The same four logits three times: as signed bars, as exponentials that make every one positive and stretch the gaps, and as one stacked bar that adds to 100% — the negative class still gets a slice. Evidence: how the top probability and the entropy move with temperature; the ladder is how peaked the same logits become from icy to uniform.
The loudest cheer gets the biggest slice — but everyone gets some
Four friends shout how much they want cake. You could give the whole cake to the loudest, but that's unfair to a friend who was almost as loud. So you cut slices: the loudest gets the biggest, the quietest the smallest, and every slice added together is exactly one cake. Even a friend who whispered gets a sliver.
Softmax turns 'who shouted loudest' into 'how big is each slice' — and the slices always make one whole cake.
pᵢ = exp(zᵢ) / Σⱼexp(zⱼ)
Icy
T = 0.25
98%
Cool
T = 0.5
84%
Neutral
T = 1
61%
Warm
T = 2
43%
Uniform
T → ∞
25%
The same four logits, +2 / +1 / +0.5 / −1, at five temperatures — linear percent for the top class