The paper introduces MedGym, a benchmark environment for dynamic medical treatment recommendation that models patient evolution in continuous time using Physics-Informed Neural Networks. It addresses gaps in existing reinforcement learning (RL) approaches by allowing evaluation of RL methods under irregular measurement intervals, personalized treatment responses, and safety considerations. MedGym enables direct comparison between discrete-time and continuous-time RL methods and supports clinically relevant metrics such as personalization and trajectory-level safety.
By Yuepeng Wang, Ken Kawano, Yoshihiko Fujisawa, Yongqi Zhou, Akifumi Wachi, Mehrshad Sadria, Lei Zhou, Richard Weiss, Katsuki Fujisawa, Ying Chen, Xin Liu, Kyoung-Sook Kim, Xiao Hu, Sebastien Gros, Xun Shen
arXiv:2603.18532v3 Announce Type: replace-cross
Abstract: The strong performance of large vision-language models (VLMs) trained with reinforcement learning (RL) has motivated similar approaches for f...
By Andrew Choi, Xinjie Wang, Zhizhong Su, Wei Xu
arXiv:2609.23697v1 Announce Type: cross
Abstract: Multi-teacher on-policy distillation allows a student to learn from complementary specialists on its own trajectories. Domain-routed approaches, howe...
By Jie Sun, Mao Zheng, Mingyang Song, Zeyuan Liu, Gengsheng Li, Houcheng Jiang, Yilin Cheng, Bichuan Feng, Yuchen Cai, Junfeng Fang, Xiang Wang
arXiv:2609.23457v1 Announce Type: new
Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) is expanding from tasks with well-defined correctness signals, such as mathematics and code, towa...
By Hao Li, Zhengkun Zhang, Gangqiang Hu, Zhen Zhang, Yude Gao, Dai Dai, Jing Liu
Exhaustively evaluating every candidate LLM configuration on every benchmark item to identify a high-performing one is costly. We formulate configuration selection as a cost-aware Bayesian bandit prob...
Recent advances in video generative models have enabled high-fidelity, temporally coherent video generation. However, these models often struggle to satisfy prompts requiring specialized knowledge, sp...
As Large Language Model (LLM) agents are applied in continuously interactive environments, driving the evolution of their own capabilities becomes a core problem for achieving long-term autonomy. Curr...
In many real-world planning tasks, agents must tackle uncertainty about the environment's state and variability in the outcomes of any chosen policy. We address both forms of uncertainty as a first st...
Automated biomedical evidence synthesis depends on retrieving published studies, but the biomedical literature is systematically skewed toward positive findings. Deeper retrieval can therefore make a...
The paper proves that using a permissive safety filter in reinforcement learning does not compromise asymptotic performance. By formalizing safety through a safety‑critical Markov decision process and a filtered MDP, the authors show that optimal policies in the filtered MDP achieve the same return as the best safe policy in the original setting. Experiments on Safety Gymnasium confirm zero violations during training and performance that matches or exceeds unfiltered baselines.
By Donggeon David Oh, Duy P. Nguyen, Haimin Hu, Jaime Fern\'andez Fisac
The paper introduces RAVEL, a retrieval‑aware online reinforcement learning framework designed to improve interactive retrieval under partial evidence. RAVEL begins with supervised question generation, directly observes the top‑4 retrieval candidates, and refines its question policy using rank feedback from the full question‑answer‑retrieval loop. Experiments on the Interactive‑PEDES dataset demonstrate that RAVEL progressively enhances retrieval performance over five interaction rounds, reallocating questioning toward localized open‑ended attributes that yield the greatest gains on challenging queries.
By Lyucheng Qian, John Yuehan Zhang, Pingyu Wang
arXiv:2609. 22041v1 Announce Type: new Abstract: Reinforcement learning is increasingly used to align image generators with reward signals, and Flow-GRPO recently extended this paradigm to flow-matching models by treating the denoising sampler as a stochastic policy that can be optimized from reward feedback.
By Yufeng Wang, Parivesh Priye, Meeshawn Marathe, Ramit Pahwa
The paper compares two leading motion‑imitation reinforcement learning pipelines—HyFyDy, which uses detailed musculotendon modeling, and MuJoCo, which focuses on computational speed. Using the same human motion‑capture and EMG data, both pipelines reproduce kinematics similarly, but HyFyDy’s muscle activation predictions align more closely with experimental EMG (RMSE 0.164, r = 0.4) than MuJoCo’s (RMSE 0.344, r = 0.11). The authors conclude that HyFyDy’s higher physiological realism makes it currently more suitable for musculoskeletal modeling, though both systems need further development for GPU‑parallelizable environments and robotic assistive‑device design.
By Ayah G. Ahmad, Claire E. Borden, Maegan Tucker
The paper introduces OverThink, a slowdown attack that forces reasoning language models (RLMs) to produce many more reasoning tokens while still giving correct answers. By injecting decoy reasoning problems—such as Markov decision processes, language translation, or graphic comprehension—into the model’s context, attackers can dramatically increase token generation (up to 46× on SQuAD and 17× on coding agents). The study evaluates the attack on both proprietary and open-source RLMs across multiple datasets, explores multimodal and coding‑agent variants, and tests several defenses, concluding that defending against OverThink is challenging and that newer RLMs are even more vulnerable due to higher per‑token costs and increased reasoning token usage.
By Abhinav Kumar, Jaechul Roh, Ali Naseh, Marzena Karpinska, Mohit Iyyer, Amir Houmansadr, Eugene Bagdasarian
The paper proposes a method for culturally aligning large language models (LLMs) using soft prompt tuning optimized via Differential Evolution (DE). Unlike traditional fine‑tuning or reinforcement learning, this approach keeps model weights frozen and requires no preference data, instead leveraging aggregated survey scores from Hofstede's Value Survey Module (VSM13). Experiments on four countries and four instruction‑tuned models show that DE‑optimized prompts reduce cultural discrepancy, improve agreement with the World Values Survey, and are preferred in blinded pairwise evaluations by LLM judges.
By Reem I. Masoud, Martin Ferianc, Philip Treleaven, Miguel Rodrigues
The paper proposes a method for learning task-relevant representations in deep reinforcement learning by maximizing rollout total correlation, which captures the correlation among all learned representations and actions across entire trajectories. It introduces two complementary lower bounds—one generative and one discriminative—along with chunk‑wise mini‑batching to improve this objective, and also proposes an intrinsic reward derived from the learned representation to enhance exploration. Experiments on challenging image‑based simulated control tasks demonstrate improved sample efficiency and robustness to white noise and natural video backgrounds compared to leading baselines.
By Bang You, Huaping Liu, Jan Peters, Oleg Arenz
ArenaFlow is a hierarchical credit propagation framework designed to improve reinforcement learning for open-ended agent tasks. It uses tournament-based relative ranking to generate trajectory-level rewards and structured reflective evaluation to identify pivotal success steps, reusable strategy skills, and skill usage attribution. The framework propagates advantages to high-confidence steps and maintains a global skill memory, enabling more targeted optimization and reusable skill priors for future exploration.
By Qiang Zhang, Ruixue Ding, Fanrui Zhang, Xi Chen, Boli Chen, Shihang Wang, Yinfeng Huang, Yi Zheng, Pengjun Xie, Kaipeng Zhang, Jiawei Liu, Zheng-Jun Zha
MIRAGE is a new inference-time framework that enhances large language models by using a Selector to choose effective conceptual perspectives and a Reasoner to solve tasks step-by-step, aggregating multiple perspectives when needed. It is inspired by human cognitive flexibility and is designed to improve performance on complex mathematical, scientific, and logical problems. Experiments on GSM8K, MATH500, MMLU-Pro, and Game-of-24 show that MIRAGE outperforms Chain-of-Thought and diverse prompting ensembles, boosting accuracy with minimal inference overhead.
By Arash Lagzian, Srinivas Anumasa, Dianbo Liu
GEM-MPC is a reinforcement learning method that blends MPPI planning with policy learning to balance exploration and exploitation in high-dimensional continuous control tasks. It trains a policy to clone the planner while also maintaining a KL-regularized policy that explores around the planner’s suggestions, thereby improving the synergy between planning and learning. The approach introduces Gated Prior Distillation, which selectively updates policies from stored planning distributions only when they offer better targets, reducing the influence of stale data without costly reanalysis. Across continuous-control benchmarks, GEM-MPC outperforms existing planning-based baselines while using lower computational budgets.
By Alvaro Serra-Gomez, Thomas Moerland
ULTRA is a unified framework for autonomous humanoid whole-body locomotion and manipulation that overcomes limitations of prior methods by combining a physics-driven neural retargeting algorithm with a multimodal controller. The retargeting algorithm translates large-scale motion capture data into physically plausible humanoid motions, while the controller learns to handle both dense motion references and sparse task specifications using a range of sensory inputs, from accurate motion-capture states to noisy egocentric vision. In simulation and on a real Unitree G1 humanoid, ULTRA demonstrates improved generalization and robustness, enabling coordinated whole-body behavior from sparse intent without relying on test-time reference motions.
By Xialin He, Sirui Xu, Xinyao Li, Runpei Dong, Liuyu Bian, Yu-Xiong Wang, Liang-Yan Gui