LLM Jaggedness Unlocks Scientific Creativity

A new study from arXiv introduces SciAidanBench, a benchmark measuring scientific creativity in large language models (LLMs). Researchers evaluated 19 base models across 8 providers (30 total variants) on open-ended scientific questions, counting unique and coherent ideas as a proxy for creative potential. The study finds that progress in LLM capabilities is jagged—uneven across tasks, domains, and model scales. Improvements in general creativity do not uniformly translate to scientific creativity, revealing jaggedness both across and within models. The work highlights the uneven nature of AI advancement and its implications for scientific discovery.