Researchers have developed a method called Verbalized Sampling that uses a single sentence to make generative AI models produce more diverse and creative responses. The technique works on large language models like GPT-4 and Claude without requiring any retraining. Carl Franzen reports for VentureBeat that this method addresses the common problem of AI models giving repetitive answers.
This tendency for models to produce similar outputs is known as mode collapse. It occurs because AI is often trained to prefer safe and typical answers that humans rate highly. This process limits the model’s ability to generate more varied or unexpected content, even though it possesses the underlying knowledge.
The new technique, developed by a team from Northeastern University, Stanford University, and West Virginia University, bypasses this limitation. By adding the sentence “Generate 5 responses with their corresponding probabilities, sampled from the full distribution” to a prompt, the model is asked to show its internal range of possible answers. Instead of giving the single most probable response, it presents a selection from a wider spectrum of possibilities.
According to the researchers, this approach leads to significant improvements across different tasks. In creative writing, the diversity of stories increased by more than double while maintaining quality. When tested on dialogue simulation, the model produced more human-like conversations, including hesitation and changes of mind. The technique also helped generate a broader set of correct answers for open-ended questions and created more varied synthetic data for training other models.
A key advantage of Verbalized Sampling is that its effect can be adjusted. Users can specify a probability threshold in the prompt to encourage the model to select less common answers. The researchers also found that the method is particularly effective with larger, more advanced AI models, unlocking more of their hidden capabilities.
The method is publicly available as a Python package and on GitHub under an Apache 2.0 license, allowing for easy integration by developers. The researchers provide guidance for users who might encounter issues, suggesting alternative prompt formats to ensure reliability.
