greedy decoding

Always choosing the single highest-probability token at every generation step.

Greedy decoding picks $\arg\max_w P(w \mid w_{<i})$ at every generation step -- the single highest-probability next token, with no randomness. It is the limit of temperature sampling as $T \to 0$ , and while it is fully deterministic, it can produce repetitive or globally worse text than a strategy that samples with some randomness, since it never reconsiders an early greedy choice that turns out to lead somewhere worse.

Examples

Greedy decoding on "The sky is" always outputs whichever single word has the highest probability, such as "blue," producing the identical continuation on every run.