News
TOME-Math, a structured corpus from 2,118 textbooks, enhances AI mathematical reasoning by 16.7 percentage points, surpassing public baselines.
A groundbreaking AI research paper, Scaling Mathematical Proof Supervision from Structured Textbooks, published on November 13, 2026, introduces TOME-Math (Textbook-Oriented Mathematical Evidence), a large-scale corpus designed to enhance mathematical reasoning in language models. Developed by lead authors Youqing Fang, Zhouqi Hua, Yinhao Tang, Yuzhe Gu, and Lingkai Kong, the study demonstrates a significant 16.7 percentage-point improvement in model performance when fine-tuned on this dataset.
Structured Mathematical Supervision
TOME-Math transforms 2,118 university-level mathematics textbooks into approximately 380,000 structured proof-supervision records. Unlike traditional datasets that flatten mathematical problems into isolated problem-solution pairs, TOME-Math preserves the curricular organization, including chapter order, proposition-proof links, dependency references, and source provenance. This structured approach enables curriculum learning across multiple mathematical subjects and provides a scalable framework for studying the impact of data organization on mathematical reasoning.
Performance Improvements
Fine-tuning a base language model on TOME-Math resulted in a 16.7 percentage-point increase in the unweighted mean score across held-out evaluations and established public benchmarks. The corpus outperformed all three public instruction-data baselines tested in aggregate, highlighting its effectiveness in advancing AI mathematical reasoning capabilities.
Implications and Future Research
The authors emphasize that structured textbooks represent a scalable and underutilized source of mathematical supervision. By releasing the TOME-Math corpus, the research team aims to facilitate further exploration of structured mathematical supervision and its applications in AI. This development underscores the potential of leveraging structured educational materials to enhance the reasoning capabilities of large language models.
Credit: Source link
