To speed up materials discovery, language models must be able to tackle problems at the research frontier. Existing textbook‑ and exam‑derived benchmarks mainly assess established knowledge and leave performance on cutting‑edge research largely untested. We introduce FrontierMat, a benchmark containing 330 problems drawn from frontier academic and industrial work in aerospace, semiconductors, embodied intelligence and nuclear fusion. Each problem is validated by leading experimentalists and industry experts. The problems present a real study and ask the model to infer a hidden outcome or draw a conclusion from selected experimental evidence. Problems are annotated with functional domain, material system and the scale of the governing mechanism; 130 require cross‑scale reasoning. The strongest general‑purpose model reaches only 53.7% accuracy, with cross‑scale items scoring 8.7 points lower than single‑scale ones. About half of the problems are solved by at most one model, and failures are mainly due to models favoring plausible answers over what was actually observed. We conduct a domain‑adaptation study using meta‑weighted continued pretraining and scale‑modular self‑distillation, raising an open‑weight model from 36.7% to 55.4%, matching the best proprietary model, while improving on public materials benchmarks and not degrading performance on general benchmarks. FrontierMat provides a testbed for measuring progress toward language models that support frontier materials R&D.
Review: FrontierMat exposes critical gaps in current models—especially in cross‑scale reasoning and fidelity to experimental observations—offering a concrete target for future domain‑specific pretraining and self‑distillation strategies.