Examining and Addressing Barriers to Diversity in LLM-Generated Ideas
Yuting Deng, Melanie Brucks, Olivier Toubia
Use ordinary personas plus CoT prompting when generating ideas with LLMs. Avoid celebrity or "creative genius" personas—they don't improve diversity. This combo beats human groups while preserving AI efficiency.
LLMs generate less diverse ideas than human groups, threatening to homogenize innovation if widely adopted. Two mechanisms drive this: individual fixation (early outputs constrain later ones) and collapsed knowledge partitioning (LLMs aggregate knowledge into one distribution instead of sampling distinct regions).
Method: Chain-of-Thought prompting reduces fixation in LLMs (not humans) by forcing structured reasoning. Ordinary personas—not "creative entrepreneurs" like Steve Jobs—improve knowledge partitioning by anchoring generation in distinct semantic regions. Combining both interventions produces idea diversity exceeding human baselines. The persona effect works because ordinary roles serve as diverse sampling cues rather than creative inspiration.
Caveats: Tested on ideation tasks. Transfer to other creative domains (design, strategy) unverified.
Reflections: Do these interventions maintain diversity over longer ideation sessions, or does fixation eventually reassert itself? · Can knowledge partitioning be improved further by dynamically rotating personas during generation? · Does the diversity advantage persist when humans iterate on LLM-generated ideas versus generating independently?