Where does mode collapse in language models really come from?

Post-training alignment narrows LLM output to a few favored responses — mode collapse — which kneecaps creative writing, social simulation, pluralistic alignment, and synthetic-data generation. Prior work blames algorithmic causes (inadequate reward models, majority-favoring optimization). This paper relocates the cause to the data: a pervasive typicality bias in preference data, where annotators systematically prefer familiar, typical text (a well-established cognitive-psychology effect). Mode collapse is thus an inherent property of preference data itself, formalized theoretically and verified on real preference datasets.

The fix follows from the diagnosis and is strikingly cheap: Verbalized Sampling (VS), a training-free prompting strategy that asks the model to verbalize a distribution over responses with their probabilities ("generate 5 jokes and their probabilities") rather than a single answer. Across creative writing, dialogue simulation, open-ended QA, and synthetic data, VS lifts diversity 1.6-2.1× over direct prompting without sacrificing factual accuracy or safety — and, tellingly, more capable models benefit more, suggesting the diversity was latent and suppressed rather than absent.

This reframes the vault's diversity-collapse thread at the source. Since Does outcome-based RL diversity loss spread across unsolved problems? locates collapse in RL dynamics, VS adds the data-level origin and a decoding-time remedy; and it offers a practical lever for Why do LLMs generate novel ideas from narrow ranges? — verbalize the distribution to surface the suppressed tail.