Greedy

Always taking the single most-likely token (argmax).

Greedy is the simplest possible decoding rule: look at the probability the model assigns to every vocabulary token and pick the one with the highest score. No dice are rolled — argmax of the logits, every step.

That makes it perfectly deterministic: the same prompt always produces the same continuation. Because a greedy run has exactly one correct output, it doubles as a reference baseline — a way to check that an optimisation (such as caching past computation) hasn't quietly changed what the model would say.

The catch is that "most likely right now" isn't always "best overall." Greedy commits to each local winner and can never back out, so it tends toward repetitive, safe text and can miss a stronger phrasing that started with a slightly lower-probability token. Temperature, top-k and top-p all exist to loosen that grip in controlled ways.

Related: From logits to probabilities · Temperature · Logits