Reproduce “ROAM-GRPO: Enforcing Reasoning Diversity in GRPO via Routed Off-mode Approach Mining” 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

Jeff Duan, Matt Thomson

ROAM-GRPO trains models to practice alternative mathematical approaches on problems they already solve. It routes solved prompts to an approach-mining track, where valid alternatives are judged for soundness and distinctness, then distilled alongside a correct anchor. In one Qwen3.5-2B LoRA experiment, delayed ROAM-GRPO reached peak pass@16 of 46.0% versus GRPO’s 41.3% on held-out Omni-MATH; training was unstable, and pass@1 was lower.

Summary by alphaXiv.

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