Editing figures in research papers is a routine yet time-consuming part of everyday research practice, where authors relabel components, rearrange panels, and restyle visuals during manuscript revisions. Automating this editing workflow under natural language instructions is challenging due to the dense nature of scientific figures, which comprise heterogeneous visual elements like schematics, plots, photos, captions, and arrows, all assembled under a tight visual grammar to support specific arguments.
To tackle this, we introduce SciDiagramEdit, a benchmark and skill-evolution framework that learns from natural paper revisions and operates on the figure's editable vector source, allowing users to inspect and co-edit individual primitives alongside the agent.
Our benchmark mines before/after figure pairs from arXiv version histories, grounded in the authors' own revision intent. To accommodate the diversity of editing instructions, we adopt agentic learning via skill evolution: an agentic proposer continually refines the agent's skill specification from execution traces over multiple epochs. The resulting skill progressively improves edit accuracy on a held-out validation set, providing evidence that natural paper revisions serve as an effective training signal for instruction-driven figure editing.
Blogger's Review: SciDiagramEdit showcases the potential of machine learning in automating scientific figure editing through the use of natural language instructions derived from paper revisions. Its innovative skill evolution framework opens new avenues for graphic automation, making it a noteworthy area for researchers to explore.