DynMuon dynamically schedules the spectral shaping parameter p in Muon-like optimizers from positive to negative values, yielding lower validation loss and 10.6-26.5% fewer steps than standard Muon across tested settings.
Polargrad: A class of matrix-gradient optimizers from a unifying preconditioning perspective, 2026
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
DynMuon: A Dynamic Spectral Shaping View of Muon
DynMuon dynamically schedules the spectral shaping parameter p in Muon-like optimizers from positive to negative values, yielding lower validation loss and 10.6-26.5% fewer steps than standard Muon across tested settings.