Most federated learning research centers on new optimizers or aggregation rules. The authors of a recent arXiv preprint instead ask whether a small per-client measurement budget can support decisions about how much personalization is worth deploying. They frame personalization depth as a cost-aware choice, with a utility that trades Macro-F1 against deployment cost.
Using shallow probes—brief, low-cost client-specific measurements—the approach allocates personalization depth across clients. The result is near-oracle performance: decisions close to those made with perfect knowledge, despite the limited budget. The study does not introduce a new federated optimizer; it isolates the allocation problem itself.
Because only the abstract is available here, details of the probing mechanism and evaluation are not yet accessible. The central claim is that shallow probes are sufficient for near-oracle depth allocation under the stated utility.