---
title: "Improve Mathematical Reasoning in Language Models by Automated Process Supervision"
license: CC-BY-4.0
license_url: https://creativecommons.org/licenses/by/4.0/
license_note: 'paper is CC BY 4.0 (verified on arXiv abs page 2026-11-12; icon by-4.0.png, http://creativecommons.org/licenses/by/4.0/); note is our own summary'
source: https://iinventedlight.com/library/notes/2406-06592-omegaprm.md
source_license: 'CC BY 4.0'
attribution: 'Summary by iinventedlight (AI-assisted). Source: Liangchen Luo, Yinxiao Liu, Rosanne Liu, and 9 coauthors, Improve Mathematical Reasoning in Language Models by Automated Process Supervision (CC BY 4.0).'
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: "Not reproduced. Abs page and HTML key sections read 2026-11-12. Code not run."
category: arxiv
topic_area: 'ai-research: reasoning'
tags: arxiv, reasoning, process-supervision, mcts, omegaprm
arxiv_id: 2406.06592
alias: 2406-06592-omegaprm
sources:
  - id: S1
    title: "Improve Mathematical Reasoning in Language Models by Automated Process Supervision"
    publisher: arXiv
    authors: "Liangchen Luo and 11 coauthors (listed on the arXiv abs page)"
    url: https://arxiv.org/abs/2406.06592
    version_read: "v2 (11 Dec 2024)"
    accessed: 2026-11-12
    license: 'CC BY 4.0'
---
# Improve Mathematical Reasoning in Language Models by Automated Process Supervision

## Summary
OmegaPRM is a divide-and-conquer Monte Carlo Tree Search method for collecting process-supervision data without human step labels. Earlier collection relied on humans or expensive per-step Monte Carlo estimates. OmegaPRM uses binary search to find the first error in a chain of thought and to balance positive and negative examples. The authors collect over 1.5 million process-supervision annotations and train process reward models. With weighted self-consistency, they raise instruction-tuned Gemini Pro from 51% to 69.4% on MATH500 and from 86.4% to 93.6% on GSM8K, and Gemma2 27B from 42.3% to 58.2% and 74.0% to 92.2% on the same sets. [S1]

## Claims
- id: C1
  claim: "OmegaPRM is a divide-and-conquer MCTS method that finds the first CoT error with binary search."
  source: S1
  evidence: "identifies the first error in the Chain of Thought (CoT) with binary search"
- id: C2
  claim: "The authors collect over 1.5 million process-supervision annotations to train process reward models."
  source: S1
  evidence: "collect over 1.5 million process supervision annotations"
- id: C3
  claim: "Instruction-tuned Gemini Pro rises from 51% to 69.4% on MATH500 under this process supervision."
  source: S1
  evidence: "from 51% to 69.4% on MATH500"

## Verification
- Code available: No official code URL is named on the abs page. Not run.
- Reproduced: not reproduced (2026-11-12). Code not run.
- Related work: See [2201-11903-chain-of-thought-prompting](https://iinventedlight.com/library/notes/2201-11903-chain-of-thought-prompting/), [2110-14168-gsm8k](https://iinventedlight.com/library/notes/2110-14168-gsm8k/), and [2402-03300-deepseekmath](https://iinventedlight.com/library/notes/2402-03300-deepseekmath/). Math-Shepherd is a pointer note in batch H.
- Limits: The headline gains are for specific models and two math sets. The method still needs many model rollouts. No official code URL is given.
- License: CC BY 4.0 on the abs page (2026-11-12; icon by-4.0.png; http://creativecommons.org/licenses/by/4.0/). Submitted 5 Jun 2024 (v1); last revised v2 (11 Dec 2024). Sources: [S1] https://arxiv.org/abs/2406.06592, v2 (11 Dec 2024). checked_on: 2026-11-12.

## When to apply
Use when asking how process rewards can be collected without human step labels, or what OmegaPRM reported on MATH500.

## Facts
- id: F1
  claim: "OmegaPRM automates process-supervision data collection with divide-and-conquer MCTS and binary search for the first error."
  confidence: medium
  date: 2026-11-12
  sources: [S1]
  license: CC BY 4.0
- id: F2
  claim: "The authors report more than 1.5 million process-supervision annotations."
  confidence: medium
  date: 2026-11-12
  sources: [S1]
  license: CC BY 4.0
- id: F3
  claim: "They report Gemini Pro MATH500 gains from 51% to 69.4% and GSM8K gains from 86.4% to 93.6%."
  confidence: medium
  date: 2026-11-12
  sources: [S1]
  license: CC BY 4.0

## Sources
- [S1] arXiv 2406.06592, https://arxiv.org/abs/2406.06592. License: CC BY 4.0. v2 (11 Dec 2024). Accessed 2026-11-12.

License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Summary by iinventedlight (AI-assisted). Paper is CC BY 4.0.

---

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