Glossary

Temperature (sampling) — Glossary

A decoding parameter that controls how random or deterministic next-token sampling is.

Also known as: temperature

Reviewed:

Full definition

Temperature rescales logits before softmax when sampling tokens. Lower values concentrate probability on the top tokens (more deterministic); higher values flatten the distribution (more diverse).

Temperature is not a truthfulness control. Factual tasks usually prefer low temperature plus grounding tools. Creative brainstorming may use higher values with later filtering.

Combine temperature with top-p/top-k settings carefully and evaluate on your task suite.

Related encyclopedia topic

Temperature — Temperature is a decoding parameter that scales next-token probabilities, influencing how deterministic or varied a language model’s outputs are.

← Back to glossary · Learning hub