The Chain-of-Thought Diversity Effect
Breaking idea generation into explicit sequential steps dramatically increases output diversity, nearly matching human group performance.
Chain-of-Thought prompting achieved 0.255 cosine similarity (where lower = more diverse), compared to 0.377 for baseline prompts and 0.243 for human groups. The technique involves:
- Generate a short list of ideas (titles only)
- Review and modify ideas to be “bolder and more different”
- Expand each into full descriptions
This staged approach forces the model to generate variety before committing to details. By separating ideation from elaboration, CoT prevents early outputs from anchoring subsequent ones.
The effect persists until about 750 ideas, when both strategies converge as the idea space becomes exhausted. CoT’s advantage is most pronounced in the first 500 ideas, exactly where it matters most for practical brainstorming.
Related: 05-atom—ai-diversity-deficit, 05-atom—idea-exhaustion-dynamics, 05-molecule—prompt-engineering-for-diversity