Greedy decoding
Always taking the highest-probability token. Repeatable, and prone to loops and flat prose on anything open-ended.
What people get wrong
It is not the same as reproducible. Batching, hardware and kernel choices change floating-point results enough to flip a close call, so the same prompt can still yield different text.
A definition is not an understanding. This is taught properly in How a model picks the next token, part of Sampling & decoding.
Learn this properlyTerms from the same topic