Can LLMs brainstorm without looking at past ideas?

When using LLMs to generate pools of creative ideas, you want breadth without sacrificing quality. This work compares inference-time strategies for diversifying candidate pools: some that anchor to seed ideas, others that don't. Across three creative tasks, semantic direction stratification—using a single planning call to organize generations across different semantic directions—outperformed anchored baselines on the diversity-quality-efficiency frontier, with population-referential divergence providing a surprisingly strong low-cost option.