A real data distribution has many modes, peaks of probability: dogs of every breed, faces of every age. Collapse happens when the generator covers only some of them. Goodfellow et al. (2014) spotted it from the start in GANs: if the generator is trained too much without updating the discriminator, it sends too many values of the noise z to the same image; with typographic humor, they called this "the Helvetica scenario" (2014, sec. 6).
Why it happens. The generator does not learn the distribution: it learns to fool the current discriminator. Its best immediate move is to produce the single output the judge now finds most realistic; when the judge learns to reject it, the generator jumps to another, and training can end up cycling among modes. Goodfellow (2016) notes that complete collapse is rare and partial collapse common (the same dog from different angles). Salimans and colleagues (2016) identified a concrete cause: the discriminator judges each example separately and cannot notice that they all look alike; their remedy was to let it compare whole batches. Chollet (2021) collects a "bag of tricks," such as injecting randomness, to keep adversarial training from getting stuck (2021, sec. 12.5).
In language models. By extension, janus (2022) called mode collapse the tendency of some instruction-tuned models to concentrate probability on a few answers: asked for a random number between 0 and 100, text-davinci-002 favored 97. Kirk and colleagues (2024) measured that RLHF reduces output diversity compared with supervised fine-tuning. This is not the same as lowering the temperature: it changes what the model considers probable, not how much risk it takes when sampling (compare a base model).
It should not be confused with model collapse (Shumailov et al., 2024), which degrades models trained on data generated by other models. The general lesson has a family resemblance to Goodhart: optimizing against a judge yields what the judge rewards, not necessarily what the judge was meant to measure.