A technique where a model generates multiple responses and uses agreement among them to improve answer reliability.