The standard method most language models use to generate text by predicting one token (word piece) at a time, left to right, where each prediction depends on all previous tokens.