Tag
Proposes Self-Generated T2T, a training method that aligns token editing training with inference by using the model's own predictions as error sources, improving accuracy on LLaDA2.1.