Novice here. I like the sound of that! I don't know the answer but I might experiment with it. Probably a researcher has tried it. You would need N times the

Novice here. I like the sound of that! I don't know the answer but I might experiment with it. Probably a researcher has tried it. You would need N times the compute per token generate of course. You could either pick to N, or sample N (with temperature adjustment to logits if needed)