An overuse penalty is a correction applied during decoding to stop a language model from repeating the same words and phrasings. Once the model picks a token, the probability of that token at the next step is reduced. The reduction grows with each repetition.
There are two common forms. A presence penalty subtracts a fixed amount if the token has appeared at all. A frequency penalty subtracts in proportion to how many times it has appeared, so it punishes persistent repetition harder. Both are exposed as numeric API parameters, usually somewhere between zero and two.
The penalty exists because generation is stepwise. The model reads its own output back as input, so once it falls into a pattern it can keep reinforcing that pattern. Without a penalty, long outputs visibly loop on the same sentence.
A worked example: in a flow writing descriptions for 20 products, with the penalty off most of them open with "This product offers users". Set a small frequency penalty and the opening lines start to vary.
Pushing the value too high backfires. As the penalty rises the model starts avoiding words the text genuinely needs. In copy where a brand name should recur, that produces odd sentences where the name never appears a second time.