New Paper Reveals Why LLMs Lack Diversity in Random Outputs