Title: FederatedSkill: Enabling Agentic Skill Evolution Through Federated Learning

Abstract:

As large language model (LLM) agents increasingly depend on skill libraries to manage complex operations, the continuous evolution of these skills has emerged as a critical mechanism for self-improvement. Nevertheless, relying on isolated, single-user task streams restricts the data diversity necessary for developing comprehensive capabilities. Although cross-user collaboration offers a solution to this data scarcity, existing methods that share raw trajectories often violate user privacy and enforce a rigid global library, which overlooks the heterogeneity among different clients.

To address these challenges, we present FederatedSkill, a privacy-centric framework designed for the collaborative evolution of agents. Rather than exchanging raw interaction histories, FederatedSkill employs semantic skill diffs—structured modifications to local libraries—as its core communication unit. On the server side, an evolution agent synthesizes these patches to dynamically map the specific capability boundaries of each client. This approach enables strictly personalized skill development, avoiding the performance compromises associated with suboptimal global averaging.

Our evaluation across 20 distinct agent task families reveals that FederatedSkill significantly outperforms self-evolving baselines. The framework achieves a success rate improvement of up to 44.4% while simultaneously reducing computational costs by 37.5%.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC