The first large dataset of real scientific figure edits comes from arXiv and GitHub.

Vision-language models can make scientific figures, but scientists must refine them step by step to get them ready for publication. Editing these figures is an important problem that researchers have not explored much. Usually, people rely on expensive proprietary tools or fake, computer-generated edits. Instead, the authors collected real revision histories from human scientists. They built a large dataset from actual figure changes and taught smaller models how to update the diagram code. Their method works by training the model to reconstruct figures and edit them together, followed by reinforcement learning that rewards visually accurate images and correct edits. In tests, their nine-billion-parameter model beat all tested baselines. Human judges ranked it above GPT-5.6-Sol and tied with Gemini-3.1-Pro, showing small models can edit scientific figures using real revision data.

Why it matters

Scientists and researchers spend a lot of time tweaking figures for papers. This tool helps them edit figures more easily using AI, saving time and making publishing smoother.

Questions answered