topology that is easily bypassed by CLS (reaching 95% ASR in milliseconds), others like Qwen-2.5 demonstrate "Early Divergence" by integrating safety mid-computation. Direct comparison with established activation-level steering methods shows that CLS achieves substantially higher attack success rates on Llama 2 (73% vs. 22.6%) and Qwen 7B (91% vs. 79.2%), demonstrating that logit-level intervention exposes alignment vulnerabilities that hidden-state methods underestimate. Beyond attacks, we show that this linearity enables bidirectional control: inverting the steering vector "hardens" models against jailbreaks without retraining. Our findings suggest that current alignment techniques create a steerable "safety axis" that serves as both a critical vulnerability and a precise primitive for defense.

The Geometry of Refusal: Linear Instability in Safety-Aligned LLMs

We examine whether large language models (LLMs) can reliably simulate historical FOMC policy decisions and whether persona-based agentic deliberation improves performance. Using strictly time-consistent vintage economic information, we evaluate multiple state-of-the-art LLMs on a three-way Hike/Hold/Cut classification task in both single-agent and multi-agent settings. Single-LLM baselines achieve nontrivial accuracy and track broad policy regime shifts, establishing a simple but strong benchmark. However, we identify a systematic behavioral asymmetry that we term \emph{Hold bias}: models disproportionately favor Hold decisions and remain reluctant to predict Cut outcomes even during easing cycles. This conservatism is especially costly around regime turning points, where reliable adaptation matters most. We further find that standard agentic workflows, including debate and consensus-style aggregation, do not mitigate this problem and often amplify caution rather than improve accuracy. Overall, our results show that plausible deliberation is not sufficient for trustworthy decision support. Progress will require agentic systems explicitly designed to diagnose and correct structural bias, rather than merely reproducing surface-level committee interaction.

The Conservative AI: Diagnosing Hold Bias and Reliability Limits in Persona-Based Monetary Policy Simulation

Existing work on LLM confidence has been predominantly framed as an estimation problem of uncertainty quantification and calibration, while the operational question of how confidence should be used to drive system decisions remains fragmented. This paper presents a comprehensive overview of $\textbf{confidence utilization}$ in LLMs. We first establish a unified definition of confidence as a control signal, decomposed along three axes (source, granularity, and functional role) to clarify existing ambiguities. We then propose a systematic taxonomy organized along the LLM lifecycle: confidence-aware training, confidence-driven inference, confidence-guided model selection, confidence-gated retrieval, and confidence-based risk management. For each family we compare methods on shared axes and highlight cross-cutting trade-offs, and finally outline open problems and promising directions toward jointly reliable, efficient, and trustworthy LLM systems.

Next from ACL 2026

Lost in Dialect: The Annotation Gap in Multilingual LLM Safety

Similar lecture

Rethinking Reward Model Evaluation Through the Lens of Reward Overoptimization