Can we weight the sampling based on a prediction of how long an answer will be? Like with another model that just says "Oh boy that word sounds like the beginn

Can we weight the sampling based on a prediction of how long an answer will be? Like with another model that just says "Oh boy that word sounds like the beginning of an essay of nonsense, I don't think that's what I want to say"?