Tag
Introduces SDABench, a benchmark evaluating LLMs on six scientific analysis capabilities across five domains, finding models struggle with tasks requiring assumption selection and mechanistic reasoning.