Muon Sublates the Edge of Stability in LLM Pretraining
Results support a split EoS picture for Muon: a stochastic loss-neutral edge survives, but it is not accompanied by a universal temporal-direction signature, and directions remain far from coherent reversal while training continues to improve.
Yan-Zhe Chen, Qifang Zhao, Xiao-Xiao Xu et al.
· 0 citations