Reproduce “Corex: Pushing the Boundaries of Complex Reasoning through Multi-Model Collaboration” 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

Qiushi Sun, Zhangyue Yin, Xiang Li, Zhiyong Wu, Xipeng Qiu, Lingpeng Kong

Corex developed a multi-model collaboration framework for Large Language Models, utilizing human-inspired 'Discuss,' 'Review,' and 'Retrieve' strategies. This approach enhances complex reasoning across mathematical, commonsense, symbolic, and semi-structured tasks, often outperforming existing methods while significantly reducing computational costs.

Key insights

  • Multi-model collaboration via human-inspired strategies (Discuss, Review, Retrieve) effectively overcomes limitations of single-model reasoning, such as cumulative errors and buggy code.
  • These collaborative modes are complementary and can be combined for further performance gains, particularly in challenging numerical tasks.
  • The framework offers substantial cost-effectiveness, achieving comparable or superior performance with significantly fewer tokens compared to majority voting methods, and enables weaker open-source models to collaborate with stronger models for enhanced capabilities.

Summary by alphaXiv.

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