Softmax temperature
The knob inside softmax that flattens or sharpens the model’s probabilities before it picks a word.
Think of it like
Like a contrast slider on a photo — turn it up and everything blends together, down and the differences pop.
Example
Set high, the model spreads its bets and gets creative; set near zero, it collapses onto the single most likely token almost every time.
How it actually works
Temperature divides the logits before softmax: values below 1 exaggerate gaps (more deterministic), above 1 shrink them (more random). It’s the mechanism underneath the “temperature” setting you see in APIs — same idea, applied at the point where scores become probabilities.
For product teams
The dial between “safe and repetitive” and “varied and risky” outputs.
For engineers
Scales logits by 1/T before softmax; T→0 approaches argmax, T>1 raises entropy of the distribution.
Related
- Temperature — The user-facing setting it implements.
- Softmax — The function it modifies.
- Entropy — What spread it changes, measured.
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