Reproduce “Fast Inference-Time Policy Refinement via Action-Affine World Models” 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
Run Wang, Yilin Mo, Liang Wu
The authors introduce A²WM, which refines a frozen imitation policy by predicting a nominal task trajectory and the effects of bounded action corrections over the planning horizon. It combines these predictions in a quadratic program and replans from observations. Across four simulated tasks, mean success was 59.51%, versus 27.18% for the frozen baseline; results are limited to simulation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.