ReinFlow fine-tunes pretrained flow-matching policies with online reinforcement learning. A learned noise-injection network enables exploration and tractable policy likelihoods during training, while preserving efficient deterministic flow inference after fine-tuning.
