Personalized Framework Reduces Negative Transfer in Federated Learning

A new decentralized federated learning framework utilizes policy-driven multi-path routing and a communication-while-aggregation protocol to prevent overconsensus bias in heterogeneous multi-task networks. The method uses a column-stochastic consensus matrix to block mismatched parameter updates and preserve local task fidelity.