Auto-FL-Research: Agentic search for Federated Learning

Auto-FL-Research, an agentic workflow for algorithmic search in federated learning, was submitted to arXiv on 1 Jul 2026 as arXiv:2607.01366. The paper by Holger R. Roth, Ziyue Xu, Chester Chen, Daguang Xu, Peter Cnudde and Andrew Feng presents a "constrained coding-agent workflow" that proposes and implements candidate FL training algorithms and records their scores, runtime and artifacts.

What is Auto-FL-Research?

Auto-FL-Research (AFR) is a constrained coding-agent workflow that searches the space of federated learning algorithmic recipes, including server aggregation rules, client update schedules, local objectives and registered model variants. Task profiles fix the mutation surface, compute budget, communication contract and final model evaluation while campaigns log candidate scores, runtime, edited files, artifacts and failure status.

AFR lets agents propose and implement candidate training algorithms within those task-profile constraints. The workflow is designed to separate changes that alter the FL training or evaluation path from fixed-surface tuning effects and from single-run artifacts that appear only under search selection.

How did AFR perform in its evaluations?

AFR was evaluated on five healthcare cross-silo FLamby tasks and on grouped-client profiles for the five fixed LEAF datasets plus the LEAF synthetic task, using five-seed repeat evaluations. The authors report gains on four FLamby tasks and on five of six LEAF profiles, while also exposing seed-sensitive and search-selected failure cases.

The paper contrasts same-budget controls with agentic search outcomes: several observed gains correspond to changes in FL recipes, some improvements are recoverable by fixed-surface scalar controls, and others fail under repeat or held-out evaluation. The submission notes the experimental package spans 8 pages, with 5 figures and 6 tables documenting these mixed outcomes.

Why it matters

AFR surfaces which improvements come from repeatable algorithmic changes and which arise from tuning or single-run artifacts. That distinction matters because federated learning workflows are sensitive to many small algorithmic choices: optimizer variants, server aggregation rules, local training schedules, normalization, regularization and model architecture. By recording edited files, failure status and repeat seeds, AFR aims to make automated search results more interpretable and more comparable across experiments.

The paper's mixed outcomes emphasize that agent-generated candidates require careful repeated and held-out evaluation to confirm real, generalizable gains rather than search-selected anomalies.

What to watch

Look for follow-up work that extends AFR's task profiles or increases repeat and held-out evaluations to confirm which agent-discovered recipes generalize. The paper highlights seed-sensitive and search-selected failure cases as concrete signals: demonstrating consistent, repeatable gains under held-out evaluation would validate agentic search in FL.

Additional details: the arXiv record lists the DOI via DataCite as pending registration and gives the identifier arXiv:2607.01366. The author list is Holger R. Roth, Ziyue Xu, Chester Chen, Daguang Xu, Peter Cnudde and Andrew Feng.