[colechristensen]

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.

[binary0010]

Or you get the original output result was perfect and the adversarial "rethinking" switches to an incorrect result.

[byzantinegene]

this seems to happen far more than i would like