TikZilla: Scaling Text-to-TikZ with High-Quality Data and Reinforcement Learning

Researchers have developed TikZilla, a new AI model that generates high-quality scientific figures from text descriptions using TikZ code. The model uses a dataset four times larger than previous versions and combines supervised learning with reinforcement learning to achieve performance matching GPT-5 while using much smaller model sizes.

• →TikZilla family includes 3B and 8B parameter models that outperform GPT-4o and match GPT-5 in generating scientific figures from text.
• →The new DaTikZ-V4 dataset is four times larger and significantly higher quality than previous versions for training text-to-figure models.
• →The two-stage training approach combines supervised fine-tuning with reinforcement learning using image encoder rewards for better visual accuracy.
• →Human evaluations with over 1,000 judgments show 1.5-2 point improvements over base models on a 5-point scale.
• →The open-source models address common issues like looping, irrelevant content, and incorrect spatial relations in generated figures.