What it measures
Human-annotated problems from US national math competitions, with 4,780 automatically checkable short-answer questions spanning six difficulty levels. The paper also studies multiple-choice prompting and inference-time sampling.
Reading the measurement
A correct final answer, a complete proof and useful research progress are different achievements. Choose the one the task actually grades. Inspect the permitted computation, attempt budget and handling of invalid arguments before comparing models.
Published evaluation
Overall short-answer accuracy over 4,780 automatically checkable problems; MATH results in the last column are a separate benchmark.
Where this fits in an evaluation suite
Final-answer accuracy, proof credit, and expert-reviewed research progress are distinct signals. A proof-aware evaluation should retain partial work and audit strong claims.
Start with a capability hypothesis, then choose a second test with a different failure mode. Keep the checkpoint, reasoning effort, agent tools, sampling policy, and scoring procedure fixed when measuring a training intervention. Record failures and missing outputs in the denominator.
A reproducible comparison
Overall short-answer accuracy over 4,780 automatically checkable problems; MATH results in the last column are a separate benchmark.
Publish the exact run configuration, item coverage, score units, and available uncertainty. If a source supplies only an aggregate or a single run, preserve that limit. A new model name or a rounded score alone does not establish a capability improvement.
PRIMARY EVIDENCE & FURTHER READING
Reviewed 2026-10-27. Explanation and interpretation by RLTasks. Benchmark ownership remains with its original authors.