Dr. Liang turned the temperature dial slowly toward zero and watched the display change. At high τ, the softmax probabilities were nearly uniform — every row's values barely distinguishable. As τ dropped, the diagonal elements grew brighter, pulling probability mass toward themselves while the margins dimmed. She knew what was coming: as τ→0, each row would collapse entirely onto its diagonal element, converging to a one-hot encoding of the argmax. The pre-softmax logits had their maximum along the main diagonal. She reached for the dial again.