The technique of adjusting randomness in a model's token selection during generation, controlled by the temperature parameter.
Lower temperature makes the model consistently pick the most likely next token; higher temperature introduces more variety and creativity at the cost of predictability.
