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

Federated learning (FL) research involves numerous algorithmic choices that are challenging to explore manually. This work introduces Auto-FL-Research (AFR), an agent-based workflow designed to automate the search for optimal FL algorithmic recipes. AFR evaluates candidate training algorithms across various healthcare and grouped-client datasets, demonstrating gains in several tasks while also identifying seed-sensitive and search-selected failure cases.