Presence penalty
A setting that nudges the model toward new words simply because it hasn’t used them yet.
Think of it like
Like a host steering the conversation to guests who haven’t spoken, regardless of how chatty the others were.
Example
Brainstorming names, a PM turns up the presence penalty so the model reaches for fresh vocabulary instead of recycling the first three ideas.
How it actually works
It applies a one-time penalty the moment a token has appeared at all — count doesn’t matter, only presence. That makes it a topic-broadener: it encourages the model to introduce new concepts rather than dwell, which is why it pairs well with brainstorming.
For product teams
Reach for it when you want breadth and novelty rather than tight, on-topic answers.
For engineers
Applies a flat logit reduction to any token that has appeared at least once in the output so far.
Related
- Frequency penalty — The count-scaled sibling.
- Sampling — Both act on the sampling 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