Figures as Programs: Recursive Generation of Editable Scientific Figures
A multi-agent system called FigTree automatically transforms a scientific paper into a structured vector figure by grounding content in the source paper, decomposing the figure into a hierarchy of local regions, generating each region as a short SVG program, and assembling the fragments. A render-critic refinement loop inspects the rendered figure and its underlying program to trace visual defects to specific statements and repair them. The system is evaluated on figure quality and editability.
FigTree is a multi-agent system that generates editable scientific figures as recursive SVG programs from a source paper. It grounds figure content in the paper, decomposes the figure into local regions, generates each region as an SVG program, and assembles them. A render-critic loop refines the figure by tracing visual defects to specific program statements. Evaluations show high-quality and editable figures.
The approach treats figure generation as program synthesis, enabling precise edits by modifying SVG code rather than raster pixels. The render-critic loop provides a feedback mechanism that links visual output to code, allowing targeted repairs. This recursive decomposition into local regions likely improves scalability and coherence.
Automating scientific figure creation could reduce time spent on manual illustration in research and publishing. Editable vector outputs are more suitable for iterative refinement and collaboration than raster images. This may lower barriers for researchers without design skills.
Potential value for academic publishers, research institutions, and software tools targeting scientific communication. Could be offered as a service or feature in document preparation platforms.
Next signals include adoption by preprint authors, integration into authoring tools, and extension to other diagram types. Further research may improve figure fidelity to complex scientific content and reduce need for human editing.