arxivcs.LGmath.OC2026-07-10
Understanding Schedule-Free Methods in Nonconvex Optimization: Rate Guarantees and Escaping Saddles
Schedule-Free methods have attracted growing interest for alleviating the burden of designing and tuning a learning rate scheduler, while matching and sometimes even outperforming optimizers with tuned schedulers. Despite their strong empirical results, their convergence theory i…