Reproduce “Figures as Programs: Recursive Generation of Editable Scientific Figures” 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
Yepeng Liu, Dasen Dai, Chengzhi Liu, Yiren Song, Hai Ci, Yu Zhang, Qi Zhang, Mike Zheng Shou and 2 more
FigTree, a multi-agent system from UC Santa Barbara and collaborators, automates the generation of editable scientific methodology figures by constructing recursive SVG programs. It achieves an overall quality score of 59.59% in VLM-based evaluation, outperforming existing raster-based methods by 9.59 points, and demonstrates improved editability and semantic accuracy.
Key insights
- Formulating figure generation as recursive SVG program construction, managed by a multi-agent hierarchy, effectively addresses the complexity and reliability challenges of creating intricate vector graphics.
- A dual-space review mechanism, where a critic inspects both the SVG code and its rendered image, enables precise identification of visual defects and accurate tracing to specific program statements for targeted, localized repairs.
- Intrinsic editability offered by vector graphics (SVG) significantly enhances the efficiency and control of iterative refinement processes compared to raster-based image generation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.