Agentic System Automates Federated Learning Algorithm Research

Federated learning (FL) research involves navigating a multitude of algorithmic decisions, from optimizer variants and server aggregation rules to local training schedules and model architectures. Manually exploring and fairly comparing these choices is resource-intensive and complex, often leading to suboptimal or inconsistent results. Auto-FL-Research (AFR) introduces an agentic workflow to automate this process. It allows coding agents to propose and implement various FL training algorithms within defined constraints, such as compute budgets and communication contracts. Each research campaign meticulously records candidate scores, runtime, code changes, and outcomes. Evaluations on healthcare and LEAF datasets demonstrate that AFR can achieve performance gains on several tasks by identifying effective FL-recipe changes. However, the research also highlights mixed outcomes, revealing instances where improvements are due to fixed-surface tuning or single-run artifacts, underscoring the need for careful analysis of agent-generated results.