Definition
Temperature scales the probability distribution before sampling. Low values (near 0) make outputs more deterministic; higher values increase diversity and surprise.
It does not make the model "smarter", it changes how boldly it samples from what it already believes.
In simple terms
Low temperature is carefully picking the safest menu item. High temperature is trying the chef's weird specials, more variety, more risk.
Where you see it
- Customer support bots use low temperature for consistent answers.
- Creative writing uses higher temperature for varied prose.
How it works
1.Model scores tokens
Logits for each vocabulary item.
2.Divide by temperature
Higher T flattens the distribution.
3.Sample
Draw the next token from the adjusted probabilities.
Why it matters
- Temperature is one of the first knobs product teams tune for reliability vs creativity.
Often confused
Temperature 0 guarantees factual correctness.
It reduces randomness but the model can still confidently produce false content.