protologue

Greedy Decoding

Also called argmax decoding, temperature 0.

Greedy decoding generates text by always choosing the single most probable next token, which makes output deterministic for a given prompt and model.

Description

It is the limit of sampling as temperature approaches zero and is often used when consistency matters, as in extraction or classification. For open-ended text it tends to produce bland or repetitive output, and for reasoning tasks sampling several answers and voting, as in self-consistency, usually beats a single greedy answer.

Sources

  1. Holtzman et al. (2019). The Curious Case of Neural Text Degeneration.

Cite this entry

Protologue. (2026). Greedy Decoding. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.4.0, PTL-0109). https://protologue.com/t/greedy-decoding/

BibTeX
@misc{protologue_greedy_decoding,
  title = {Greedy Decoding},
  author = {{Protologue}},
  year = {2026},
  howpublished = {Protologue: A Taxonomy of Prompting and LLM Techniques, v1.4.0},
  note = {Entry PTL-0109},
  url = {https://protologue.com/t/greedy-decoding/}
}

Markdown JSON