Researchers have proposed a way to stop quantum federated learning systems from forcing every client to use an identical model structure.
Federated learning lets multiple devices train a shared model without pooling their raw data, and quantum federated learning applies the same idea to quantum neural networks. Existing approaches typically require every client to use the same circuit design, known as an ansatz, even though client data is often unevenly distributed across classes. The new framework, called PAS-QFL, splits each client's quantum neural network into a shared ansatz, picked using a cross-client stability check and aggregated centrally as usual, and a private ansatz that functions as a personalized decision layer, selected locally with a metric called Macro-F1. Clients train both pieces on their own data but upload only the shared parameters, keeping federated aggregation intact while letting each device keep its own private structure.
That distinction matters because class imbalance is the norm in real-world data, not the exception, and a one-size-fits-all ansatz can produce unstable, unfair results across clients. The approach also echoes personalized-layer techniques already established in classical federated learning, now carried over to the quantum case.
The paper reports that PAS-QFL improves average Macro-F1 over fixed-ansatz baselines, but it doesn't publish the actual figures, so there's no way to judge from this alone how large that improvement is. Quantum hardware capable of running federated learning at any meaningful scale is still mostly theoretical, so treat this as an architecture proposal, not a deployable system.