Reproduce “SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning” 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
Peng Xia, Jianwen Chen, Hanyang Wang, Jiaqi Liu, Kaide Zeng, Yu Wang, Siwei Han, Yiyang Zhou and 5 more
SkillRL distills successful and failed interaction trajectories into a hierarchical bank of general and task-specific skills, then updates that bank from validation failures as its policy learns. A cold-start supervised fine-tuning stage teaches the policy to use the skills. On ALFWorld, it reports 89.9% success versus GRPO’s 77.6%; the headline setup uses OpenAI o3 as teacher.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.