Softmax
The math step that turns raw scores into clean percentages that add up to 100% — how the model commits to probabilities.
Think of it like
Converting a pile of raw vote counts into a pie chart where the slices sum to a whole.
Example
It takes logits like 8, 2, 1 and turns them into something like 95%, 4%, 1%.
How it actually works
Softmax exponentiates each score and divides by the total, so bigger scores dominate and everything lands between 0 and 1. Temperature is literally a divisor applied before this — high temperature flattens the resulting percentages, low temperature sharpens them toward the top pick.
For product teams
Rarely surfaced to users; it’s the plumbing behind “how confident” a model looks.
For engineers
Normalizes logits into a probability distribution; temperature scales the input logits.
Related
- Logits — Consumes the raw logits.
- Temperature — Temperature works by stretching or squeezing this step.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome