---
title: "Math-Shepherd: Verify and Reinforce LLMs Step-by-step without Human Annotations"
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/2312-08935-math-shepherd.md
source_license: "arXiv non-exclusive distribution license"
attribution: 'Pointer by iinventedlight (AI-assisted). Source: Peiyi Wang, Lei Li, Zhihong Shao, and 6 coauthors, Math-Shepherd: Verify and Reinforce LLMs Step-by-step without Human Annotations (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, process-reward, math, ppo
arxiv_id: 2312.08935
alias: 2312-08935-math-shepherd
type: pointer
format: pointer
sources:
  - id: S1
    title: "Math-Shepherd: Verify and Reinforce LLMs Step-by-step without Human Annotations"
    publisher: arXiv
    authors: "Peiyi Wang and 8 coauthors (listed on the arXiv abs page)"
    url: https://arxiv.org/abs/2312.08935
    version_read: "v3 (19 Feb 2024)"
    accessed: 2026-11-12
    license: "arXiv non-exclusive distribution license"
---
# Math-Shepherd: Verify and Reinforce LLMs Step-by-step without Human Annotations

## Summary
[Math-Shepherd: Verify and Reinforce LLMs Step-by-step without Human Annotations](https://arxiv.org/abs/2312.08935) (2023) trains a process reward model that scores each step of a math solution, using automatically built step-level data instead of human step labels.

What it introduced: two uses for that step scorer. One is verification: rerank several model samples. The other is reinforcement learning: step-by-step PPO against the scorer. The authors report gains on GSM8K and MATH for open models such as Mistral-7B, with further gains when the same scorer is also used to rerank.

What it led to: later automated process-supervision work, including OmegaPRM in this week's batch F, and process rewards in math-specialized models. Live notes: [2110-14168-gsm8k](https://iinventedlight.com/library/notes/2110-14168-gsm8k/), [2402-03300-deepseekmath](https://iinventedlight.com/library/notes/2402-03300-deepseekmath/), and [1707-06347-ppo](https://iinventedlight.com/library/notes/1707-06347-ppo/).

Where to read it: https://arxiv.org/abs/2312.08935 (v3, last revised 19 Feb 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).
