Edit2TikZ - Restoration Score: leaderboard
Edit2TikZ (arXiv 2608.13441v1, Table 3): 1,548 scientific-figure edits where a multimodal model gets a rendered figure and an instruction and must write compilable TikZ for the revised figure. Restoration Score is the judged preservation of non-target text, objects, relations and layout, 0-100, zero for render failures; GPT-5.6-Terra judges. Higher is better.
Metric: RS (0-100; non-target preservation, zero for render failures). Source: arxiv.org. Saturation forecast: Around February 2027. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 74.51 | #64 |
| 2 | GPT-5.6 Sol | 63.24 | #22 |
| 3 | GPT-5.6 Terra | 62.29 | #40 |
| 4 | Claude Opus 4.8 | 62.09 | #39 |
| 5 | GPT-5.6 Luna | 56.26 | #79 |
| 6 | Claude Sonnet 4.6 | 53.85 | #94 |
| 7 | Qwen 3.7 Plus | 49.93 | #98 |
| 8 | Qwen 3.6 27B | 48.57 | #182 |
| 9 | Seed 2.1 Pro | 47.94 | #73 |
| 10 | Qwen 3.5 9B | 32.35 | #391 |
| 11 | Qwen 3.5 4B | 26.1 | #514 |
| 12 | gemma-4-E4B-it | 24.64 | #577 |
| 13 | Qwen 3.5 2B | 6.09 | #1114 |
| 14 | Gemma 4 E2B (IT) | 3.17 | #779 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=edit2tikz-restoration-score · How It Works · Data refreshed daily, snapshot 2026-11-05.