No Train No Gain:Revisiting Efficient Training Algrthm for Transformer-BasedLM (opens in new tab)

Recent paper highlights the difficulty of creating a new optimizer as drop-in replacement. Sophia and Lion were recently proposed as superior alternatives to Adam, but appeared worse in an independent eval