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

arXiv cs.CL · 6 hours ago

Researchers identified that typicality bias in preference data during alignment training causes large language models to suffer from mode collapse, reducing their output diversity. Verbalized Sampling, a prompting technique that asks models to generate multiple responses with probabilities, increased creative writing diversity by 1.6-2.1x over standard prompting without harming accuracy. This inference-time method addresses mode collapse by leveraging the model's pre-trained generative capabilities without requiring retraining.