Teacher Forcing
Also known as: Teacher-Forcing
Teacher forcing is a training trick for models that generate one token at a time: at each step, feed the decoder the correct previous token from the reference, not the token the model just predicted.
Without it, an early mistake becomes the next input and can cascade. With it, every step is a clean “given the right history so far, predict the next word” problem. At inference there is no reference, so the model must consume its own outputs instead.