Someone once posted that they got GPT to produce much higher quality output by instructing it to make an argument, critique that argument, make a new argument with the critique in mind, and so on.

Maybe stable, precise reasoning is just a result of multiple language models arguing with each other.