ReinFlow was jointly proposed in September 2025 by a research team from Carnegie Mellon University, Tsinghua University, and other universities and institutions. The relevant research results were published in the paper "...".ReinFlow: Fine-tuning Flow Matching Policy with Online Reinforcement LearningIt has been selected for NeurIPS 2025.

ReinFlow is the first online reinforcement learning algorithm capable of stably fine-tuning a range of flow matching policies for a class of flow matching policies in continuous robot control. Based on rigorous RL theory, this paradigm injects learnable noise into the deterministic path of the flow policy, transforming the flow into a discrete-time Markov process, thereby enabling accurate and direct probability calculation. This transformation facilitates exploration and ensures training stability, allowing ReinFlow to stably fine-tune various flow model variants, especially with very few or even just one denoising step.