Daily papers

Thursday, 12 November 2026 · 6 min read

Structured textbook-based supervision improves mathematical reasoning in large language models

TOME-Math converts 2,118 university-level math textbooks into 380K structured proof QA pairs, preserving curricular and dependency structure. Fine-tuning a 30B model on this corpus raises average benchmark scores by 16.7 points over the base and outperforms public baselines.

Based on the paper Scaling Mathematical Proof Supervision from Structured Textbooks by Youqing Fang, Zhouqi Hua, Yinhao Tang, Yuzhe Gu and 6 others

Challenges in scaling mathematical supervision beyond isolated problem-solution pairs

Mathematical reasoning in language models depends heavily on the quality and organization of training data. Existing approaches typically flatten diverse sources into isolated problem-solution pairs, losing the rich curricular and evidential structure present in original textbooks.

Textbooks organize mathematical knowledge hierarchically: definitions introduce concepts, lemmas and theorems build on them, proofs establish validity, and exercises reinforce understanding. This structure encodes dependencies and pedagogical progression that isolated QA pairs discard.

Prior work has used textbooks as sources of math data but rarely preserves their native structure such as chapter order, proposition dependencies, and source provenance. This limits the ability to scale natural-language proof supervision while leveraging the inherent organization of mathematical knowledge.

Sign in to read the whole paper

You have read the opening. The rest is free with a Bezen account, and no card is needed.

In the rest of this paper

  • How the method works, step by step, with a flow chart
  • The figure from the paper, and how to read it
  • How it was tested, what was found, and where it falls short
Structured textbook-based supervision improves mathematical reasoning in large language models · Bezen