Temperature (LLM non-determinism)
A setting that controls how random a model’s word choices are; even at zero, answers can still differ between runs.
Low temperature makes output more focused and repeatable, high makes it more varied. Batching and hardware differences on the provider’s side mean the same request can still return different text, and some reasoning models fix the setting themselves. So tests on model output cannot be exact string comparisons: check properties, run each case several times, and validate structure with schemas.
Related terms
- Evals: A repeatable test set that scores an AI feature, so changes can be measured instead of guessed.
- Prompt regression: An answer that used to be right goes wrong after a change to the prompt or the model.
- Structured outputs: The model’s answer is constrained to match a JSON schema you provide.
