Research
Expert audit finds physics benchmark errors can more than double frontier-model scores
Experts re-graded six physics benchmarks and found large grader and benchmark errors, showing why model-selection pipelines need auditable evaluation quality.