TLDR; Train an "energy" model that checks if the output is correct (rather than just outputting something), and gradient descent to find good outputs. Using transformers.