The paper investigates on‑policy distillation (OPD) as a preparatory step for reinforcement learning (RL). It shows that students initialized with OPD achieve higher final RL performance than those trained directly with RL or with supervised fine‑tuning followed by RL, even when OPD offers little immediate accuracy gain. The study also finds that the choice of distillation objective (reverse‑KL vs forward‑KL) and the source of trajectories influence OPD’s effectiveness at different stages of RL training.
By Shuai Dong, Yongfu Zhu, Yuqi Xu, Weichu Xie, Liuwenpu, Ziyue Wang, Kaiwen Tuo, Congcong Wang, Siyuan Wang, Wenqi Shao, Shuai Yang, Ji Zhao, Caoyuan Ma, Wenzheng Chang, Taiqiang Wu, Xinlei Yu, Hongrui Wu, Xiaoxuan He, Fangke Chen, Dianyi Wang, Kanghui Tian, Sirry Chen, Xingyu Liu, Xiangnan Wu, Jiawei Guo, Haowen Hou, LingHan Chen, Zhongyu Wei, Jiaqi Wang
The paper introduces Planned Test-Time Scaling (PTTS), a method that replaces independent sampling of reasoning branches with a coordinated joint policy. PTTS uses a planner to generate distinct solution outlines for each branch and an executor to produce full solutions, thereby improving coverage of complementary reasoning modes. Two variants—PTTS‑ZS (zero‑shot) and PTTS‑RL (reinforcement‑learned)—demonstrate significant gains on five mathematical reasoning benchmarks, with PTTS‑RL achieving up to a 13.4‑point improvement in pass@64 over repeated sampling.
By Xueqing Wu, Langxing Bai, Hritik Bansal, Po-Nien Kung, Shuo Li, Hao Liu, Nanyun Peng, Kai-Wei Chang
Quantum Reinforcement Learning for Cost and Delay Tradeoffs in Quantum Cloud Orchestration proposes QRLQ, a scheduling framework that integrates parameterised quantum circuits with a dueling double deep Q‑network to balance execution cost and delay in quantum‑as‑a‑service environments. Simulation results show QRLQ outperforms heuristic baselines, achieving 5‑11% lower mean cost and up to 82% lower mean delay while maintaining fidelity within 2% of a fidelity‑greedy policy. Compared to a classical deep reinforcement learning baseline, QRLQ delivers comparable performance with 72% fewer trainable parameters.
By An N. H. Phan, Dang Van Huynh, Muhammad Usman, Hoa T. Nguyen
MobileGym is a browser-hosted, lightweight simulation platform designed for mobile GUI agent research. It offers verifiable outcome signals via deterministic, JSON-based state judging and supports scalable online reinforcement learning with hundreds of parallel instances on a single server. The platform includes a declarative task-definition framework, a structured AnswerSheet protocol, and a benchmark of 416 parameterized tasks across 28 apps, demonstrating strong sim-to-real transfer in a case study.
By Dingbang Wu, Rui Hao, Haiyang Wang, Shuzhe Wu, Han Xiao, Zhenghong Li, Bojiang Zhou, Zheng Ju, Zichen Liu, Lue Fan, Zhaoxiang Zhang
The paper introduces VHD-Play, a pipeline that first samples and solves a mathematical model before generating agentic reinforcement learning environments, ensuring that dynamics and evaluation are aligned from the outset. This approach yields 3,300 diverse environments at a low cost and significantly improves the performance of a large language‑model agent (Qwen3.6‑35B‑A3B) across multiple diagnostic families and external benchmarks. The study demonstrates that stateful interaction is a key factor in learning gains and that scaling the training substrate can further enhance performance.
By Xinjie Shen, Wei Fan, Xudong Guo, Jianhong Tu, Yang Su, Chuqiao Kuang, Yinger Zhang, Dayiheng Liu
Omni-Decision is an omni-modal agent that uses evidence-ledger planning to manage noisy multimodal observations. It replaces the growing dialogue history with a compact evidence ledger that tracks missing, confirmed, and conflicting evidence, allowing the planner to operate on a streamlined context. The system achieves state‑of‑the‑art accuracy on OmniGAIA and WorldSense while operating at a fraction of the cost of larger models.
By Ming Ma, Yi Zhu, Yiran Zhong, Feida Zhu, Yuhao Wang, Junhan Shi, Lingrui Mei, Tianming Yang, Steven Hoi
The paper introduces a new algorithm for the Multi‑Armed Bandit problem that prioritizes selecting the arm with the lowest variance rather than the highest expected reward, using a softmax policy parameterization. It constructs an unbiased estimate of the minimal‑variance objective by drawing two independent samples from the chosen arm and proves convergence under natural conditions. Numerical experiments demonstrate the algorithm’s practical behavior and provide implementation guidance, while also addressing general risk‑aware trade‑offs between average reward and variance.
By Gabriel Turinici
EBRL is an asynchronous embodied reinforcement learning training system that overlaps rollout and training stages, pipelines simulation and generation across environment groups, and eliminates synchronization stalls. It employs a fine‑grained resource manager that pools CPU cores and GPU streaming multiprocessors, dynamically adjusting resource quotas and batch sizes based on stage profiles and runtime feedback. Experiments on RLinf with four policies and four simulation benchmarks show EBRL improves rollout throughput by 1.30–3.47× and training convergence by 2.5× over state‑of‑the‑art embodied RL systems.
By Liang Mi, Weijun Wang, Bowen Gao, Tianze Yu, Zixu Hao, Han Xiao, Xin Ding, Mingzhe Huang, Xin He, Lu Shi, Hao Wu, Haipeng Dai, Guihai Chen, Yunxin Liu, Ting Cao
Limiting‑Kernel Q(λ) (LKQL) is an off‑policy value estimator that blends n‑step truncation with a long‑horizon approximation based on the limiting kernel. It maintains the computational efficiency of n‑step methods while improving policy evaluation accuracy, especially for long‑horizon tasks. The authors prove faster convergence of LKQL’s operator under aperiodicity and near‑on‑policy conditions, and demonstrate empirical gains on MuJoCo continuous‑control benchmarks.
By Tolga Ok, Arman Sharifi Kolarijani, Peyman Mohajerin Esfahani, Mohamad Amin Sharifi Kolarijani
ForgetMimic is a motion-level unlearning method for reinforcement learning-based humanoid control. It selectively degrades performance on a chosen subset of motions while preserving the policy’s effectiveness on the remaining motions. Experiments on Unitree G1 and H2 robots across 12 motions show that the method successfully removes memory of designated motions without affecting other behaviors.
By Xukun Luan, Zhongxiang Lei, Chen Gong, Shaowei Li, Yuanguo Bi, Jinyan Liu
The paper introduces Hierarchical Reinforcement and Collective Learning (HRCL), a framework that combines multi‑agent reinforcement learning (MARL) with decentralized coordination. HRCL uses MARL at a high level to generate strategic guidance that limits the decision space for low‑level agents, enabling efficient short‑term coordination while considering long‑term effects. Experiments on synthetic, energy‑management, and drone‑swarm scenarios demonstrate faster convergence and significant reductions in system‑wide and individual costs compared to standalone MARL.
By Chuhao Qin, Evangelos Pournaras
arXiv:2609.27987v1 Announce Type: new
Abstract: Large language models (LLMs) have made substantial progress on medical question-answering, yet effective medical dialogue also requires learning to ask...
By Chenxuan Li, Jiayi Wan, Xinrong Chen, Zhongyu Zhao, Xuecheng Shang, Peixing Wan
arXiv:2603.08166v2 Announce Type: replace
Abstract: Automated Drug Combination Extraction (DCE) from large-scale biomedical literature is crucial for advancing precision medicine and pharmacological...
By Zhijun Wang, Ling Luo, Dinghao Pan, Huan Zhuang, Lejing Yu, Yuanyuan Sun, Hongfei Lin
The paper introduces Reinforcement Learning with Decomposed Subtasks (RLDS), a method that splits trajectory rewards into per‑subtask shares before policy updates, replacing the scalar advantage used in Group Relative Policy Optimization (GRPO). RLDS employs Subtask‑Decomposed Advantage Estimation (SDAE) to compute group‑relative advantages and distribute credit to tokens based on subtask importance, focusing on steps where a reflection marks a subtask as consequential. Experiments on four benchmarks—FrozenLake, HotpotQA, ScienceWorld, and DeepResearch—show that RLDS improves performance on high‑heterogeneity tasks (ScienceWorld and FrozenLake) and is more compute‑efficient than scalar GRPO for long rollouts.
By Mattie Terzolo, Mikolaj Sacha, Ayan Sinha, Andrew Rabinovich
arXiv:2609.27869v1 Announce Type: new
Abstract: Long-horizon multimodal agents rely on specialized capabilities for perception, retrieval, reasoning, verification, and execution. Existing designs typ...
By Wenhao Yuan, Chenchen Lin, Jian Chen, Jinfeng Xu, Shuo Yang, Edith Cheuk-Han Ngai
The paper introduces Anchored Neighborhood Optimization (ANO), a new policy‑optimization method that directly designs a smooth, bounded gain field for the probability‑ratio surrogate objective. ANO anchors the identity map at a ratio of one, peaks at a specified trust‑region boundary, and limits the influence of extreme off‑policy samples while providing a bounded, redescending pull on outliers. Empirical results show ANO consistently outperforms existing methods on Atari and MuJoCo benchmarks, and it remains robust under aggressive learning‑rate settings.
By Yiheng Zhang, Yiming Wang, Kaiyan Zhao, Zhenglin Wan, Jiayu Chen, Leong Hou U
The paper introduces Wasserstein‑Tilted Flow Maps (WTF), a simulation‑free reinforcement learning method that fine‑tunes pre‑trained flow‑based generative models by adding an optimal transport regularizer derived from the model’s drift. Unlike traditional KL‑reward tilting, WTF transports individual samples toward higher reward, framing the problem as a deterministic optimal control task on the flow map. Experiments on ImageNet‑256 and text‑to‑image demonstrate that WTF achieves higher reward and comparable or better diversity while reducing training compute by up to 280×.
By Abbas Mammadov, Jerry Y. Huang, Justin Lin, Partha Kaushik, Sheel Shah, Kartik Nair, Yee Whye Teh, Nicholas M. Boffi
arXiv:2609. 27594v1 Announce Type: new Abstract: We study the problem of computational efficiency for linear bandits in high-dimensional settings with a finite arm set.
By Hantao Yang, Hong Xie, Defu Lian
The paper introduces RECAP, a redundancy-aware credit assignment method that improves reasoning efficiency in large language models by assigning credit to each reasoning step based on its downstream role and contribution to the correct answer. RECAP uses a semantic dependency graph to measure structural responsibility and evaluates step efficacy via changes in gold-answer log-likelihood, enabling step-specific updates without requiring a separate reward model or concise trajectories. Experiments on two 7B models across four mathematical reasoning benchmarks show that RECAP enhances the accuracy-efficiency trade-off, boosting pass@1 by 2.0–3.7 percentage points while cutting reasoning tokens by 8–31% compared to GRPO.
By Yuqing Zhou, Hong Wang, Manqing Mao, Zhuoer Wang, Samson Koelle, Jie Yuan, Yanjun Lin, James Feng, Nikki Lijing Kuang, Ziwei Zhu, Wei Niu
arXiv:2609.27389v1 Announce Type: cross
Abstract: Audio language models understand what is said far better than how it sounds. Closing this gap takes more than data. Detailed acoustic annotation is c...
By Yuxiang Wang, Shengbo Cai, Yingda Shen, Ming-Hao Hsu, Qinke Ni, Liqiang Zhang, Teddy Sun, Steve Yevs, Zhizheng Wu