Researchers propose Auto-FL-Research, an agentic workflow for automating federated learning algorithm design

Federated learning (FL) research relies on numerous small but consequential algorithmic choices—such as optimizer variants, server aggregation rules, local training schedules, normalization, regularization, and model architecture—that are costly to explore manually and difficult to compare fairly. These choices can alter the FL training or evaluation path, complicating fair comparisons across candidate changes.

The authors introduce Auto-FL-Research (AFR), a constrained coding-agent workflow designed to automate the search for FL algorithmic recipes. Within AFR, agents may propose and implement candidate training algorithms, including server aggregation rules, client update schedules, local objectives, and registered model variants. Task profiles in AFR fix the mutation surface, compute budget, communication contract, and final model evaluation to ensure consistent comparisons.

Each AFR campaign records candidate scores, runtime, edited files, artifacts, and failure status, enabling systematic tracking of algorithmic variants and their outcomes. The evaluation covers five healthcare cross-silo FLamby tasks and grouped-client profiles for the five fixed LEAF datasets plus the LEAF synthetic task.

Five-seed repeat evaluations report gains on four FLamby tasks and five of six LEAF profiles, but also expose seed-sensitive and search-selected failure cases. Same-budget controls indicate that some gains correspond to genuine FL-recipe changes, while others are recovered by fixed-surface scalar controls or fail under repeat or held-out evaluation.

The mixed outcomes are presented as part of the contribution, demonstrating how agent-generated candidates can be categorized into repeated FL mechanisms, fixed-surface tuning effects, and selected single-run artifacts. This categorization underscores the importance of rigorous controls and repeated evaluation in automated algorithm design.