Less Is More in Federated Continual Learning: RieSelect for Conflict-Aware Layer Selection in LLMs

Federated continual learning (FCL) of large language models on edge devices is constrained by a communication--stability--plasticity trilemma. We reveal a less-is-more phenomenon: beyond a moderate layer upload ratio, stability loss offsets saturated plasticity gains, so overall continual performance no longer improves. Moreover, layer-wise conflict is heavy-tailed and concentrates in a few layers; denser uplink increasingly includes these layers, which disproportionately drives forgetting and motivates selective sparse communication. Therefore, we introduce RieSelect, which treats stability as staying within a Fisher-metric safe basin around historical solutions. Under this safe-basin constraint, we derive a layer-wise conflict score and a closed-form certified safe step size for finite local updates, and formulate selective uplink as a knapsack-based utility--risk selection, balancing plasticity gains against stability risks. Extensive experiments show that, under a per-round uplink budget, RieSelect achieves the best performance across task orders. Beyond this matched-budget setting, under standard communication protocols, RieSelect improves average accuracy by 18.99–28.14 points while reducing total uplink by 53–115$\times$.

Large AI models often need to keep learning from new data on many user devices, such as phones or edge machines, without collecting private data in one place. A common approach is to let each device send its model changes to a central server, but sending more changes is not always better: it can waste communication resources and may even cause the model to forget useful abilities learned earlier. In this work, we find that this problem is often caused by a small number of model parts whose changes conflict with past knowledge. We propose RieSelect, a method that lets each device send only the changes that are both useful for learning new tasks and unlikely to damage old knowledge. It also carefully limits risky changes before they are shared. Experiments show that this selective strategy helps large models learn new tasks while better retaining previous skills, using much less communication than existing methods. This makes continual improvement of large AI models more practical for bandwidth-limited and privacy-sensitive settings.