Reproduce “Beyond Division of Labor: How Roles Shape Feedback and Recovery in Multi-Agent Systems” 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

Chih-Hsuan Yang, Xinyang Li, Rajeev Thakur, Feng Yan

The authors test how role instructions affect a three-agent Planner–Executor–Reviewer system’s work and access to evaluator feedback. Across text-based reasoning tasks, removing roles largely erased specialization without lowering first-pass accuracy. In an exploratory gemma-4 Omni-MATH comparison, role-free teams recovered 0.609 of initially wrong answers, versus 0.222 with roles. This reflects oracle-assisted recovery, not autonomous self-correction.

Summary by alphaXiv.

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