Simple Prompt Addition Doubles AI Output Diversity While Maintaining Quality, Researchers Find

Researchers discover that adding a single sentence to AI prompts doubles output diversity by up to 2.1× while maintaining quality, with the technique called Verbalized Sampling bypassing AI models' tendency to give repetitive answers and showing particular success in creative writing and dialogue tasks.

• Researchers at Northeastern University, Stanford University and West Virginia University discover that adding the sentence 'Generate 5 responses with their corresponding probabilities, sampled from the full distribution' to AI prompts increases output diversity by up to 2.1× while maintaining quality
• The method called Verbalized Sampling (VS) works by bypassing AI models' tendency toward 'mode collapse' where they give repetitive, safe answers due to human feedback training that favors familiar responses
• VS shows improved performance across creative writing, dialogue simulation, and data generation tasks, with larger models like GPT-4 and Claude showing even greater diversity gains than smaller models