Reproduce “Experience-Handoff Reinforcement Learning for Recursive Self-Improvement” 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

Doyoung Kim, Byeongguk Jeon, Kailin Zhang, Sherry Yang

ACE-RL trains a model to solve tasks and write task-local playbooks that guide later attempts. It uses an actor-reflector-curator harness and scores candidate edits by running a fresh actor attempt through the task verifier, without a learned reward model. On 57 AppWorld dev tasks, it reaches RSI@3 of 0.667 ± 0.063 versus 0.597 ± 0.066 for matched-harness actor-only GRPO.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.