Reproduce “CODA: Coordinating the Cerebrum and Cerebellum for a Dual-Brain Computer Use Agent with Decoupled Reinforcement Learning” 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

Zeyi Sun, Yuhang Cao, Jianze Liang, Qiushi Sun, Ziyu Liu, Zhixiong Zhang, Yuhang Zang, Xiaoyi Dong and 3 more

CODA introduces a trainable compositional framework for autonomous GUI agents, drawing inspiration from the cerebrum and cerebellum to enable both high-level planning and precise execution. The system achieves state-of-the-art performance on specialized scientific computing tasks through decoupled reinforcement learning and a two-stage training curriculum.

Key insights

  • Decoupling planning from execution, inspired by the human brain's cerebrum-cerebellum architecture, allows for specialized optimization of each component.
  • Applying reinforcement learning exclusively to the Planner while keeping the Executor fixed is a data-efficient approach for acquiring domain-specific strategic reasoning.
  • A "specialist-to-generalist" training paradigm, where a generalist model learns from aggregated successful trajectories of specialist teachers, effectively enhances generalization and broader domain knowledge.

Summary by alphaXiv.

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