Temperature

A sampling parameter that changes how concentrated a model's next-token distribution becomes. Lower temperatures favor tokens that are already more probable and make output more predictable; higher temperatures spread probability across alternatives and increase variation.

Temperature controls diversity, not truth, knowledge, or safety. Lowering it can make a wrong answer more consistent; raising it may suit ideation and be unsuitable for structured extraction. The setting should be measured on the task alongside format, context, model, and other decoding controls.