source: arxiv artificial intelligence: auto-fl-research: agentic search for federated learning algorithms
level: research
federated learning research involves many small algorithmic decisions like optimizer choice, aggregation rules, and local training schedules. exploring these manually is slow and makes fair comparisons hard. auto-fl-research (afr) is a constrained coding-agent workflow that automates this search. agents propose and implement candidate training algorithms, including server aggregation, client update schedules, local objectives, and model variants. task profiles lock down the mutation surface, compute budget, communication rules, and final evaluation to keep experiments consistent.
each campaign records candidate scores, runtime, edited files, artifacts, and failure status. the system was tested on five healthcare cross-silo flamby tasks. by fixing the evaluation path, afr ensures that changes in one part of the pipeline do not silently alter other steps. this makes it easier to compare different algorithmic recipes fairly. the agent-based approach reduces the manual effort needed to iterate on fl designs.
the workflow is built for researchers who need to explore many fl configurations without hand-tuning every variant. it treats algorithm design as a search problem where agents operate within strict boundaries. this can surface unexpected combinations that perform well. the recorded metadata helps track what was tried and why it succeeded or failed. afr shows how coding agents can be applied to systematic experimentation in machine learning.
why it matters: it automates the tedious trial-and-error of federated learning design, helping researchers find better algorithms faster and with fairer comparisons.
source: arxiv artificial intelligence: auto-fl-research: agentic search for federated learning algorithms