Autoregressive generation
Producing text one token at a time, feeding each new token back in as input for the next.
Also: autoregressive model
An autoregressive model writes by repeated prediction: it reads the prompt, picks a next token, appends it, and runs again on the longer text. Each step depends on every step before it, which is why answers appear word by word and why generation cannot simply be spread across many chips at once. It also means an early mistake carries forward. Nearly all large language models generate this way. Diffusion language models are one attempt to write many tokens in parallel instead.