Researchers have found a way to make large language models less predictably boring.
A team introduced CreativityNeuro, a technique that steers model weights to improve divergent thinking without requiring any behavioral data, retraining, or gradient-based fine-tuning. On the Divergent Association Task - a vocabulary-space creativity benchmark - it lifted model performance by up to 14 human percentile points. In a 720-person human evaluation across two other creativity tests, models using CreativityNeuro scored higher on originality, surprise, and overall creativity, and those gains held up on longer, more open-ended tasks. The method also measurably reduced mode collapse, the tendency for a model to funnel toward a narrow band of predictable outputs.
This matters because mode collapse is not just an academic nuisance - it is why every AI brainstorming tool starts sounding the same after a few sessions. If weight steering can genuinely diversify outputs at inference time without a costly retraining cycle, that is a meaningful lever for developers building creative applications on top of existing models. The comparison to activation steering is telling: activation steering matched CreativityNeuro on one benchmark but failed to generalize to the harder open-ended tasks, suggesting that changing the weights themselves, not just the activations, is what makes the difference stick.
The "artificial hivemind" problem has been discussed for years; most proposed fixes involve prompt engineering or expensive fine-tuning. CreativityNeuro is notable for being neither - though peer review will ultimately determine how well these benchmarks translate to real creative work.