I use a process called sample-and-vote when having LLMs return executable solutions. I imagine we will see this kind of design a lot with probabilistic computing.