The function that turns raw scores into a probability distribution summing to one.
Broad technical consensus on what this term means. Disagreements that invoke it are about the world, not the word.
Softmax exponentiates every logit and divides by the total, producing values between zero and one that sum to one. It appears twice in a transformer: over the vocabulary at the output head, and inside attention to turn comparison scores into mixing weights.
Because it exponentiates, small differences in logits become large differences in probability. That sensitivity is why a model can be overwhelmingly confident about one token and why temperature, which scales logits before this step, has such a pronounced effect.