CreativityNeuro Improves LLM Divergent Thinking via Weight Steering
July 3, 2026
CreativityNeuro uses contrastive weight steering to enhance LLM originality without additional training data. The method improves performance on the Divergent Association Task by up to 14 human percentile points and reduces mode collapse in open-ended generation.
HOW THIS AFFECTS YOU
●
researcherYou can apply this data-free weight steering to mitigate the artificial hivemind effect in models.
●
designerThis enables more diverse and less repetitive outputs for generative creative tools.