A new approach to federated learning stops building one personalized model per user and instead builds one per task.
Federated learning lets multiple devices or organizations train a shared model without pooling their raw data, but merging updates from clients with wildly different data tends to drag down everyone's performance, and current fixes that personalize a model per client often struggle once that client faces a new task or mixes several tasks in local training. FedRouter's answer is to attach small adapter modules to each client and cluster them twice: locally, to match adapters to a client's own task data, and globally, to group similar adapters across different clients into task-specific models. An evaluation router then sends each test sample to whichever adapter cluster fits it best. Tested on a multitask benchmark, FedRouter beat prior personalized federated learning methods by up to 6.1% when a client's own tasks interfered with each other, and by up to 136% when clients faced tasks they had not seen before.
That generalization number is the one worth sitting with. Most personalized federated learning work optimizes for the client a model already knows, a reasonable bet until that client's usage shifts, say a company rolling out a new internal tool or a phone keyboard meeting a new language. Routing by task rather than by identity is a structurally different bet, and it matters more as federated systems move from clean benchmark splits into messier real-world data.
It is still a benchmark result from a preprint, tested on one multitask dataset, not a product running on anyone's phone. Whether task clustering holds up when the number of tasks is unknown or constantly shifting, which is the normal state of real devices, is the next thing to watch.