Understanding the Emergence of Seemingly Useless Features in Next-Token Predictors

Publication
ICLR