Reproduce “Federated In-Context LLM Agent 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

Panlong Wu, Kangshuo Li, Junbao Nan, Fangxin Wang

The FICAL framework from The Chinese University of Hong Kong, Shenzhen introduces a federated learning approach for LLM agents that enables collaborative tool-use learning by exchanging natural language knowledge compendiums instead of model parameters. This method drastically reduces communication overhead by a factor of 3.33 x 10^5 while maintaining competitive tool-using accuracy, outperforming baselines under various data heterogeneity and quantization settings.

Key insights

  • Exchanging small, natural language knowledge compendiums instead of voluminous model parameters drastically reduces communication costs, achieving O(1) complexity with respect to model size.
  • Distilling abstract tool usage knowledge from private data maintains privacy while enabling collaborative learning of diverse tool-using skills.
  • A RAG-based module effectively manages extensive global knowledge compendiums, providing contextually relevant information to LLM agents and enhancing tool learning without overwhelming context windows.

Summary by alphaXiv.

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