Reproduce “Math-PUMA: Progressive Upward Multimodal Alignment to Enhance Mathematical Reasoning” 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
Wenwen Zhuang, Xin Huang, Xiantao Zhang, Jin Zeng
Math-PUMA introduces a three-stage progressive upward multimodal alignment framework to enhance MLLMs' mathematical reasoning capabilities, particularly with diagrams. The approach effectively reduces the performance gap between text-rich and vision-rich mathematical problems, achieving state-of-the-art results on benchmarks like MATHVERSE, outperforming previous open-source models by approximately 10% accuracy.
Key insights
- MLLM performance in mathematical reasoning can be systematically improved by explicitly aligning multimodal understanding.
- Leveraging text-rich inputs to supervise vision-rich inputs via KL divergence effectively reduces performance disparity across modalities.
- A structured, progressive training approach (text foundation, followed by alignment, then instruction tuning) is effective for building robust multimodal reasoning capabilities.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.