Muon Optimizer Accelerates Grokking
Empirical evaluation across seven modular arithmetic tasks comparing Muon vs. AdamW and softmax variants (standard softmax, stablemax, sparsemax). Muon reduced the mean grokking epoch from 153.09 to 102.89 (t = 5.0175, p = 6.33e-08), accelerating the transition from memorization to generalization.




