CreativityNeuro boosts LLM divergent thinking by 14 percentile points
New weight-steering technique improves creativity scores without any retraining.
CreativityNeuro, developed by Samuel Schapiro, Core Francisco Park, Felix Sosa, and Lav R. Varshney, is a novel approach to enhancing the creative output of large language models. Rather than requiring behavioral data, fine-tuning, or gradient-based training, it directly steers model weights in weight space to encourage divergent thinking. In tests, the method raised LLM scores on the Divergent Association Task (a vocabulary-based creativity metric) by up to 14 percentile points compared to baseline models. It also proved effective in a large-scale human evaluation (720 participants) using the Alternative Uses Test and Task Task, where it significantly improved ratings for originality, surprise, and overall creativity.
Importantly, CreativityNeuro demonstrated a measurable reduction in mode collapse—the tendency for LLMs to produce repetitive or overly similar responses to open-ended prompts. The researchers found that activation steering (altering model activations during inference) matched CreativityNeuro on the DAT but failed to transfer to longer, more open-ended tasks. Weight-space steering, by contrast, generalized robustly, making it a practical, plug-in enhancement for improving LLM creativity across diverse domains. The work was accepted at the ICML 2026 Workshop on Creativity & Generative AI.
- Improves LLM divergent thinking by up to 14 percentile points on the Divergent Association Task (DAT).
- Large-scale human evaluation (N=720) shows significant gains in originality, surprise, and creativity on the Alternative Uses Test and Task Task.
- Data-free method: no retraining, no gradient fine-tuning; weight-space steering generalizes better than activation steering.
Why It Matters
A lightweight, generalizable method to make LLMs more creative without costly retraining, useful for brainstorming and idea generation tools.