That’s already what we currently have. Look for “temperature” in the context of LLMs. They are already tuned to have variance.