Definition
Plain language
Peeking at what's coming later before committing to a choice now.
As stated in the literature
Conditioning a decision on future candidates or states rather than only the prefix already produced; standard left-to-right decoding has none.
Why it matters: Without lookahead, a model cannot tell whether a list it is halfway through will end up complete, so it has no basis for deciding when to stop.
For example, a chess player who considers three possible replies before moving is using lookahead, while a model writing text left to right commits to each word before seeing what follows.
Heard on the show
“Listing is a different animal, because a language model writes one token at a time, conditioned only on what it has already written, with no lookahead over the candidates it hasn't sampled yet.”Episode 233 — Why a Model Can Grade an Answer But Not Write the Answer Key