The research is surprisingly strong in favor of this.

Du et al. (ICML 2024) showed that when you get multiple LLMs to debate each other, not just share answers, but actively challenge each other's reasoning, then factual accuracy goes up and hallucinations go down. The wildest finding was that in some cases, every model started with the wrong answer but converged on the correct one through debate. The process itself generated correctness that no individual model had.

This is why xAI shipped a 4-agent debate inside Grok 4.20 three days ago. One of the leading AI labs looked at every way to improve output quality and landed on structured debate.