---
format: "aidr-story-markdown/v1"
id: "5605712242da3270bef1e6ec6ae53983bc07b6f07c3378c3c7fab74e10a7caec"
canonical_url: "https://aidr.today/56057122?lang=en"
title: "Scaling Mathematical Proof Supervision from Structured Textbooks"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-11-12T00:00:00.000Z"
category: "Data"
topics: ["reasoning","data-engineering"]
source_urls: ["https://huggingface.co/papers/2610.40394"]
summary: "Researchers introduced TOME-Math, a corpus converting 2,118 university-level mathematics textbooks into approximately 380K structured proof-supervision records, which improved held-out evaluation scores by 16.7 percentage points over the base model when used for fine-tuning."
---

# Scaling Mathematical Proof Supervision from Structured Textbooks

> [Open the canonical story](<https://aidr.today/56057122?lang=en>)

**Published:** 2026-11-12T00:00:00.000Z
**Category:** Data
**Topics:** reasoning, data\-engineering

## Summary

Researchers introduced TOME\-Math, a corpus converting 2,118 university\-level mathematics textbooks into approximately 380K structured proof\-supervision records, which improved held\-out evaluation scores by 16\.7 percentage points over the base model when used for fine\-tuning\.

## Sources

- [Story source](<https://huggingface.co/papers/2610.40394>)