Concept

Autoregressive

Generative model that produces output one token at a time, each conditioned on prior tokens.

Definition

Autoregressive models factorize a sequence's probability as a product of conditionals: each new token is sampled given all tokens generated so far. GPT, Claude, Llama and most chat models are autoregressive decoders. This makes generation inherently sequential and explains why throughput is measured in tokens per second.

Common use cases

  • Text generation
  • Code completion
  • Story writing

Related models

gpt
claude
llama

Related terms

    Autoregressive — AI Glossary | Railwail