Definition
Plain language
Picking the single option a model rates highest.
As stated in the literature
The decoding operation that selects the token with maximum probability from a model's output distribution, equivalent to greedy decoding at each step.
Why it matters: It's the simplest possible decoding strategy and the reference point against which sampling, beam search, and other methods are compared.
For example, with greedy decoding the model takes argmax over its output distribution and emits whichever single token had the highest probability.
Heard on the show
“The argmax changes after.”Episode 032 — A Sticky-Note for Every Layer: Letting Transformers Remember What They Were Just Thinking