Paper #7: Structured Creativity Methods for Multi-Agent LLMs: Brainwriting Outperforms Disney and Double Diamond on LLM-Judged Originality

Large language models (LLMs) are increasingly used to support idea generation, yet most applications rely on simple one-shot prompting. We investigate whether structured creativity methods from human group ideation (Brainwriting, the Disney Method, and the Double Diamond) can systematically improve LLM creative output along both dimensions of creativity, namely originality and usefulness. In a factorial exper- iment, we compare three methods in two configurations each (multi-agent with four separate LLM instances vs. a single LLM simulating all roles) against two baselines. The design crosses three generator models, five applied creative tasks, and five LLM-as-judge evaluators (600 evaluator ratings). Multi-agent Brainwriting achieves the highest LLM-judged originality. Even simulated single-LLM Brainwriting captures roughly 70% of this gain. Disney and Double Diamond yield no significant originality improvements. Usefulness does not differ significantly across conditions. Under LLM-as-judge evaluation, process structure appears to matter more than agent architecture: the iterative build-on-others design of Brainwriting, not merely the presence of multiple agents, drives the observed originality gains.