EdiTikZ: Scientific Figure Editing from Revision Trajectories

EdiTikZ is a family of open multimodal language models for instruction-guided scientific figure editing and image-to-TikZ reconstruction. The models are trained through multi-task supervised fine-tuning (SFT) on DaEdiTikZ for editing and DaTikZ-V4 for reconstruction. They are subsequently optimized for editing using multi-reward reinforcement learning (RL) with complementary rewards for rendered fidelity and edit application. EdiTikZ achieves state-of-the-art scientific figure editing and reconstruction performance among open models.

Overview

EdiTikZ introduces three main contributions:

1. DaEdiTikZ – a scalable framework for recovering realistic scientific figure revision trajectories from naturally occurring collections of related figures. DaEdiTikZ contains 391K plausible edit pairs and 781K directed editing instances mined from arXiv, GitHub, and TeX Stack Exchange.

2. Editing-specific post-training – a training pipeline that combines joint editing and reconstruction during SFT with subsequent editing-focused reinforcement learning. The RL stage uses GDPO with complementary rewards for rendered fidelity and edit application.

3. Strong open models – compact Qwen3.5-based models with 4B and 9B parameters that outperform previous open approaches and remain competitive with substantially larger proprietary systems.

Models

• EdiTikZ-4B
• EdiTikZ-4B-RL
• EdiTikZ-9B
• EdiTikZ-9B-RL

All four models support instruction-guided editing and image-to-TikZ reconstruction.

Datasets

• DaEdiTikZ – redistributable training subset with paired figures, TikZ programs, and bidirectional editing instructions.
• DaEdiTikZ-Bench – human-curated benchmark containing 345 figure pairs and 690 directed editing instances.
• DaTikZ-V4 – large-scale dataset used for image-to-TikZ reconstruction training.

Code release

The complete source code for data construction, edit-instruction generation, training, and evaluation will be released soon. Models and redistributable datasets are already available through our Hugging Face collection.

Citation

If EdiTikZ is useful for your research or applications, please cite:

@misc{greisinger2026editikzscientificfigureediting,
    title={EdiTikZ: Scientific Figure Editing from Revision Trajectories}, 
    author={Christian Greisinger and Zhixue Zhao and Steffen Eger},
    year={2026},
    eprint={2609.01409},
    archivePrefix={arXiv},
    primaryClass={cs.AI},
    url={https://arxiv.org/abs/2609.01409}, 
}