Unlocking More Creativity Out Of AI Models

Large language models often give repetitive answers after alignment training, a problem called mode collapse. Reporting in a joint paper between Northeastern, Stanford and West Virginia Universities, researchers found that this happens partly because human raters prefer “typical” or familiar responses, a typicality bias built into training data. To fix this, they propose Verbalized Sampling (VS), a new prompting method that asks the model to generate several answers along with their probabilities (for example, “Give me 5 jokes with their likelihoods”). This helps the model recover its natural variety without retraining. Tests on creative writing, dialogue, and question answering show VS increases diversity up to twice as much as normal prompts, while keeping accuracy and safety stable.

Cornell University, arXiv – Northeastern University / Stanford University / West Virginia University, Jiayi Zhang et al – Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity – 1st October 2025