---
title: "Qwen2.5-Math Technical Report: Toward Mathematical Expert Model via Self-Improvement"
license: CC-BY-4.0
license_url: https://creativecommons.org/licenses/by/4.0/
license_note: 'paper is arXiv non-exclusive distribution license (verified on arXiv abs page 2026-11-12); this pointer note is our own words and is CC BY 4.0; we link to the paper rather than reproduce it'
source: https://iinventedlight.com/library/notes/2409-12122-qwen2.5-math.md
source_license: "arXiv non-exclusive distribution license"
attribution: 'Pointer by iinventedlight (AI-assisted). Source: An Yang, Beichen Zhang, Binyuan Hui, and 13 coauthors, Qwen2.5-Math Technical Report: Toward Mathematical Expert Model via Self-Improvement (arXiv non-exclusive distribution license). This note links to the paper rather than reproducing it.'
author: 'iinventedlight (AI-assisted)'
date: 2026-11-12
status: published
approved_by: "Cole Withington, 2026-11-12 14:44 PT, Librarian chat widget t51s316: \"Yes, publish all 100 second-list arXiv notes\""
checked_on: 2026-11-12
as_of: 2026-11-12
confidence: medium
untested: 'Pointer note only. Paper not reproduced. Abs page and license checked 2026-11-12.'
category: arxiv
topic_area: 'ai-research: reasoning'
tags: arxiv, reasoning, math, qwen, self-improvement
arxiv_id: 2409.12122
alias: 2409-12122-qwen2.5-math
type: pointer
format: pointer
sources:
  - id: S1
    title: "Qwen2.5-Math Technical Report: Toward Mathematical Expert Model via Self-Improvement"
    publisher: arXiv
    authors: "An Yang and 15 coauthors (listed on the arXiv abs page)"
    url: https://arxiv.org/abs/2409.12122
    version_read: "v1 (18 Sep 2024)"
    accessed: 2026-11-12
    license: "arXiv non-exclusive distribution license"
---
# Qwen2.5-Math Technical Report: Toward Mathematical Expert Model via Self-Improvement

## Summary
[Qwen2.5-Math Technical Report: Toward Mathematical Expert Model via Self-Improvement](https://arxiv.org/abs/2409.12122) (2024) describes the Qwen2.5-Math and Qwen2.5-Math-Instruct models at 1.5B, 7B, and 72B, trained with a self-improvement loop rather than a single supervised pass.

What it introduced: self-improvement across pretraining, post-training, and inference. A prior math model generates pretraining data; a reward model is built from massive sampling and is used to iterate supervised data; that stronger model then updates the reward model; reinforcement learning is applied at the end; and the same reward model can guide sampling at inference. The instruction models support Chinese and English and can use chain-of-thought or tool-integrated reasoning. Evaluation covers ten math datasets, from grade-school sets such as GSM8K to contest sets such as AIME24.

What it led to: a strong open math line sitting next to DeepSeekMath and later Qwen3 reasoning models. Live notes: [2402-03300-deepseekmath](https://iinventedlight.com/library/notes/2402-03300-deepseekmath/) and [2110-14168-gsm8k](https://iinventedlight.com/library/notes/2110-14168-gsm8k/). The Qwen2.5 and Qwen3 reports are pointer notes in this week's batch F.

Where to read it: https://arxiv.org/abs/2409.12122 (v1, 18 Sep 2024).

The paper is posted on arXiv under the arXiv non-exclusive distribution license. This note links to it rather than reproducing it. The note itself is CC BY 4.0.

License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Pointer by iinventedlight (AI-assisted). Paper license: arXiv non-exclusive distribution; this note links to it and does not reproduce it.

---

Agents: [/agents.md](/agents.md) | [/.well-known/agent-library.json](/.well-known/agent-library.json).
