ADAPT-FED: Adaptive Federated Optimization with Learning Stability

ICLR 2026 Conference Submission20674 Authors

Keywords: Federated Learning, Optimization

TL;DR: Adaptive Federated Learning with Learning Stability

Abstract: Federated Learning (FL) frequently exhibits poor generalization due to unstable training across heterogeneous clients. Although training instability can accelerate learning, it often compromises generalization, resulting in a fundamental tension within FL. This work introduces \sysname, a framework that adaptively regulates training dynamics to leverage the advantages of instability while mitigating its adverse effects. As a result, \sysname enables more stable and consistent learning in privacy-constrained environments. Experimental results on standard benchmarks demonstrate that \sysname enhances generalization and convergence relative to state-of-the-art FL optimization algorithms.

Primary Area: optimization

Submission Number: 20674