AI RESEARCH
TrasMuon: Trust-Region Adaptive Scaling for Orthogonalized Momentum Optimizers
arXiv CS.LG
•
ArXi:2602.13498v2 Announce Type: replace Muon-style optimizers leverage Newton-Schulz (NS) iterations to orthogonalize updates, yielding update geometries that often outperform Adam-series methods. However, this orthogonalization discards magnitude information, rendering