SciDiagramEdit for Figure Editing

SciDiagramEdit uses natural paper revisions, editable vector source, and skill evolution to study instruction-driven editing of scientific figures.

TL;DR — SciDiagramEdit is a benchmark and skill-evolution framework for natural-language editing of scientific figures. It mines before/after figure pairs from arXiv version histories, grounds them in authors' revision intent, and edits the figure's editable vector source so users can inspect and co-edit primitives. The abstract reports improved edit accuracy on a held-out validation set but provides no numbers.

Problem and contribution

SciDiagramEdit addresses a specific but common research workflow: revising figures as a manuscript changes. The abstract grounds the motivation directly: "Editing the figures in a research paper is a routine and time-consuming part of everyday research practice." It gives concrete examples of edits authors make: they "relabel components, rearrange panels, and restyle visuals."

The contribution is a benchmark and learning framework that turns natural paper revisions into training and evaluation signal for instruction-driven scientific figure editing. The abstract's central formulation is: "we present SciDiagramEdit, a benchmark and skill-evolution framework" that "learns from natural paper revisions."

Full analysis, extracted claims, numerical results, entity graph, FAQ, related work, applications, and BibTeX are available via x402 micropayment.