Study Finds Same AI Model Debating Itself Outperforms Mixed-Model Pairs

A field test of the open-source AdversarialDebate framework produced a counterintuitive result: a homogeneous GPT-plus-GPT pairing scored higher and made more concessions per debate than heterogeneous pairings like GPT-plus-Gemini and Gemini-plus-Mistral. The GPT-plus-GPT pair achieved an average score of 0.688 with a 57% verdict rate, significantly outperforming GPT-plus-Gemini, which scored only 0.357 with a 4% verdict rate. The findings suggest that model diversity alone does not guarantee better critical reasoning in multi-agent AI debate systems. A subsequent v0.2.1 update further refined the thesis, indicating that the presence or absence of the Mistral model — not diversity versus homogeneity — was the more decisive factor. The project is publicly available on GitHub and PyPI under the name adversarial-debate.