Auto-FL-Research: Agentic Search for Federated Learning Algorithms

AFR is an agentic search tool that autonomously discovers effective federated learning algorithms by proposing and implementing candidate training algorithm changes.

AFR uses a constrained coding-agent workflow to propose and implement candidate training algorithms, including aggregation rules and update schedules, while fixing mutation surfaces and budgets.

AFR showed gains on 4 of 5 FLamby healthcare tasks and 5 of 6 LEAF decentralized profiles, while exposing seed-sensitive failures.

Yes, AFR accelerates federated learning research for sensitive sectors like healthcare and finance, where data is distributed and privacy is critical.