Reproduce “SAPD: Learning from Agent Experience without Environment Interaction” 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
Juncheng Diao, Haibin Wang, Wei Shi, Yifan Ding, Qingbin Li, Rongxiang Weng, Xunliang Cai
SAPD reuses recorded interaction histories to train agents without executing new actions. For each decision context, the student generates candidate responses and a fixed teacher scores them; step scores are centered against candidates from the same prefix. In WebShop, SAPD took 10.7 seconds per step versus 116.7 for online OPD, a reported 10.9× speedup, not evidence of fewer tokens or attempts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.