TOME-Math Raises AI Math Scores by 16.7 Percentage Points

A dataset assembled from 2,118 university mathematics textbooks lifted a base language model’s average score by 16.7 percentage points, giving researchers a structured source for training mathematical reasoning.

The gain is the unweighted mean across held-out evaluations and established public benchmarks. In the aggregate comparison, TOME-Math scored above each of the three public instruction-data baselines tested.

Short for Textbook-Oriented Mathematical Evidence, TOME-Math contains about 380,000 proof-supervision records. The records retain information about where material appears in a book, which propositions connect to proofs, how mathematical ideas refer to related material and which text each record came from.

That structure keeps parts of a textbook’s organization with the proof material. The corpus is designed to support curriculum-style learning across multiple areas of mathematics and further research into whether textbook organization affects AI reasoning.

The reported 16.7-point gain summarizes results across evaluations; it does not mean every individual test score rose by that amount. The comparison with the three public instruction-data baselines is also reported at the aggregate level.

The paper was published Nov. 13, 2026.