Researchers propose a new approach to federated learning called Split Federated Learning with Client-Specific Sufficiency Estimation. This method aims to improve the efficiency of distributed model training by adapting to client heterogeneity. The proposed framework, FedSGA, uses private prompt tokens and shallow sufficiency estimation to determine when a split is sufficient for each client. This approach reduces unnecessary client-side computation while improving model performance. A harmonization module also projects activations from different depths into a shared semantic space, making it easier to compare heterogeneous client interfaces. --- Why it matters: This matters because it addresses the issue of client heterogeneity in federated learning, which can lead to suboptimal results when using a single split point for all clients. By adapting to each client's specific needs, FedSGA can improve model performance and reduce computation costs. Source: https://arxiv.org/abs/2608.15639