Glossary · Term

greedy decoding

← all terms

Definition

Plain language

The simplest way for a language model to generate text — at each step it just picks the single most likely next word.

As stated in the literature

A decoding strategy that selects the argmax token at each step from the model's output distribution; the standard choice for factual QA and the regime in which commitment failures are most visible.

Why it matters: It's the cheapest and most reproducible way to generate text, but it commits early and can lock in mistakes that more exploratory decoders would avoid.

For example, with greedy decoding a model asked '2 + 2 = ' will pick whatever single token has the highest probability at each step, even if the second-most-likely token would lead to a more accurate continuation.

Heard on the show

“On a small Gemma model, success goes from thirty-three percent with greedy decoding up to thirty-nine with progress-advantage selection.”
Episode 173 — The Free Step-Level Grader Hiding in Every RL Training Run

Mentioned in 4 episodes

  1. 173
    The Free Step-Level Grader Hiding in Every RL Training Run
  2. 140
    When a Reasoning Model Says "Let Me Double-Check" After It's Already Decided
  3. 070
    When Models Know the Answer But Say the Wrong Thing Anyway
  4. 026
    What RL Actually Does to Language Models, at the Token Level

Related terms