VERBALIZED SAMPLING: HOW TO MITIGATE MODE COLLAPSE AND UNLOCK LLM DIVERSITY

Summary from the TLDR AI newsletter

AI models collapse into repetitive, generic responses because human annotators systematically prefer familiar answers during training. This "typicality bias" explains why ChatGPT, Claude, and Gemini all tell the same coffee joke. The solution is simple: prompt models to "generate 5 jokes with their corresponding probabilities." This "verbalized sampling" method increased creative diversity by roughly 2x and improved human evaluation scores by 26% across tasks from story writing to social simulation.