Reproduce “FederatedSkill: Federated Learning for Agentic Skill Evolution” 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

Jingbo Yang, Guanyu Yao, Yang Zhang, Ramana Rao Kompella, Gaowen Liu, Shiyu Chang

A framework named FederatedSkill enables privacy-preserving and personalized collaborative evolution of LLM agent skills for heterogeneous clients. The approach leverages semantic skill patches for communication and a server-side evolution agent for tailored updates, achieving up to a 44.4 percentage point increase in task success rate and a 37.5% reduction in API costs while demonstrating low sensitive data leakage.

Key insights

  • Communicating 'skill diffs' (semantic patches) is an effective and privacy-preserving unit for federated learning in agent skill evolution, abstracting away sensitive raw execution data.
  • Personalization is critical for collaborative skill evolution among heterogeneous LLM agents; a universal skill library often performs suboptimally compared to tailored updates.
  • A server-side agent with an explicit model of client capabilities (via capability matrix and memory) can intelligently curate and transfer relevant skills, preventing library bloat and promoting efficient learning.

Summary by alphaXiv.

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