buildbox labs
understanding AI capability in the physical sciences
Scientific discovery happens where theory meets reality. Measurements surprise us, materials vary, and experiments produce results we didn’t expect. Those differences often tell us what to investigate next.
We benchmark frontier models across long-running, multi-agent scientific investigations, starting with power electronics and electrical engineering tasks. Our tasks hinge on real research, with experimental evidence and physical constraints shaping the decisions models have to make.
our approach
Our benchmarks and testing environments are grounded in scientific reasoning, specialized software, and experimental evidence. We examine both results and methods: checking calculations, comparing predictions with measurements, and testing the validity of conclusions. Recognizing uncertainty (and knowing when more evidence is needed) is part of scientific capability, too.
work with us
We’re looking for scientists and AI researchers to help define meaningful tests of scientific capability. If you have a research problem that could reveal what AI can or cannot do, .