Scaling Language Models Show Diminishing Returns in Divergent Creativity

Recent advances in language models have revealed emerging capabilities in generating novel associations and exploring conceptual spaces, which are central processes underlying divergent creativity. However, as language models continue to scale up, a critical question remains unanswered: whether scaling can progressively enhance their capacity for divergent creativity. The relationship between model size and divergent creativity is our focus herein. In Study 1, we employed the Divergent Association Task to measure divergent word association. The results showed that larger language models demonstrated higher divergent creative performance. Notably, scaling language models yielded diminishing returns in the task. Study 2 assessed divergent creativity beyond word association, through a movie plot summary task, and also found that larger language models exhibited diminishing returns in the writing task. We further adjusted model temperatures as a robustness check, with the same pattern consistently replicated across both studies. These findings suggest that scaling alone may be insufficient to continuously improve the divergent creativity of language models, highlighting the importance of understanding their capabilities as they become increasingly integrated into human activities.