Reproduce “Lean4AgenticRL: Advancing Agentic RL with Verifiable Multi-level Process Rewards” 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

Ruida Wang, Gerald Friedland, Tong Zhang, Chris Kong

Lean4AgenticRL trains agents with task-specific workflows and rewards for final outcomes, workflow stages, and tool calls. A planning agent explores and plans each task; Lean checks some properties, while language-model judges assess others. On TerminalBench-2.1, its 27B model scores 77.3%, versus 69.6% for Qwen-3.8-27B under mean@5 evaluation; this comparison does not establish which component caused the difference.

Summary by alphaXiv.

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