A research team has released SFBench, a benchmark designed to test whether AI can tell plausible science from nonsense.
The dataset contains 197 materials science claims, each scored on a five-point feasibility scale by subject matter experts. Unlike most benchmarks, the claims were written from scratch rather than pulled from existing publications — a deliberate choice to cut down on the risk that large language models have already seen the answers during training. The explanations are fully open-ended, not multiple choice or short-answer, and GPT models were used to establish baseline results.
That design matters because most AI benchmarks have a contamination problem: if the test data overlaps with training data, scores tell you more about memorization than reasoning. By commissioning novel claims from domain experts, SFBench tries to close that loophole in a domain — materials science — where plausible-sounding but physically impossible claims are easy to generate and hard to debunk without real expertise.
Whether the benchmark holds up over time is a separate question. Once SFBench is public, it becomes training fodder for the next generation of models — the same contamination risk it was built to avoid.