Muon improves large-scale training by applying a spectral-norm steepest-descent update to matrix parameters, but practical models also contain parameter bloc…
机构:上交
来源:arXiv 2610.01395 | AI4Papers 论文推荐平台