United States

Reinforcement Learning with Verifiable Rewards (RLVR) and Reinforcement Learning from Internal Feedback (RLIF) often fail to benefit from test-time compute due to entropy collapse and the resulting loss of reasoning diversity. We show that this collapse is driven not by uniform entropy decay, but by premature overconfidence at a small number of structurally critical decision points. Based on a token-level analysis of GRPO-style policy optimization, we propose SCOPE (Structural Collapse-aware Optimization via Partial Entropy control), which assigns each generated token a redistribution score and applies selective KL regularization to only the top $\sim 5$\% of tokens under this score. Across model scales and architectures on math reasoning benchmarks, SCOPE consistently improves performance under both RLVR and RLIF settings, demonstrating that targeted entropy control at a vanishingly small subset of tokens is sufficient to sustain reasoning diversity and effective test-time scaling.

ACL 2026

A Few Bad Apples Spoil the Bunch: Preventing Global Entropy Collapse Driven by a Small Set of Tokens in LLM Reasoning

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.

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.

poster

Large Language Models (LLMs) often struggle with computational efficiency and error propagation in multi-step reasoning tasks. While recent advancements on prompting and post-training have enabled LLMs to perform step-wise reasoning, they still tend to explore unproductive solution paths without effective backtracking or strategy adjustment. In this paper, we propose Meta-Reasoner, a new framework that empowers LLMs to ``think about how to think''. It optimizes the inference process by dynamically adapting reasoning strategies in real-time. Our approach employs contextual multi-armed bandits (CMABs) to learn an adaptive policy. It learns to evaluate the current state of LLM's reasoning and determine optimal strategy that is most likely to lead to a successful outcome during inference, like whether to backtrack, switch to a new approach, or restart the problem-solving process. This meta-guidance helps avoid unproductive paths exploration during inference and hence improves computational efficiency. We evaluate Meta-Reasoner on math problems (e.g., Game-of-24, TheoremQA) and scientific tasks (e.g., SciBench). Results show that our method outperform previous SOTA methods by 9-12% in accuracy, while reducing inference time by 28-35% under the same compute budget. Additional experiments on creative writing demonstrate the generalizability of our approach to diverse reasoning-intensive tasks.

 Meta-Reasoner: Dynamic Guidance for Optimized Inference-time Reasoning in Large Language Models

Post-training LLMs with Reinforcement Learning, specifically Group Relative Policy Optimization (GRPO), has emerged as a paradigm for enhancing mathematical reasoning. However, standard GRPO relies on scalar correctness rewards that are often non-injective with respect to semantic content: distinct reasoning paths receive identical rewards. This leads to a Diversity-Quality Inconsistency, where the policy collapses into a narrow set of dominant modes while ignoring equally valid but structurally novel strategies.

To bridge this gap, we propose Diversity-aware Reward Adjustment (DRA), a theoretically grounded framework that calibrates the reward signal using the semantic density of sampled groups. By leveraging Submodular Mutual Information (SMI), DRA implements an Inverse Propensity Scoring (IPS) mechanism that effectively de-biases the gradient estimation. This creates a repulsive force against redundancy, driving the policy to achieve better coverage of the high-reward landscape.

Our method is plug-and-play and integrates seamlessly with GRPO variants. Empirical evaluations on five math benchmarks demonstrate that DRA-GRPO consistently outperforms strong baselines, achieving an average accuracy of 58.2% on DeepSeek-R1-Distill-Qwen-1.5B with only 7,000 training samples and $55 cost, highlighting the critical role of diversity calibration in data-efficient alignment.

DRA-GRPO: Your GRPO Needs to Know Diverse Reasoning Paths for Mathematical Reasoning

While LLMs have demonstrated remarkable potential in Question Answering (QA), evaluating personalization remains a critical bottleneck. Existing paradigms predominantly rely on surface-level similarity or manual heuristics, often lacking sufficient data-driven validation. We address this by mining Community-Individual Preference Divergence (CIPD)—where individual choices override consensus—to distill six key personalization factors as evaluative dimensions. Accordingly, we introduce CoPA, a benchmark with 1,985 user profiles for fine-grained, factor-level assessment. By quantifying the alignment between model outputs and user-specific cognitive preferences inferred from interaction patterns, CoPA provides a more comprehensive and discriminative standard for evaluating personalized QA than generic metrics. The code is available at https://github.com/bjzgcai/CoPA.

CoPA: Benchmarking Personalized Question Answering with Data-Informed Cognitive Factors

Emotional interaction is increasingly crucial for conversational AI, yet current systems lack a self-emotion determination mechanism to drive the streaming text-to-speech (TTS) synthesis. We propose an emotion-planning framework that determines the emotion prior to the textual generation, grounding the downstream emotional TTS in a streaming manner. The framework is implemented by a plug-and-play LLM module, initialized from pretrained LLMs, and trained by reinforcement learning (RL) with emotions as the actions. A hybrid reward is employed which combines imitation signals with theory-driven scoring, in which the theory of Plutchik's wheel of emotions is adopted. By experiments on DailyDialog, EmoryNLP, IMEOCAP, and MELD, our method outperforms prompting and finetuning baselines on both emotion determination and response quality. We finally implement an entire streaming pipeline for real-time deployment, with the speech quality confirming the framework's emotional alignment, contextual coherence, and expressive fluency. Codes, cases, and demos are available in https://sixingdeguo.github.io/EmoQ-page/.

Self-EmoQ: Plutchik-Guided Value-based Planning to Drive Streaming Emotional TTS

Chain-of-Thought (CoT) prompting has been shown to be effective in eliciting structured reasoning (i.e., CoT reasoning) from large language models (LLMs). Regardless of its popularity, recent studies expose its failures in some reasoning tasks, raising fundamental questions about the nature of CoT reasoning. In this work, we propose a data distribution lens to understand when and why CoT reasoning succeeds or fails. We hypothesize that CoT reasoning reflects a structured inductive bias learned from in-distribution data, enabling models to conditionally generate reasoning trajectories that approximate those observed during training. As such, the effectiveness of CoT reasoning is fundamentally governed by the nature and degree of distribution discrepancy between training data and test queries. Guided by this lens, we dissect CoT reasoning via three dimensions: task, length, and format. To test the hypothesis, we introduce DataAlchemy, an abstract and fully controllable environment that trains LLMs from scratch and systematically probes them under various distribution conditions. Through rigorous controlled experiments, we reveal that CoT reasoning is a brittle mirage when it is pushed beyond training distributions, emphasizing the ongoing challenge of achieving genuine and generalizable reasoning.

Is Chain-of-Thought Reasoning of LLMs a Mirage? A Data Distribution Lens

Mid-training has become critical for enhancing the knowledge and reasoning ability of large language models (LLMs), especially through the utilization of large-scale synthetic data. However, existing data synthesis methods often generate simplistic and homogeneous QA pairs, with limited scale and diversity. To address this, we propose BoostQA, a novel framework designed to synthesize large-scale, diverse, and high-quality QA data for mid-training. BoostQA introduces model probes during mid-training for the first time and implements STEM-focused multi-grade synthesis to boost data diversity as well as high-difficulty synthesis to alleviate difficulty degradation, followed by answer refinement to further improve quality. Extensive experiments by mid-training Llama-3 8B demonstrate that using only 20B-token BoostQA data achieves a significant average improvement of **12.74%** on MMLU and CMMLU over the pre-training baseline. After mid-training on 500B tokens, including 100B-token BoostQA data, our model achieves SOTA average results across benchmarks among mainstream models of comparable size. BoostQA also demonstrates robust scalability, with performance consistently improving as model size, data volume, and initial FLOPs scale.

 Large-Scale Diverse Synthesis for Mid-Training

The deployment of Large Vision-Language Models (LVLMs) for real-world document question answering is often constrained by dynamic, user-defined policies that dictate information disclosure based on context. While ensuring adherence to these explicit constraints is critical, existing safety research primarily focuses on implicit social norms or text-only settings, overlooking the complexities of multimodal documents. In this paper, we introduce Doc-PP (Document Policy Preservation Benchmark), a novel benchmark constructed from real-world reports requiring reasoning across heterogeneous visual and textual elements under strict non-disclosure policies. Our evaluation highlights a systemic Reasoning-Induced Safety Gap: models frequently leak sensitive information when answers must be inferred through complex synthesis or aggregated across modalities, effectively circumventing existing safety constraints. Furthermore, we identify that providing extracted text improves perception but inadvertently facilitates leakage. To address these vulnerabilities, we propose DVA (Decompose–Verify–Aggregation), a structural inference framework that decouples reasoning from policy verification. Experimental results demonstrate that DVA significantly outperforms standard prompting defenses, offering a robust baseline for policy-compliant document understanding.

 Doc-PP: Document Policy Preservation Benchmark for Large Vision-Language Models

Task-oriented dialogue (TOD) systems are vital for facilitating complex, goal-directed interactions across sectors like customer support and online retail. However, they face persistent limitations: labor-intensive manual metadata tuning and sparse reinforcement learning (RL) rewards that fail to diagnose invocation errors. To address this, we propose ToolDNA, a dynamic adaptation framework enabling autonomous co-evolution of policy networks and tool metadata via RL, anchored by two synergistic loops. An RL loop optimizes policies by generating rollout trajectories (reasoning, actions, descriptive updates) from user inputs, with multi-dimensional rewards refining invocations. A tool metadata loop—coordinated by a dedicated Tool Manager—evolves metadata through policy-generated candidates during rollouts and Feedback LLM-derived refinements from historical data. These mutually reinforcing loops close traditional reward gaps, forming a closed-loop trial-error-reflection cycle for self-improvement. Extensive experiments on a real-world dataset of 3,100 customer service dialogues confirm ToolDNA's superiority, with notable gains over baselines: it achieves +11% problem resolution and +54% accuracy over commercial LLMs with prompt engineering; +25%/+35% over supervised fine-tuning; and +15%/+15% over traditional RL baseline. Linguistic analysis corroborates evolved metadata retain semantic intent while enhancing parseability. Case studies in two typical contexts, i.e., car inventory search and loan calculation, further validates its ability to resolve critical ambiguities. ToolDNA pioneers scalable self-improvement for robust, deployable tool-augmented agents with minimal human oversight. We release our code to facilitate future research.

ToolDNA: Autonomous Evolution of Tool Metadata for Robust Dialogue Agents

Retrieval-Augmented Generation (RAG) combines the language understanding and reasoning capabilities of large language models (LLMs) with external retrieval to produce domain-grounded responses. Effectively adapting RAG systems to domain-specific settings requires specialized, context-rich training data beyond general-purpose question-answering datasets. Here, we propose RAGen, a scalable and modular data-centric framework for generating domain-grounded question–answer–context (QAC) triples tailored to diverse RAG adaptation strategies. These QAC triples serve as training signals for multiple RAG adaptation approaches; in this work, we demonstrate their use for contrastive fine-tuning of embedding models and supervised fine-tuning of LLMs under retrieved contexts. RAGen generates QAC triples by identifying key concepts within documents, producing diverse questions guided by Bloom’s Taxonomy–inspired principles, and pairing them with precise answers extracted from relevant contexts. Its modular pipeline incorporates semantic chunking, hierarchical concept extraction, multi-chunk retrieval, and curated distractor contexts to encourage robust reasoning. Designed for scalability, RAGen efficiently handles large and evolving document corpora without redundant processing, making it particularly suitable for dynamic domains like enterprise knowledge bases.

Domain-Specific Data Generation Framework for RAG Adaptation

Psychiatric narratives encode patient identity not only through explicit identifiers but also through idiosyncratic life events embedded in clinical structure. Existing de-identification approaches, including PHI masking and LLM-based synthetic rewriting, operate at the text level and offer limited control over which semantic elements are preserved or altered. We introduce \textbf{Anonpsy}, a de-identification framework that reformulates the task as graph-guided semantic rewriting. Anonpsy (1) converts each narrative into a semantic graph encoding clinical entities, temporal anchors, and typed relations; (2) applies graph-constrained perturbations that modify identifying context while preserving clinical structure; and (3) regenerates text via graph-conditioned LLM generation. Evaluated on 90 clinician-authored psychiatric case narratives, Anonpsy preserves diagnostic fidelity while achieving consistently low re-identification risk under expert, semantic, and GPT-5-based evaluations. Compared with a strong LLM-only rewriting baseline, Anonpsy yields substantially lower semantic similarity and identifiability. These results demonstrate that explicit structural representations combined with constrained generation provide an effective approach to de-identification for psychiatric narratives.

Downloads

Next from ACL 2026

Meta-Reasoner: Dynamic Guidance for Optimized Inference-time Reasoning in Large Language Models

Similar lecture

Rethinking Entropy Interventions in RLVR: An Entropy Change Perspective