ModelsSelected12:25Chinese contentSG-TULA:非凸非光滑采样的驯化次梯度算法新采样算法SG-TULA能处理非光滑非凸问题,在GPT-2预训练上干赢了AdamW和Muon,还带理论保证。#SG-TULA#Langevin算法#GPT-2#采样aarXiv cs.LG@Iosif Lytras et al. (3)SourceRead laterReadWorth followingUsefulFollow SG-TULA