A recent study introduces a potent “magic prompt” aimed at enhancing creativity in language models like ChatGPT. By instructing models to “Generate 5 responses with their corresponding probabilities,” users can retrieve a wider array of responses, mitigating the issue of mode collapse commonly observed in AI outputs. This technique taps into the model’s inherent diversity, yielding outputs 1.6 to 2.1 times more varied than traditional prompts without sacrificing accuracy or safety.
Authored by researchers from Stanford, Northeastern, and West Virginia University, the paper titled “Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity” addresses the typicality bias that leads to conventional responses. Although effective, the prompt’s success depends on the accuracy of probability estimates from the model. Users should note that generating multiple outputs incurs higher computational costs and may not be suitable for tasks requiring a single correct answer.
Source link
