Generative Online Reinforcement Learning

Chubin Zhang, Zhenglin Wan, Feng Chen, Fuchao Yang, Lang Feng, Yaxin Zhou, Xingrui Yu, Yang You, Ivor Tsang, Bo An
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:158816-158837, 2026.

Abstract

Reinforcement learning (RL) faces a persistent tension: policies that are stable to optimize (e.g., Gaussians) are often too simple to represent the multimodal action distributions required for complex control. Conversely, expressive generative policies—such as diffusion and flow matching—can be difficult to optimize in online RL due to intractable likelihoods and gradients propagating through long sampling chains. We address this tension with a key structural principle: decoupling optimization from generation. Building on this, we introduce GoRL (Generative Online Reinforcement Learning), an algorithm-agnostic framework that trains expressive policies from scratch by confining policy optimization to a tractable latent space while delegating action synthesis to a conditional generative decoder. Using a two-timescale alternating schedule and anchoring decoder refinement to a fixed prior, GoRL enables stable optimization while continuously expanding expressiveness. Empirically, GoRL consistently outperforms unimodal and generative baselines across diverse continuous-control tasks. Notably, GoRL achieves returns exceeding 870 on HopperStand, more than $3\times$ the strongest baseline; on high-dimensional humanoid tasks, it further outperforms the strongest non-GoRL baseline by over an order of magnitude.

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-zhang26fu, title = {Generative Online Reinforcement Learning}, author = {Zhang, Chubin and Wan, Zhenglin and Chen, Feng and Yang, Fuchao and Feng, Lang and Zhou, Yaxin and Yu, Xingrui and You, Yang and Tsang, Ivor and An, Bo}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {158816--158837}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/zhang26fu/zhang26fu.pdf}, url = {https://proceedings.mlr.press/v306/zhang26fu.html}, abstract = {Reinforcement learning (RL) faces a persistent tension: policies that are stable to optimize (e.g., Gaussians) are often too simple to represent the multimodal action distributions required for complex control. Conversely, expressive generative policies—such as diffusion and flow matching—can be difficult to optimize in online RL due to intractable likelihoods and gradients propagating through long sampling chains. We address this tension with a key structural principle: decoupling optimization from generation. Building on this, we introduce GoRL (Generative Online Reinforcement Learning), an algorithm-agnostic framework that trains expressive policies from scratch by confining policy optimization to a tractable latent space while delegating action synthesis to a conditional generative decoder. Using a two-timescale alternating schedule and anchoring decoder refinement to a fixed prior, GoRL enables stable optimization while continuously expanding expressiveness. Empirically, GoRL consistently outperforms unimodal and generative baselines across diverse continuous-control tasks. Notably, GoRL achieves returns exceeding 870 on HopperStand, more than $3\times$ the strongest baseline; on high-dimensional humanoid tasks, it further outperforms the strongest non-GoRL baseline by over an order of magnitude.} }
Endnote
%0 Conference Paper %T Generative Online Reinforcement Learning %A Chubin Zhang %A Zhenglin Wan %A Feng Chen %A Fuchao Yang %A Lang Feng %A Yaxin Zhou %A Xingrui Yu %A Yang You %A Ivor Tsang %A Bo An %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-zhang26fu %I PMLR %P 158816--158837 %U https://proceedings.mlr.press/v306/zhang26fu.html %V 306 %X Reinforcement learning (RL) faces a persistent tension: policies that are stable to optimize (e.g., Gaussians) are often too simple to represent the multimodal action distributions required for complex control. Conversely, expressive generative policies—such as diffusion and flow matching—can be difficult to optimize in online RL due to intractable likelihoods and gradients propagating through long sampling chains. We address this tension with a key structural principle: decoupling optimization from generation. Building on this, we introduce GoRL (Generative Online Reinforcement Learning), an algorithm-agnostic framework that trains expressive policies from scratch by confining policy optimization to a tractable latent space while delegating action synthesis to a conditional generative decoder. Using a two-timescale alternating schedule and anchoring decoder refinement to a fixed prior, GoRL enables stable optimization while continuously expanding expressiveness. Empirically, GoRL consistently outperforms unimodal and generative baselines across diverse continuous-control tasks. Notably, GoRL achieves returns exceeding 870 on HopperStand, more than $3\times$ the strongest baseline; on high-dimensional humanoid tasks, it further outperforms the strongest non-GoRL baseline by over an order of magnitude.
APA
Zhang, C., Wan, Z., Chen, F., Yang, F., Feng, L., Zhou, Y., Yu, X., You, Y., Tsang, I. & An, B.. (2026). Generative Online Reinforcement Learning. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:158816-158837 Available from https://proceedings.mlr.press/v306/zhang26fu.html.

Related Material