Reproduce “Hindsight Credit Assignment over Trees for Agentic RL” 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
Penghong Zhao, Yihao Chen, Maoheng Yuan, Yunjian Zhang, Abudukelimu Wuerkaixi, Haoran Xu, Man Lv, Yaolong Wang and 3 more
TriGRPO trains long-horizon agents by branching from an active trajectory when it reaches a state matched to one in a successful run, inserting that run’s action, and letting the branch continue autonomously. On ALFWorld with Qwen2.5-3B-Instruct, it reaches 93.8% success using 0.338B rollout tokens, versus GiGPO’s 87.6% and 0.360B. The method requires exact state matching and resettability.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.