---
title: "LMM-R1: Empowering 3B LMMs with Strong Reasoning Abilities Through Two-Stage Rule-Based RL"
canonical_url: "https://www.modelscope.cn/papers/125170"
md_url: "https://www.modelscope.cn/papers/125170.md"
arxiv_id: 2503.07536
published: 2025-03-11
last_updated: 2025-03-11
authors:
  - "Yingzhe Peng"
  - "Gongrui Zhang"
  - "Miaosen Zhang"
  - "Zhiyuan You"
  - "Jie Liu"
  - "Qipeng Zhu"
  - "Kai Yang"
  - "Xingzhong Xu"
  - "Xin Geng"
  - "Xu Yang"
model_name: LMM-R1
model_developer: "东南大学、蚂蚁集团、香港中文大学、复旦大学"
domain:
  - "自然语言处理"
  - "计算机视觉"
  - "多模态大模型"
  - "强化学习"
  - "推理能力增强"
type:
  - "自然语言处理"
  - "计算机视觉"
  - "多模态大模型"
  - "强化学习"
  - "推理能力增强"
  - "Computation and Language"
  - "Artificial Intelligence"
arxiv_url: "https://arxiv.org/abs/2503.07536"
pdf_url: "https://arxiv.org/pdf/2503.07536"
code_link: "https://github.com/TideDra/lmm-r1"
---

# LMM-R1: Empowering 3B LMMs with Strong Reasoning Abilities Through Two-Stage Rule-Based RL

> Enhancing reasoning in Large Multimodal Models (LMMs) faces unique challenges from the complex interplay between visual perception and logical reasoning, particularly in compact 3B-parameter architectures where architectural constraints limit reasoning…

「LMM-R1: Empowering 3B LMMs with Strong Reasoning Abilities Through Two-Stage Rule-Based RL」是 ModelScope 魔搭社区收录的论文，arXiv 2503.07536，作者为 Yingzhe Peng, Gongrui Zhang, Miaosen Zhang et al.，发表于 2025-03-11，属于 自然语言处理、计算机视觉、多模态大模型 领域。

- **ArXiv**: 2503.07536
- **Published**: 2025-03-11
- **Authors**: Yingzhe Peng, Gongrui Zhang, Miaosen Zhang, Zhiyuan You, Jie Liu, Qipeng Zhu, Kai Yang, Xingzhong Xu, Xin Geng, Xu Yang
- **Model**: LMM-R1
- **Developer**: 东南大学、蚂蚁集团、香港中文大学、复旦大学
- **Domain**: 自然语言处理, 计算机视觉, 多模态大模型, 强化学习, 推理能力增强
- **ArXiv URL**: https://arxiv.org/abs/2503.07536
- **PDF**: https://arxiv.org/pdf/2503.07536
- **Code**: https://github.com/TideDra/lmm-r1

Source: https://www.modelscope.cn/papers/125170

---

> LMM-R1：用两阶段规则基础强化学习提升3B参数规模多模态模型的推理能力

## 摘要

本文提出了一种名为LMM-R1的两阶段规则基础强化学习框架，旨在增强大型多模态模型（LMMs）的推理能力。当前3B参数规模的LMM在进行复杂推理时面临两大挑战：数据不足和基础推理能力弱化。为此，LMM-R1通过两个阶段来解决这些问题。第一阶段是基础推理增强（FRE），使用高质量的纯文本数据进行规则基础的强化学习，以提升模型的基础推理能力；第二阶段是多模态泛化训练（MGT），将这些推理能力扩展到各种多模态任务中。实验结果显示，LMM-R1在多个基准测试上取得了显著的性能提升，特别是在复杂的足球游戏任务中表现尤为突出。该方法不仅提高了模型的推理能力，还增强了其视觉理解能力，为小型LMM的应用提供了新的可能性。

## Abstract

Enhancing reasoning in Large Multimodal Models (LMMs) faces unique challenges from the complex interplay between visual perception and logical reasoning, particularly in compact 3B-parameter architectures where architectural constraints limit reasoning capacity and modality alignment. While rule-based reinforcement learning (RL) excels in text-only domains, its multimodal extension confronts two critical barriers: (1) data limitations due to ambiguous answers and scarce complex reasoning examples, and (2) degraded foundational reasoning induced by multimodal pretraining. To address these challenges, we propose \textbf{LMM-R1}, a two-stage framework adapting rule-based RL for multimodal reasoning through \textbf{Foundational Reasoning Enhancement (FRE)} followed by \textbf{Multimodal Generalization Training (MGT)}. The FRE stage first strengthens reasoning abilities using text-only data with rule-based RL, then the MGT stage generalizes these reasoning capabilities to multimodal domains. Experiments on Qwen2.5-VL-Instruct-3B demonstrate that LMM-R1 achieves 4.83\% and 4.5\% average improvements over baselines in multimodal and text-only benchmarks, respectively, with a 3.63\% gain in complex Football Game tasks. These results validate that text-based reasoning enhancement enables effective multimodal generalization, offering a data-efficient paradigm that bypasses costly high-quality multimodal training data.
