CreativityNeuro: Steering Language Model Weights to Improve Divergent Thinking and Reduce Mode Collapse
06:00 · July 3, 2026 · arXiv cs.AI RSS
Divergent thinking is a crucial aspect of creativity, yet large language models (LLMs) tend to consistently generate similar responses to open-ended questions, in what has been termed the artificial hivemind effect. Here, we introduce CreativityNeuro, a data-free method for enhancing divergent thinking in LLMs via contrastive weight steering. We evaluate our method across multiple creativity assessments and report several main findings. On the Divergent Association Task (DAT), a vocabulary-space creativity test, CreativityNeuro improves performance by up to 14 human percentile points. Next, in a large-scale human evaluation (N=720) on the Alternative Uses Test (AUT) and the Task Task, CreativityNeuro achieves significant improvements in originality, surprise, and creativity, transferring to longer-form and more open-ended tasks. Importantly, we find that across all three tasks, CreativityNeuro demonstrably reduces measures of mode collapse. Moreover, activation steering achieves comparable performance to CreativityNeuro on the DAT, but it does not transfer to the AUT and Task Task, demonstrating the effectiveness of weight-space steering in generalizing to unseen tasks. In conclusion, CreativityNeuro improves divergent thinking and reduces mode collapse without requiring behavioral data, re-training, or gradient-based fine-tuning, providing a straightforward way to enhance LLM performance in creative domains.
Summary
CreativityNeuro is a data-free technique that steers the weights of large language models to increase divergent thinking while curbing the artificial hivemind effect, in which models repeatedly produce similar answers to open-ended prompts. The approach relies on contrastive prompt sets—one set encouraging creative responses and the other favoring conventional ones—to compute per-weight importance scores across layers. It then isolates a sparse subset of creativity-specific parameters by subtracting those most active under non-creative prompts and applies a scaled multiplicative perturbation to those weights only.
Evaluations show consistent gains. On the Divergent Association Task, a vocabulary-based test of remote associations, the method lifts model performance by as much as 14 human percentile points. In a 720-participant human study using the Alternative Uses Test and the more open-ended Task Task, responses generated after steering received higher ratings for originality, surprise, and overall creativity. The same interventions measurably reduced indicators of mode collapse across all three tasks.
Unlike activation steering, which matches CreativityNeuro on the DAT but fails to transfer to longer-form tasks, weight-space adjustments generalize to unseen prompt distributions. The procedure requires no behavioral datasets, gradient updates, or retraining, distinguishing it from prompting frameworks, temperature tuning, and reinforcement-learning approaches that depend on labeled preference data.
Why it matters
This research provides Dutch AI researchers and practitioners with a resource-efficient method to improve LLM creativity and mitigate mode collapse. Its data-free, weight-steering approach aligns with the Netherlands' focus on sustainable, controllable, and innovative AI development.
