Multi-model debate refers to having two or more models each answer the same question and read, rebut, and supplement each other, after which a model or a person summarizes; its theoretical basis is that different models differ in training data, alignment strategies, and reasoning styles, which can expose the blind spots of a single model talking to itself.