DynMuon: A Dynamic Spectral Shaping View of Muon

In three linesDynMuon extends Muon by replacing update M with U·Σ^p·V† using dynamic parameter p. Theory shows positive p accelerates signal contraction early by emphasizing high-curvature directions, while mildly negative p reallocates update strength to low-curvature directions late in training. DynMuon reduces steps to target loss by 10.6-26.5% versus Muon.

Summary generated by Claude — human-verified