Most maths training data for language models arrives as loose question-and-answer pairs, cut away from the books and the reasoning they came from. A new arXiv preprint by researchers at the University of Science and Technology of China, Shanghai AI Laboratory, Fudan University and Shanghai Jiao Tong University argues that the structure is worth keeping. Their corpus, TOME-Math, turns 2,118 university textbooks across 29 subjects into 380,292 proof records, and a model trained on it gains more than ones trained on far larger public datasets.
The pipeline records where each item sits in its book, which theorem a proof belongs to, what it depends on and where it was taken from. It also creates "gap" questions out of steps that a textbook proof passes over quickly. The team fine-tuned Qwen3-30B-A3B-Base for one epoch on 364,631 records, after holding back 64 books for testing. Across six benchmarks (their own TOME-LongForm and TOME-Objective sets, ProofNet, MiniF2F, MATH-500 and AIME 2026), the average rose from 29.17% to 45.83%. The model came first on five of the six and second on AIME 2026.
The comparison with public data is the more interesting part. OpenMathInstruct-1 holds 1.82M examples, about 4.8 times as many, yet the TOME-trained model beat it on all six benchmarks, and it also finished ahead of MetaMathQA and MathInstruct. Controlled runs then test the structure directly. Training through chapters in book order, with gap questions first, beat shuffled data in all 18 combinations of subject and data size, and the lead grew to 7.72 points at the largest budget. Spreading the same budget across 16 books rather than 8 added 3.3 points on average.
The finding fits a wider lesson in model training: how data is organised can matter as much as how much of it there is. The base model belongs to the same Qwen family as Qwen3.6, which we review, and the recipe is aimed at open models of that kind rather than closed APIs. For people choosing a model for proof work, it suggests that a model's maths ability depends heavily on the shape of its fine-tuning data, which vendors rarely disclose.
There are clear limits. This is a preprint and has not been peer reviewed. The long-form proofs are graded by a model, gpt-5.6-luna, reviewing each answer eight times, with a human calibration check rather than full human marking. The authors concede that the public datasets differ from theirs in several ways, so the headline comparison alone does not prove that structure is the cause. Only one base model was tested. The corpus release is promised without a date or licence, and the paper does not explain how the rights to 2,118 copyrighted textbooks will be handled.