Definition
Mode collapse is a failure pattern in which a generative system produces a much narrower range of outputs than the true diversity of the underlying data or task, repeatedly converging on the same answer, phrasing, or pattern. In studies comparing many models side by side, this shows up starkly: dozens of independently trained systems will overwhelmingly output the same single word or phrase for a given category rather than a healthy spread of plausible alternatives. It signals that a model has learned a narrow set of high-probability shortcuts rather than genuinely representing the diversity present in its training distribution.