Large language model agents can evolve reusable natural‑language skills from execution experience. In real deployments this experience is scattered across data owners who cannot pool raw data, motivating federated skill evolution.
A naïve design lets each client evolve the entire skill document and upload it, which suffers from three drawbacks: the document may still contain client‑specific information; transmitting an almost unchanged document wastes bandwidth; and a single client’s contribution is hard to attribute on the server side.
FedSkillGit replaces full‑document updates with commits. Each client submits a bounded set of patches targeting named rules together with a summary of its local experience for the round. The pair passes a two‑stage privacy filter before leaving the client.
The server deduplicates all patches and experiences, updates a global experience document (read only by optimizers), and merges the surviving patches into a candidate skill. The candidate is distributed to clients, which evaluate it on their own validation sets and report scores. The candidate is published only when the reported scores improve over the current version.
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.
Review: By using incremental patches and experience deduplication, the framework dramatically reduces communication overhead and provides traceable contribution attribution, offering a practical path for real‑world federated deployments.