Reproduce “Synthesizing Multimodal Geometry Datasets from Scratch and Enabling Visual Alignment via Plotting Code” 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
Haobo Lin, Tianyi Bai, Chen Chen, Jiajun Zhang, Bohan Zeng, Wentao Zhang, Binhang Yuan
GeoCode trains geometry-solving models to predict plotting code that represents points, segments, circles, and annotations, then solve from that structured scene description. In a matched Qwen3-VL-7B-Instruct comparison on Test-mini, code-plus-debias supervision achieves 26.49% answer accuracy, versus 20.52% for caption-plus-debias. This supports the diagram-recovery account but does not establish causation or generalize beyond the tested scope.
Key insights
- Leveraging the unique property of geometry problems—being both symbolically solvable and numerically verifiable—allows for the generation of large-scale, high-quality multimodal data with guaranteed correctness.
- Explicit supervision through plotting code fosters stronger visual-symbolic alignment and geometric perception, outperforming implicit natural-language-based alignment by transforming visual understanding into a structured prediction task.
- Models trained on synthetically generated, structurally rich data, combined with explicit alignment, learn transferable geometric representations that generalize well to challenging, out-of-distribution problems.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.