Reproduce “CompassPlay: Rewarding the Proposer for Where It Moves the Solver” 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

Sophia Xiao Pu, Ximeng Sun, Jiang Liu, Jialian Wu, Emad Barsoum, Zicheng Liu, William Yang Wang

CompassPlay rewards tasks by aligning the solver’s loss gradient on a verified answer with a reference gradient for a target capability. In coding self-play, mean accuracy gains were 1.50 points on coding benchmarks and 2.65 on mathematics benchmarks versus AZR. In Lean, it reached a coverage threshold with 40% fewer GPU-hours; alignment remains a proxy, not measured learning progress.

Summary by alphaXiv.

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