Link. “each token generated by a language model is by definition predicted by the preceding tokens”