SciAidanBench: Evaluating LLM Scientific Creativity

Large language models (LLMs) have demonstrated impressive general creativity, but their capacity for scientific creativity—generating novel, domain-relevant ideas—remains underexplored. We introduce SciAidanBench, a benchmark of 155 open-ended scientific questions designed to measure scientific idea generation in LLMs. We evaluate 19 models across seven vendors and observe a strong positive correlation (r = 0.81) between general and scientific creativity, though models consistently produce fewer scientific ideas per prompt. Vendor-specific variations and performance gaps between non-reasoning and reasoning variants of models are observed. This preliminary investigation provides initial findings on evaluating the scientific creativity of LLMs.