Reproduce “Model Swarms: Collaborative Search to Adapt LLM Experts via Swarm Intelligence” 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

Shangbin Feng, Zifeng Wang, Yike Wang, Sayna Ebrahimi, Hamid Palangi, Lesly Miculicich, Achin Kulshrestha, Nathalie Rauschmayr and 4 more

Model Swarms introduces a collaborative search algorithm leveraging swarm intelligence to adapt Large Language Model experts. The method enables tuning-free adaptation of LLMs to various tasks and preferences, achieving state-of-the-art results even with limited data (e.g., 200 examples) and activating implicit expertise in weaker models.

Key insights

  • Collaborative search can discover new capabilities, solving problems that none of the initial experts could, evidenced by an average "Correctness Emergence" of 44.8%.
  • Initially "weak" LLM experts can contribute to and become the best-performing adapted models, with 56.9% of the final best models originating from the bottom half of the initial pool.
  • The diversity of the initial expert pool is crucial, leading to a 35.3% performance improvement when moving from the least to the most diverse sets.

Summary by alphaXiv.

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