Reproduce “Asymmetric On-Policy Distillation Learns In-Context Exploration” 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
Sriyash Poddar, Yanda Bao, Jacob Krantz, Matthew Chang, Xavier Puig, Roozbeh Mottaghi, Natasha Jaques, Abhishek Gupta
The authors introduce ASTEROID, which trains a partially observed agent to search by pairing histories from its own rollouts with actions from a privileged expert. This teaches the student to use observations and feedback rather than only imitate successful demonstrations. In simulated manipulation, it learns successful cube-pick or peg-insertion policies with about 100× fewer environment steps than RL baselines, a measure of training interactions rather than elapsed time.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.