EdiTikZ: Scientific Figure Editing from Revision Trajectories
A paper titled 'EdiTikZ: Scientific Figure Editing from Revision Trajectories' was published on arXiv on 2026-09-01. It introduces DaEdiTikZ, a dataset of 391K TikZ edit pairs mined from arXiv, GitHub, and TeX SE, with 781K directed edit instructions inferred by a VLM. It also introduces DaEdiTikZ-Bench, a human-refined benchmark with 790 instances, and trains two Qwen3.5-based EdiTikZ models (4B and 9B) using reconstruction and editing learning followed by reinforcement learning with rewards for rendered fidelity and edit application.
The paper presents EdiTikZ, a method for scientific figure editing using revision trajectories. It leverages naturally occurring edits to create a large-scale dataset and benchmark, and trains compact models that can edit TikZ figures based on instructions.
The approach uses a VLM to infer edit instructions from rendered figures and TikZ code, and employs reinforcement learning with complementary rewards for rendered fidelity and edit application. The models are based on Qwen3.5 and come in 4B and 9B sizes.
This work addresses a gap in scientific figure editing, which is important for producing publication-ready figures. It offers a scalable alternative to costly proprietary agentic systems by using naturally occurring revision data.
The dataset and models could enable more efficient scientific figure editing, reducing reliance on expensive proprietary systems and potentially improving productivity in academic and research settings.
Potential next signals include further evaluation of the 9B model, release of the dataset and benchmark, and possible applications in scientific writing tools.