Reproduce “RSIPO: Recursive Self-Improving Policy Optimization for Agentic Reinforcement Learning via Skill-Policy Co-Evolution” 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
Fan Zhang, Hao Wu, Jinyang Wu, Hongyu Qu, Zheng Lian, Xiao Luo, Yefeng Zheng, Bo Zhang and 2 more
RSIPO trains agents to write useful skills and calibrates trust in a skill-conditioned teacher. It splits task rollouts into evidence and control groups, replays rewritten skills from the same initial state, and uses replay gains to improve skill writing and adjust distillation. On ALFWorld, Qwen2.5-7B-Instruct achieves 98.6% average success, versus 96.1% for SEED; replay adds computation and requires resettable environments.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.