FedSkillGit: Federated Skill Evolution via Version Control

Key points

• Large language model agents can improve their capabilities by evolving reusable natural-language skills from execution experience.1
• In real-world deployments, this experience is distributed across data owners who cannot pool their raw data, which motivates federated skill evolution.1
• The natural first design lets each client evolve the whole skill document and upload the result for server-side aggregation.1
• Experiments on three datasets and three language model families show that FedSkillGit outperforms state-of-the-art baselines by up to 22% and approaches centralized evolution without sharing raw client data.1

Sources (1)

• [1]FedSkillGit: Federated Skill Evolution via Version Control
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 15, 06:29 AM

  We propose FedSkillGit, a federated skill evolution framework that replaces full-document updates with commits.

  Large language model agents can improve their capabilities by evolving reusable natural-language skills from execution experience.

Extractive summary: sentences quoted from the sources.