Yong-Tao Wu

From Spectral to Frobenius: When Muon Loses Its Spectral Implicit Bias

It is proved that Muon can recover its spectral max-margin bias with Newton-Schulz iterations growing logarithmically over training, and disentangle the roles of Newton-Schulz iteration count, the stabilization constant, and the Newton-Schulz polynomial, revealing distinct asymptotic behaviors, including directions tha...