EdiTikZ: Scientific Figure Editing from Revision Trajectories
EdiTikZ leverages DaEdiTikZ dataset and RL to achieve scientific figure editing; 9B model outperforms GPT-5.6-Sol.
Key Findings
Methodology
The study introduces DaEdiTikZ, a dataset with 391K TikZ edit pairs and 781K edit instructions, and trains Qwen3.5 models via joint learning and RL. The models integrate figure reconstruction and editing while optimizing fidelity and edit application through rendering feedback.
Key Results
- Result 1: EdiTikZ-9B surpasses all baselines in automatic evaluation, scoring higher than GPT-5.6-Sol in editing quality.
- Result 2: In human evaluation with 4,320 ratings, the 9B model performs on par with Gemini-3.1-Pro.
- Result 3: Under OOD testing, the 9B model remains competitive within the 2K sequence-length regime.
Significance
This work addresses a gap in scientific figure editing by providing a new supervision source for multimodal models. The release of DaEdiTikZ and EdiTikZ models advances automation in academic publishing and research collaboration.
Technical Contribution
Key contributions include: 1) a large-scale dataset derived from natural revision trajectories; 2) a joint learning framework for figure reconstruction and editing; 3) reinforcement learning with rendering feedback and edit-specific rewards.
Novelty
This is the first work to use scientific revision trajectories as a supervision source, offering greater realism and scalability compared to synthetic methods.
Limitations
- Limitation 1: Limited ability to handle complex geometric edits, leading to rendering errors.
- Limitation 2: Dataset reliance on TikZ may hinder generalization to other graphic languages.
- Limitation 3: Edit instruction generation depends on VLMs, potentially introducing biases.
Future Work
Future directions include extending to other graphic languages, improving robustness for complex edits, and exploring more efficient training methods.
AI Executive Summary
Scientific figure editing is crucial for academic publishing, yet existing methods rely on expensive proprietary systems or synthetic data, falling short of practical needs.
EdiTikZ addresses this challenge by introducing the DaEdiTikZ dataset and reinforcement learning techniques. The dataset comprises 391K TikZ edit pairs and 781K edit instructions, extracted from arXiv, GitHub, and TeX SE. Using this dataset, the authors trained two Qwen3.5 models (4B and 9B) through joint learning and rendering-feedback optimization, achieving high-quality figure editing.
Experimental results demonstrate that the EdiTikZ-9B model outperforms all baselines in both automatic and human evaluations, with strong generalization under OOD scenarios. This study provides a novel solution for scientific figure editing and lays the groundwork for further advancements in this domain.
Deep Analysis
Background
Scientific figures are essential for academic communication, with TikZ being a popular choice due to its precision and interpretability. However, its complex syntax poses challenges for both humans and models. Prior work has focused on generating TikZ from text or images, with limited attention to editing existing figures.
Core Problem
Existing methods rely on synthetic data or proprietary systems, failing to capture real-world editing needs. Editing scientific figures requires preserving original content while applying precise modifications, posing significant challenges for models.
Innovation
Key innovations include: 1) leveraging scientific revision trajectories to build a realistic editing dataset; 2) proposing a joint learning framework for reconstruction and editing; 3) optimizing editing quality through rendering feedback.
Methodology
- �� Dataset construction: Extract TikZ edit pairs from arXiv and other platforms, generating edit instructions.
- �� Model training: Train Qwen3.5 models via multi-task learning, combining reconstruction and editing tasks.
- �� Reinforcement learning: Optimize edit application and visual fidelity using rendering feedback.
Experiments
Experiments use the DaEdiTikZ dataset, with baselines including GPT-5.6-Sol and Gemini-3.1-Pro. Evaluation metrics cover edit application, source preservation, and visual quality, with OOD testing conducted.
Results
EdiTikZ-9B outperforms all baselines in automatic evaluation, achieves parity with Gemini-3.1-Pro in human ratings, and demonstrates robustness in OOD scenarios.
Applications
The method can be applied to academic publishing, research collaboration, and education, enabling efficient scientific figure editing.
Limitations & Outlook
The model struggles with complex geometric edits, and its reliance on TikZ may limit generalization. Future work should address these issues.
Plain Language Accessible to non-experts
Imagine building a LEGO model. After finishing, your friend suggests changes like swapping colors or adding new pieces. EdiTikZ acts like a smart assistant that understands these suggestions and quickly makes the adjustments for you. Its uniqueness lies in its ability to not only 'see' your LEGO model but also modify it exactly as requested.
ELI14 Explained like you're 14
Think of playing Minecraft and building a castle. Your friend says, 'Make the tower blue and add a flag!' You might take time to do it, but EdiTikZ is like a super helper that instantly makes these changes for you, even better than you could! Cool, right?
Glossary
TikZ
A LaTeX-based graphics language widely used for academic figure creation.
Core language for generating and editing scientific figures.
VLM
Vision-language model combining image and text understanding for multimodal tasks.
Used to generate edit instructions and evaluate editing quality.
DaEdiTikZ
A large-scale dataset with 391K TikZ edit pairs for training and evaluating scientific figure editing models.
Serves as the training data source for EdiTikZ.
Reinforcement Learning
A machine learning approach optimizing behavior through reward signals.
Used to enhance edit application and rendering fidelity.
OOD Testing
Evaluating model performance on unseen data distributions.
Validates EdiTikZ's generalization capabilities.
Open Questions Unanswered questions from this research
- 1 How to generalize to other graphic languages for broader applicability?
- 2 How to improve the model's ability to handle complex geometric edits?
- 3 How to reduce biases in edit instruction generation?
Applications
Immediate Applications
Academic Figure Optimization
Helps researchers efficiently adjust figures for academic publishing.
Educational Tools
Assists students in learning TikZ, reducing the learning curve.
Long-term Vision
Cross-Language Figure Editing
Supports multiple graphic languages, advancing scientific collaboration automation.
Abstract
Vision-language models (VLMs) have shown strong performance in generating scientific figures from text or images. However, producing publication-ready figures requires iterative refinement, making scientific figure editing an important yet largely unexplored task. Existing approaches rely on costly proprietary agentic systems, focus primarily on evaluation, or construct training supervision from synthetically generated edits. Instead, we leverage naturally occurring scientific revision and development trajectories as a scalable source of supervision. To this end, we introduce DaEdiTikZ, the first large-scale dataset of revision-derived scientific figure edits, constructed by mining 391K plausible TikZ edit pairs from arXiv, GitHub, and TeX SE and inferring 781K directed edit instructions with a VLM conditioned on rendered figures and TikZ code. We further introduce DaEdiTikZ-Bench, a human-refined benchmark with 790 instances, and train two compact Qwen3.5-based EdiTikZ models (4B and 9B) by jointly learning reconstruction and editing, followed by reinforcement learning (RL) with complementary rewards for rendered fidelity and edit application. Automatic evaluation places our 9B model above all tested baselines, while human evaluation with 9 annotators and 4,320 ratings places it above GPT-5.6-Sol and on par with Gemini-3.1-Pro. Under severe out-of-distribution shifts, it remains competitive with GPT-5.6-Sol near its 2K training sequence-length regime. Models and datasets will be released.