Drift, Then Spread: Learning One-Step Generative Policies for Online Reinforcement Learning

Name: Drift, Then Spread: Learning One-Step Generative Policies for Online Reinforcement Learning
Uploaded: 2026-10-18T19:44:24.125652+00:00
Duration: 971 s
Description: DRIFTSPREAD proposes a one-step generative policy for online reinforcement learning by combining critic-guided Drift and entropy-induced Spread updates derived from a Maximum Entropy objective. This method achieves superior performance over diffusion and flow baselines on MuJoCo, offering higher return and lower training costs while maintaining fast inference.

DRIFTSPREAD proposes a one-step generative policy for online reinforcement learning by combining critic-guided Drift and entropy-induced Spread updates derived from a Maximum Entropy objective. This method achieves superior performance over diffusion and flow baselines on MuJoCo, offering higher return and lower training costs while maintaining fast inference.

Open the full episode