Autoregressive
Generating a sequence one token at a time, each conditioned on all the tokens produced so far.
Autoregressive generation produces a sequence one element at a time, feeding each output back in as input for the next — the model predicts token t+1 from tokens 1..t. It is how GPT-style LLMs generate text.
Worked example: to write “the cat sat”, the model predicts “the”, then “cat” given “the”, then “sat” given “the cat” — which is why generation is inherently sequential and the decode phase is latency-bound. Gotcha: because each token depends on the previous, you cannot fully parallelize generation the way you can the prefill of the prompt — the bottleneck techniques like speculative decoding attack; diffusion-LLMs are an alternative that generate non-sequentially.