Improving Factuality and Reasoning in Language Models Through Multiagent Debate

Something like this is what I've been waiting to read about. The agents can debate and that way approach the truth. Will it really work?