Reproduce “R$^2$Orch: Recursive Self-Improvement for Agent Orchestration via Self-Play within a Skill-Internalization Loop” 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
Mingda Zhang, Yun Huang, Qiang Huang, Yanjin Li, Rui Qian, Qichao Zhao, Tiesunlong Shen, Na Zhao and 2 more
R²Orch trains a Director to build graphs of collaborating agents, whose frozen Workers execute. It trains on tasks scored with a candidate skill hidden, while randomized with-skill and without-skill rollouts estimate whether that skill helps. Across 12 benchmarks, it leads tested baselines at zero skills; on AIME 2026, it scores 81.33%, versus 70.67% under current-scaffold learnability. Initial skills were customized by hand per dataset.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.