When a language model attends to the same previously-generated token multiple times during inference.