SDABench Benchmark Tests LLMs Across Six Scientific Reasoning Capabilities

Researchers have introduced SDABench, a new benchmark designed to evaluate large language models on six distinct scientific capabilities: descriptive, exploratory, inferential, predictive, causal, and mechanistic analysis. Unlike traditional benchmarks that focus on code execution or workflow completion, SDABench assesses how well LLMs support different types of scientific claims across five domains — Biology, Chemistry, Environment, Geography, and Physics. The dataset includes 527 real-world instances and 6,000 synthetic instances, offered in both multiple-choice and open-ended formats. A five-stage error analysis framework is also built in, helping identify precisely where models break down during scientific reasoning. An evaluation of 15 LLMs found that while models handled descriptive tasks reasonably well, performance dropped significantly on tasks requiring deeper scientific reasoning and rigor.