Temperature scaling softens the prediction distribution by dividing by a learned temperature parameter T before the softmax, but its applicability to LLMs is limited by the unstructured nature of the output.