AI models can't generate diverse ideas because they don't know what's valid

Researchers found that large language models (LLMs) struggle to produce diverse outputs. It turns out the problem is not how they pick words, but how they assign probabilities to options. This means simply changing the sampling method will not make an AI more creative or better at exploring scientific ideas.

Developers thought they could make AI models more creative by tweaking how they sample words. This paper shows the problem is more fundamental: the models themselves do not reliably distinguish between good and bad options. They also put too much weight on a few common ones. This means building truly diverse AI will require rethinking how these models learn to assign probabilities.