“Best-of-N sampling, where a model generates multiple responses and a reward model selects the best one, is a common technique used by OpenAI to improve output quality at the cost of increased inference compute.”