Stale Information Matters: Efficiently Tackling Data Heterogeneity in Asynchronous Federated Learning with Model Calibration

Federated averaging (FedAvg) and its variants are widely used in federated learning (FL), but their synchronous aggregation suffers from the straggler problem and thus poor training efficiency. Asynchronous FL (AFL) mitigates these delays by removing strict synchronization, yet its performance is often degraded by client drift under non-independent and identically distributed (non-IID) data, biasing the global model toward frequently updated clients and away from rarely participating ones. We propose FedQue, a semi-asynchronous FL method in which the server maintains a model queue and reuses historical updates to calibrate the global model and correct client drift. FedQue aggregates both recent and buffered updates, so that each global step moves in a direction that better reflects the whole client population rather than a few fast devices. Our non-convex analysis shows that the proportion of stale information, captured by the queue staleness, balances local gradient variance and sensitivity to model initialization, and that FedQue attains the standard stochastic gradient descent convergence rate in the non-convex setting. Building on this theory, we design a data-driven procedure that adaptively selects queue staleness during training using online statistics of update variability and staleness, without modifying client-side code or communication patterns. Extensive experiments on multiple datasets across different non-IID levels and client scales, demonstrate that FedQue consistently improves both accuracy and time-to-target, achieving up to 2.14× higher training efficiency than state-of-the-art AFL algorithms under highly heterogeneous data, while maintaining competitive or better final test performance.