Autoregressive
Writing one token at a time, each new word chosen from all the words so far — including the ones it just wrote.
Think of it like
Predictive text that keeps going: it guesses the next word, commits, then guesses again from the fuller sentence.
Example
When a model streams its answer word by word on screen, you’re literally watching autoregression happen.
How it actually works
Each step feeds the model everything generated up to now and asks for the next token. That token gets appended and the loop repeats. It’s why generation is sequential and can’t fully parallelize, and why an early wrong turn can snowball.
For product teams
It’s why responses stream in and why latency scales with output length.
For engineers
Factorizing p(sequence) as a product of next-token conditionals, decoding left to right.
Related
- Next-token prediction — The mechanism behind next-token prediction.
- Tokens — Emits one token at a time.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome