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

2025-10-01

AI Summary

Post-training alignment frequently induces mode collapse in large language models (LLMs), primarily due to typicality bias inherent in preference data. This work is the first to identify—on the data level—that such bias is a key driver of reduced response diversity. To address it, we propose Verbalized Sampling: a training-free, inference-time prompting method that explicitly elicits from the model multiple candidate responses alongside their estimated probability distributions. Grounded in cognitive psychology theories of typicality and validated through empirical analysis, our approach employs a structured prompt template to mitigate bias during decoding. Experiments across creative writing, dialogue simulation, open-ended question answering, and synthetic data generation demonstrate 1.6–2.1× improvements in response diversity, with no degradation in factual accuracy or safety. Notably, gains scale with model capability. This work establishes a novel “training-free, prompt-driven” paradigm for enhancing generative diversity in LLMs.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Cognitive Modeling & Cognitive Systems: Computational Creativity

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Search and Retrieval-Augmented AI: Large language models for search
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

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, 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 1.6-2.1x over direct prompting. We further observe an emergent trend that more capable models benefit more from VS. In sum, our work provides a new data-centric perspective on mode collapse and a practical inference-time remedy that helps unlock pre-trained generative diversity.

Problem

• Addresses mode collapse in LLMs caused by typicality bias in preference data
• Introduces Verbalized Sampling to enhance diversity without sacrificing accuracy
• Improves creative writing, dialogue, and QA through training-free prompting strategy

Innovation

• Verbalized Sampling prompts probability distribution verbalization
• Training-free strategy mitigates mode collapse in LLMs
• Method improves diversity across creative writing tasks