A blunt setting that discourages the model from saying the same thing again. It fixes visible looping and, pushed too hard, stops the model repeating words it genuinely needs, such as a name or a technical term used throughout a document.
Technical view
Divides or subtracts from the logits of previously generated tokens, in variants including presence and frequency penalties and n-gram blocking. It treats all repetition as undesirable, which damages structured output, code and any text with legitimate recurring vocabulary. Repetition is usually a symptom of low-temperature decoding or a weak model, so adjusting sampling is generally the better fix.
Related
Part of AI Terminology Index, section: Language Models Specifically.