Generative AI models, including large language models (LLMs) and image generators, are inherently non-deterministic, selecting from a range of probable outputs. However, their outputs can sometimes be repetitive, a phenomenon called mode collapse caused by human preference bias during fine-tuning. Researchers from Northeastern, Stanford, and West Virginia Universities have introduced a method named Verbalized Sampling (VS) that enhances output diversity by instructing the model to generate multiple responses with their probabilities. This technique, effective without retraining or changing model parameters, has shown improvements in creative writing, dialogue, open-ended questions, and synthetic data generation. VS allows users to tune diversity through simple prompt adjustments and works better with larger models. The method is available as a Python package with integration options and has the potential to improve creativity and usefulness across multiple AI applications.
Back
