I mean, those approaches do improve results. Some lower complexity translation tasks will reliably return a factual response. These are statistical models so sampling and dropping odd-man-out responses can get to 100% factual responses for a growing category of prompts.