Advancing Multimodal Reasoning: From Optimized Cold Start to Staged Reinforcement Learning

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Chen, Shuang, Guo, Yue, Su, Zhaochen, Li, Yafu, Wu, Yulun, Chen, Jiacheng, Chen, Jiayu, Wang, Weijie, Qu, Xiaoye, Cheng, Yu
Format: Preprint
Veröffentlicht: 2025
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866912853778235392
author Chen, Shuang
Guo, Yue
Su, Zhaochen
Li, Yafu
Wu, Yulun
Chen, Jiacheng
Chen, Jiayu
Wang, Weijie
Qu, Xiaoye
Cheng, Yu
author_facet Chen, Shuang
Guo, Yue
Su, Zhaochen
Li, Yafu
Wu, Yulun
Chen, Jiacheng
Chen, Jiayu
Wang, Weijie
Qu, Xiaoye
Cheng, Yu
contents Inspired by the remarkable reasoning capabilities of Deepseek-R1 in complex textual tasks, many works attempt to incentivize similar capabilities in Multimodal Large Language Models (MLLMs) by directly applying reinforcement learning (RL). However, they still struggle to activate complex reasoning. In this paper, rather than examining multimodal RL in isolation, we delve into current training pipelines and identify three crucial phenomena: 1) Effective cold start initialization is critical for enhancing MLLM reasoning. Intriguingly, we find that initializing with carefully selected text data alone can lead to performance surpassing many recent multimodal reasoning models, even before multimodal RL. 2) Standard GRPO applied to multimodal RL suffers from gradient stagnation, which degrades training stability and performance. 3) Subsequent text-only RL training, following the multimodal RL phase, further enhances multimodal reasoning. This staged training approach effectively balances perceptual grounding and cognitive reasoning development. By incorporating the above insights and addressing multimodal RL issues, we introduce ReVisual-R1, achieving a new state-of-the-art among open-source 7B MLLMs on challenging benchmarks including MathVerse, MathVision, WeMath, LogicVista, DynaMath, and challenging AIME2024 and AIME2025.
format Preprint
id arxiv_https___arxiv_org_abs_2506_04207
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Advancing Multimodal Reasoning: From Optimized Cold Start to Staged Reinforcement Learning
Chen, Shuang
Guo, Yue
Su, Zhaochen
Li, Yafu
Wu, Yulun
Chen, Jiacheng
Chen, Jiayu
Wang, Weijie
Qu, Xiaoye
Cheng, Yu
Machine Learning
Artificial Intelligence
Computation and Language
Computer Vision and Pattern Recognition
Inspired by the remarkable reasoning capabilities of Deepseek-R1 in complex textual tasks, many works attempt to incentivize similar capabilities in Multimodal Large Language Models (MLLMs) by directly applying reinforcement learning (RL). However, they still struggle to activate complex reasoning. In this paper, rather than examining multimodal RL in isolation, we delve into current training pipelines and identify three crucial phenomena: 1) Effective cold start initialization is critical for enhancing MLLM reasoning. Intriguingly, we find that initializing with carefully selected text data alone can lead to performance surpassing many recent multimodal reasoning models, even before multimodal RL. 2) Standard GRPO applied to multimodal RL suffers from gradient stagnation, which degrades training stability and performance. 3) Subsequent text-only RL training, following the multimodal RL phase, further enhances multimodal reasoning. This staged training approach effectively balances perceptual grounding and cognitive reasoning development. By incorporating the above insights and addressing multimodal RL issues, we introduce ReVisual-R1, achieving a new state-of-the-art among open-source 7B MLLMs on challenging benchmarks including MathVerse, MathVision, WeMath, LogicVista, DynaMath, and challenging AIME2024 and AIME2025.
title Advancing Multimodal Reasoning: From Optimized Cold Start to Staged Reinforcement Learning
topic Machine Learning
Artificial Intelligence
Computation and Language
Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2506.04207