Research map / Large language models / Instruction and fine-tuning
TikZilla: Scaling Text-to-TikZ with High-Quality Data and Reinforcement Learning
Christian Greisinger, Steffen Eger
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add RL with domain reward after SFT (ablation) | Qwen2.5-3B · DaTikZ-V4 test | AVG: 0.324 -> 0.385 | Improved |
| add SFT before RL (ablation) | Qwen2.5-3B · DaTikZ-V4 test | AVG: 0.269 -> 0.385 | Improved |
| RL directly without SFT (ablation) | Qwen3-8B · DaTikZ-V4 test | AVG: 0.251 -> 0.357 | Improved |
| add RL (ablation) | Qwen models · DaTikZ-V4 test | compilation rate: TikZilla-3B 89% -> 98% | Improved |
| use VLM descriptions instead of raw captions at inference (ablation) | GPT-4o · DaTikZ-V4 test | AVG: 0.270 -> 0.315 | Improved |
| use VLM descriptions instead of captions for SFT (ablation) | Qwen2.5-3B · DaTikZ-V4 test | AVG: 0.279 -> 0.289 | Improved |
| mix captions with descriptions in SFT (ablation) | Qwen2.5-3B · DaTikZ-V4 test | AVG: 0.324 -> 0.313 | Hurt |
| add LLM debugging of uncompilable code to training data (ablation) | Qwen2.5-3B SFT · DaTikZ-V4 test | AVG: 0.288 -> 0.324 | Improved |
| use domain reward R_Sim instead of DreamSim (ablation) | SFT+RL Qwen2.5-3B · DaTikZ-V4 test | AVG: 0.474 -> 0.498 | Improved |
| retrain reward encoder on DaTikZ-V4 instead of V3 (ablation) | reward model · DaTikZ-V4 test | AVG: 0.487 -> 0.498 | Improved |
| increase SFT data from 0 to 25% then to 100% (ablation) | Qwen2.5-3B · DaTikZ-V4 test | AVG | Improved |
| evaluate TikZilla out of distribution on SPIQA | TikZilla-3B-RL · SPIQA | DreamSim: GPT-5 0.432 -> 0.637 | Improved |
Closest papers by meaning
- AutomaTikZ: Text-Guided Synthesis of Scientific Vector Graphics with TikZ (ICLR 2024 · similarity 0.93)
- DeTikZify: Synthesizing Graphics Programs for Scientific Figures and Sketches with TikZ (NeurIPS 2024 · similarity 0.91)
- ScImage: How good are multimodal large language models at scientific text-to-image generation? (ICLR 2025 · similarity 0.88)
- Breaking the SFT Plateau: Multimodal Structured Reinforcement Learning for Chart-to-Code Generation (ICLR 2026 · similarity 0.86)
- Text-Printed Image: Bridging the Image-Text Modality Gap for Text-centric Training of Large Vision-Language Models (CVPR 2026 · similarity 0.86)
- SketchVL: Policy Optimization via Fine-Grained Credit Assignment for Chart Understanding and More (CVPR 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
