Researchers from Wenhao Yan et al. explore a new approach to federated learning on non-IID data. They investigate how regularization survives averaging in this context and propose an update masking method that leverages the benefits of noise added to parameter updates. The study shows that when each client draws its own mask, federated averaging weakens the charge for moving in sharp directions by a factor related to the cohort size. However, using the same mask across all clients can mitigate this effect. --- Why it matters: This research matters because it sheds light on the relationship between regularization and averaging in federated learning. Understanding how to effectively preserve regularization in distributed training settings is crucial for developing robust and generalizable models. Source: https://arxiv.org/abs/2608.23286