Diagram-MMU - Code Parsing: leaderboard
The diagram-to-code track of Diagram-MMU (arXiv 2608.12262): a model reads a scientific diagram and writes the TikZ that reproduces it, scored by re-rendering the output and comparing it back to the source figure; the published number is the track’s "All" column across its figure types.
Metric: All Score (%). Source: vi-ocean.github.io. Saturation forecast: Not forecast. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Kimi K2.5 | 56.97 | #144 |
| 2 | Qwen 3 VL 235B A22B | 55.98 | #238 |
| 3 | Qwen 3.5 397B A17B | 52.72 | #146 |
| 4 | Claude Opus 4.6 | 52.41 | #63 |
| 5 | GPT-5.2 | 51.88 | #110 |
| 6 | Seed 2.0 Pro | 50.89 | #118 |
| 7 | Gemini 3.1 Pro (Preview) | 48.38 | #58 |
| 8 | InternVL3-38B | 41.13 | #405 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=diagram-mmu-code-parsing · How It Works · Data refreshed daily, snapshot 2026-10-20.