Reproduce “DeFlow: Decoupling Behavior-Prior Modeling and Value Maximization for Offline Policy Extraction” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhancun Mu, Chi Zhang
DeFlow improves offline reinforcement learning by refining actions from a multi-step flow prior. A small network corrects each proposed action using its state and value estimates, while stop-gradient prevents critic updates from changing the prior. In OGBench cube-double, DeFlow scores 40 ± 5, versus 29 ± 2 for FQL and 25 ± 3 for SORL, with results selective across categories.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.