A sharpened version of a language model's output distribution that emphasizes high-probability tokens, used to elicit better reasoning.