United States

Reinforcement Learning with Verifiable Rewards (RLVR) serves as a cornerstone technique for enhancing the reasoning capabilities of Large Language Models (LLMs). However, its training is often plagued by \emph{entropy collapse}, a rapid decline in policy entropy that limits exploration and undermines training effectiveness. While recent works attempt to mitigate this issue via several heuristic entropy interventions, the underlying mechanisms remain poorly understood. In this work, we conduct comprehensive theoretical and empirical analyses of entropy dynamics in RLVR, offering two main insights: (1) We derive a tight approximation for token-level entropy change at each update step, revealing four governing factors and providing a unified theoretical framework of how existing methods influence entropy; (2) We reveal a fundamental limitation of recent approaches: they rely on heuristic adjustments to one or two of these factors, leaving other relevant factors unconsidered, thus inherently limiting their effectiveness. Motivated by these findings, we propose STEER, a principled entropy-modulation method that adaptively reweighs tokens based on theoretically-estimated entropy variations. Extensive experiments across six mathematical reasoning and three coding benchmarks demonstrate that STEER effectively mitigates entropy collapse and consistently outperforms state-of-the-art baselines.

ACL 2026

Rethinking Entropy Interventions in RLVR: An Entropy Change Perspective

### Welcome to IJCNLP-AACL 2025! 
 It is a great honor to host this joint conference in Mumbai, India, from December 20 to 24, 2025. The joint conferences of IJCNLP and AACL are organized with alternating leadership in the Asia-Pacific region. The event is run by the Asian Federation of Natural Language Processing (AFNLP) in odd years, and by AACL in even years, while it is organized solely by ACL when the annual ACL meeting is held in the region. This year, the conference is primarily organized by AFNLP. 
*Kentaro Inui
MBZUAI, UAE
General Chair, IJCNLP-AACL 2025* 
Read full message and download the Conference Handbook [**here**](https://drive.google.com/file/d/1UTwxkAqSqI-GAoJC3wE1zZt5VP1Y8GX0/view?usp=sharing).

IJCNLP-AACL 2025

The 14th IJCNLP & 4th AACL will be held in Mumbai, India from December 20th to December 24th, 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.

Ensuring fairness in social survey simulation is critical, as biased outputs can misrepresent underrepresented groups. This issue is growing as large language models (LLMs) are increasingly used for this task. However, standard fine-tuning based on Empirical Risk Minimization (ERM) often under-optimizes minority groups, causing substantial subgroup disparities. Distributionally robust Optimization (DRO) methods reduce worst-case errors, but their strict worst-case selection can lead to noisy and unstable optimization under demographic sparsity. These issues create intertwined challenges for fairness, convergence and stability. We propose SAFO, a dynamic utility–fairness optimization framework for LLM-based survey simulation that explicitly targets both fairness and training stability. SAFO combines (i) an Optimizer that preserves mean-loss utility, (ii) an Adversary that performs temperature-controlled, EMA-smoothed and loss-driven group reweighting, and (iii) a Nash-inspired Regulator that adaptively adjusts the utility–fairness trade-off by tracking weak-group gains and collateral utility damages. Experiments on three large-scale survey datasets from China, the U.S., and Europe show that SAFO consistently improves minority performance and social-welfare metrics. It reduces worst-group gaps by up to 12.7%, maintains overall accuracy with a mean change of less than 0.3% and lowers variance across random seeds. Our code is available at https://github.com/PiLab-ZJU/SAFO.

SAFO: Stable Adaptive Fairness Optimization for LLM-Based Social Survey Simulation

Realizing endogenous narrative evolution in LLM-based multi-agent systems is hindered by the inherent stochasticity of generative emergence. In particular, long-horizon simulations suffer from social memory stacking, where conflicting relational states accumulate without resolution, and narrative-spatial dissonance, where spatial logic detaches from the evolving plot. To bridge this gap, we propose EvoSpark, a framework specifically designed to sustain logically coherent long-horizon narratives within Endogenous Interactive Agent Societies. To ensure consistency, the Stratified Narrative Memory employs a Role Socio-Evolutionary Base as living cognition, dynamically metabolizing experiences to resolve historical conflicts. Complementarily, a Generative Mise-en-Scène mechanism enforces Role-Location-Plot alignment, synchronizing character presence with the narrative flow. Underpinning these is the Unified Narrative Operation Engine, which integrates an Emergent Character Grounding Protocol to transform stochastic sparking into persistent characters. This engine establishes a substrate that expands a minimal premise into an open-ended, evolving story world. Experiments demonstrate that EvoSpark significantly outperforms baselines across diverse paradigms, enabling the sustained generation of expressive and coherent narrative experiences.

EvoSpark: Endogenous Interactive Agent Societies for Unified Long-Horizon Narrative Evolution

Recent advances in Vision-Language Models (VLMs) have enabled mobile agents to perceive and interact with real-world mobile environments based on human instructions. However, the current fully autonomous paradigm poses potential safety risks when model understanding or reasoning capabilities are insufficient. To address this challenge, we first introduce \textbf{InquireBench}, a comprehensive benchmark specifically designed to evaluate mobile agents' capabilities in safe interaction and proactive inquiry with users, encompassing 5 categories and 22 sub-categories, where most existing VLM-based agents demonstrate near-zero performance. In this paper, we aim to develop an interactive system that actively seeks human confirmation at critical decision points. To achieve this, we propose \textbf{InquireMobile}, a novel model inspired by reinforcement learning, featuring a two-stage training strategy and an interactive pre-action reasoning mechanism. Finally, our model achieves an 46.8% improvement in inquiry success rate and the best overall success rate among existing baselines on InquireBench. The project page is available at \url{https://bit-aqh.github.io/InquireMobile/homepage/}.

InquireMobile: Teaching VLM-based Mobile Agent to Request Human Assistance via Reinforcement Fine-Tuning

While Chain-of-Thought (CoT) reasoning empowers Large Language Models (LLMs) to tackle complex tasks, its reliance on discrete token decoding imposes an inherent Discreteness Bottleneck, limiting expressiveness within a restricted vocabulary space. Existing continuous reasoning approaches, such as SoftCoT, mitigate this but typically rely on external auxiliary models, resulting in complex deployment and fractured inference pipelines. To address these challenges, we propose Self-SoftCoT, a self-contained framework that enables a frozen LLM to internally generate and consume latent thoughts without external assistants. By establishing a single-stream "Thinking → Speaking" closed-loop, we decouple latent planning from explicit generation. Furthermore, we adopt Group Sequence Policy Optimization (GSPO) to stabilize learning and employ Position-Aware Independent Projection to mitigate representation homogenization. Experimental results on five reasoning benchmarks demonstrate that our method significantly improves the reasoning performance of frozen LLMs. Specifically, our Qwen2.5-based model uses only N=2 soft tokens to outperform the SoftCoT baseline (N=4), improving the average accuracy from 75.06% to 78.42%. Similarly, LLaMA-3.1 performance increases from 70.52% to 74.55%.

Self-SoftCoT: A Self-Consistent Framework via Position-Aware Latent Space Reinforcement Learning

Reinforcement learning (RL) paradigms have demonstrated strong performance on reasoning-intensive tasks such as code generation. However, limited trajectory diversity often leads to diminishing returns, which constrains the achievable performance ceiling. Search-enhanced RL alleviates this issue by introducing structured exploration, which remains constrained by the single-agent policy priors. 
Meanwhile, leveraging multiple interacting policies can acquire more diverse exploratory signals, but existing approaches are typically decoupled from structured search. We propose \textbf{MARS$^2$} (Multi-Agent Reinforced Tree-Search Scaling), a unified RL framework in which multiple independently-optimized agents collaborate within a shared tree-structured search environment. MARS$^2$ models the search tree as a learnable multi-agent interaction environment, enabling heterogeneous agents to collaboratively generate and refine candidate solutions within a shared search topology. To support effective learning, we introduce a path-level group advantage formulation based on tree-consistent reward shaping, which facilitates effective credit assignment across complex search trajectories. Experiments on code generation benchmarks show that MARS$^2$ consistently improves performance across diverse model combinations and training settings, demonstrating the effectiveness of coupling multi-agent collaboration with tree search for enhancing reinforcement learning. Our code is publicly available at \url{https://github.com/TsinghuaC3I/MARTI}.

MARS$^2$: Scaling Multi-Agent Tree Search via Reinforcement Learning for Code Generation

Large language models (LLMs) have demonstrated strong reasoning capabilities through step-by-step chain-of-thought (CoT) reasoning. Nevertheless, at the limits of model capability, CoT often proves insufficient, and its strictly sequential nature constrains test-time scalability. A potential alternative is divide-and-conquer (DAC) reasoning, which decomposes a complex problem into subproblems to facilitate more effective exploration of the solution space. Although promising, our analysis reveals a fundamental misalignment between general-purpose post-training and DAC-style inference, which limits the model’s capacity to fully leverage this potential. To bridge this gap and fully unlock LLMs’ reasoning capabilities on the most challenging tasks, we propose an end-to-end reinforcement learning (RL) framework to enhance their DAC-style reasoning capacity. At each step, the policy decomposes a problem into a group of subproblems, solves them sequentially, and addresses the original problem conditioned on the subproblem solutions, with both decomposition and solution integrated into RL training. Under comparable training settings, our DAC-style framework endows the model with a higher performance ceiling and stronger test-time scalability, surpassing CoT by 8.6% in Pass@1 and 6.3% in Pass@32 on competition-level benchmarks. The code is available at the [provided link](https://github.com/MasterVito/DAC-RL).

Training LLMs for Divide-and-Conquer Reasoning Elevates Test-Time Scalability

We investigate the efficacy of Large Language Models (LLMs) in detecting implicit and explicit hate speech, examining how models with minimal safety alignment (uncensored) compare with more heavily aligned (censored) counterparts in a deployed-model setting when deployed using political personas. While uncensored models are often framed as offering a less constrained perspective, our results reveal a trade-off: censored models outperform their uncensored counterparts in both accuracy and robustness, achieving 69.0\% versus 64.1\% strict accuracy. However, this higher performance is also associated with greater resistance to persona-based influence, while uncensored models are more malleable to ideological framing. Furthermore, we identify critical failures across all models in understanding nuanced language such as irony. We also find alarming fairness disparities in performance across different targeted groups and systemic overconfidence that renders self-reported certainty unreliable. These findings challenge the notion of LLMs as objective arbiters and highlight the need for more sophisticated auditing frameworks that account for fairness, calibration, and ideological consistency. Taken together, these results point to censorship-as-deployed rather than safety alignment in isolation as the more appropriate frame for interpreting model differences.

Confident, Calibrated, or Complicit: Safety Alignment and Ideological Bias in LLM Hate Speech Detection

Word sense disambiguation (WSD) is a foundational task in natural language processing. 
Recent research has reformulated WSD for large language models (LLMs) as a generative task, where the model produces a definition to convey the intended meaning of an ambiguous word in context.
In practice, most existing approaches implement this formulation through straightforward supervised fine-tuning, which tends to prioritize superficial context-to-gloss memorization over true contextual sense discrimination, leading to degraded performance on less frequent senses (LFS), particularly in unseen settings.
To address this issue, we propose WSDPO, a training framework for generative WSD with chain-of-thought (CoT) and preference optimization. 
WSDPO consists of three stages: (1) disambiguation-aware CoT construction, which produces training data containing explicit disambiguation steps for the later stage;
(2) disambiguation-guided supervised fine-tuning, which explicitly trains the model to discriminate word sense before generating the final definition; and
(3) preference-based optimization, which further strengthens the model’s ability to generate sense-faithful definitions by optimizing it using preference pairs constructed from multiple sampled CoT outputs.
Extensive experiments across benchmark datasets and multiple backbone LLMs demonstrate that WSDPO achieves substantial performance gains on rare and unseen settings, and exhibits strong generalization in standard evaluation settings.

WSDPO: A Generative Word Sense Disambiguation Framework with Chain-of-Thought and Preference Optimization

Fine-tuning large language models (LLMs) is an effective approach to enhancing their performance on specialized downstream tasks. Among the various techniques, low-rank adaptation has garnered significant attention due to its ability to maintain the full performance of fine-tuning while enhancing computational efficiency. However, existing approaches often rely on manually specified and fixed hyperparameters to identify the trainable components within weight matrices, resulting in suboptimal performance and low parameter efficiency. This paper presents a novel Learnable Low-Rank Adaptation (LeLoRA) framework that utilizes dynamically learned fine-tuning strategies to facilitate the effective adaptation of LLMs. Our framework integrates an LLM with a policy network that automatically and adaptively generates matrix-specific adaptation strategies to identify the trainable components of each weight matrix, taking into account their unique characteristics, such as singular values and matrix norms. A reinforcement learning-based optimization algorithm is then employed to iteratively update the LLM and the policy network, ensuring that the generated strategies adapt in real time to the evolving states of the LLM. Extensive experiments have been conducted across various natural language processing and multimodal tasks. The results across ten different LLMs, ranging from 125M to 70B parameters, provide compelling evidence that LeLoRA consistently outperforms existing baselines in adapting LLMs. Moreover, analytical experiments provide valuable insights into the effectiveness of the generated strategies.

LeLoRA: Learnable Low-Rank Adaptation of Large Language Models

Editing large language models is challenging as incorporating new knowledge often requires sequential parameter updates while maintaining model capability. In this work, we experimentally observe that sequential knowledge updating under the locate-then-edit framework can introduce safety risks, regardless of whether the knowledge being edited is benign or malicious. We propose a novel model editing approach that estimates safety transforms and identifies corresponding safety direction in the neural activation space, and then aligns neural activation updates and network parameter updates under the safety constraints, resulting in a safety-aware model editing approach. We evaluate our approach on open-source LLMs, Llama-3-8B-Instruct, Qwen3-4B-Instruct and Qwen2.5-14B-Instruct, using the benchmark datasets ZsRE and COUNTERFACT, as well as the malicious dataset Mal-KSet. Experimental results demonstrate that our approach effectively reduces unsafe responses to malicious queries while preserving the effectiveness of model editing.

Downloads

Next from ACL 2026

SAFO: Stable Adaptive Fairness Optimization for LLM-Based Social Survey Simulation

Similar lecture

How Well Does First-Token Entropy Approximate Word Entropy as a Psycholinguistic Predictor?

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from ACL 2026

SAFO: Stable Adaptive Fairness Optimization for LLM-Based Social Survey Simulation

Similar lecture

How Well Does First-Token Entropy Approximate Word Entropy as a Psycholinguistic Predictor?

Downloads