自回归

Generating text one token at a time, each predicted from the previous ones.

LLMs are mostly autoregressive: each step predicts the next token from what has been generated, looping until an end marker. This naturally supports conditional generation and streaming, but speed is limited by serial decoding.