CreativityNeuro Enhances LLM Divergent Thinking and Reduces Mode Collapse

Large language models (LLMs) often suffer from an "artificial hivemind effect," where they generate similar responses to open-ended prompts, limiting their creative output. This research introduces CreativityNeuro, a novel data-free method designed to enhance divergent thinking in LLMs through contrastive weight steering. This technique directly manipulates the model's weights to encourage more varied and original responses. The method was evaluated across multiple creativity assessments. On the Divergent Association Task (DAT), CreativityNeuro improved performance by up to 14 human percentile points. A large-scale human evaluation (N=720) on the Alternative Uses Test (AUT) and the Task Task further demonstrated significant improvements in originality, surprise, and overall creativity for longer, more open-ended tasks. Crucially, CreativityNeuro effectively reduces "mode collapse," a phenomenon where models repeatedly produce similar outputs. The study also found that while activation steering showed comparable performance on the DAT, it failed to generalize to the AUT and Task Task, highlighting the superior effectiveness of weight-space steering for broader creative applications. This method offers a straightforward way to boost LLM creativity without requiring new data, retraining, or gradient-based fine-tuning.