Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity

Post-training alignment often reduces LLM diversity, leading to a phenomenon known as mode collapse. Unlike prior work that attributes this effect to algorithmic limitations, we identify a fundamental, pervasive data-level driver: typicality bias in preference data, whereby annotators systematically favor familiar text as a result of well-established findings in cognitive psychology. We formalize this bias theoretically, verify it on preference datasets empirically, and show that it plays a central role in mode collapse. Motivated by this analysis, we introduce Verbalized Sampling (VS), a simple, training-free prompting strategy to circumvent mode collapse. VS prompts the model to verbalize a probability distribution over a set of responses (e.g., “Generate 5 jokes about coffee and their corresponding probabilities”). Comprehensive experiments show that VS significantly improves performance across creative writing (poems, stories, jokes), dialogue simulation, open-ended QA, and synthetic data generation, without sacrificing factual accuracy and safety. For instance, in creative writing, VS increases diversity by 1.6-2.1× over direct prompting. We further observe an emergent trend that more capable models benefit more from VS.

Introduction. Post-training alignment methods like RLHF can unintentionally cause mode collapse (Janus, 2022; O’Mahony et al., 2024; Kirk et al., 2024b), whereby the model favors a narrow set of responses (the “mode”) over all plausible outputs, as shown in Figure 1. This significantly reduces output diversity (Padmakumar & He, 2024; West & Potts, 2025) and limits LLMs’ effectiveness in various applications such as creative writing (Lu et al., 2025b), social simulation (Anthis et al., 2025b), pluralistic alignment (Kirk et al., 2024a), and synthetic data generation (Zhu et al., 2025a). Existing work often attributes mode collapse to algorithmic causes such as inadequate reward models (Chakraborty et al., 2024) or the majority-favoring optimization process (Xiao et al., 2024). In this paper, we show that the issue is more fundamental: mode collapse is an inherent property of preference data itself. We identify typicality bias, the human tendency to prefer more typical text, as a pervasive data-level cause for mode collapse.