Reproduce “Vision-R1: Incentivizing Reasoning Capability in Multimodal Large Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Wenxuan Huang, Bohan Jia, Zijie Zhai, Shaosheng Cao, Zheyu Ye, Fei Zhao, Zhe Xu, Xu Tang and 2 more
VISION-R1 introduces a framework that incentivizes complex reasoning capabilities in Multimodal Large Language Models (MLLMs) by combining cold-start initialization with Reinforcement Learning. The method, featuring "Modality Bridging" for high-quality data generation and Progressive Thinking Suppression Training, enabled Vision-R1-7B to achieve 73.5% accuracy on the MathVista benchmark, representing an average 10% improvement over its base model while exhibiting human-like cognitive processes.
Key insights
- High-quality, human-like CoT data, constructed without manual annotation via "Modality Bridging," is crucial for effectively initializing MLLMs to learn complex reasoning capabilities.
- A hybrid approach combining cold-start initialization with R1-like RL training proves more effective for incentivizing complex multimodal reasoning than direct RL alone, especially given data and computational limitations.
- Progressive Thinking Suppression Training (PTST) successfully manages the "Overthinking Optimization Problem" by progressively relaxing reasoning length constraints, allowing models to leverage extended CoT for improved accuracy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.