Mode collapse

In generative modeling, a failure to represent or produce enough of the target distribution's distinct modes; practitioners also use it loosely for repetitive outputs.

The term originated with generative adversarial networks, where a generator may cover only a small subset of the data distribution. Applied to language products, it often describes outputs that converge on the same structure, phrases, or stance across prompts. Repetition alone does not identify the cause: decoding settings, prompt templates, narrow examples, training data, or preference tuning can all reduce variation.

Builder example

Diversity matters when the task calls for distinct concepts, voices, or solutions. It is not automatically desirable for invoices, safety notices, or structured records, where consistency may be the requirement. Evaluate variation against the task rather than maximizing it in the abstract.

Common confusion: Mode collapse (one model producing repetitive outputs) is a different problem from model collapse (future models degrading because they trained on AI-generated data). The names are confusingly similar, but the causes and solutions are distinct.