No, they just need to be trained to have adversarial self review "thinking" processes.

You ask an LLM "What's wrong with your answer?" and you get pretty good results.