VisEditBench - Correctness Repair: leaderboard
VisEditBench (arXiv 2608.10408v1, Table 2): zero-shot visualization code editing pass rate on the 529 tasks with the Correctness Repair editing intent. A pass needs executable code, task accuracy of at least 4.5 of 5, readability and visual quality of 4.0 and visual similarity of 90 under a GPT-4o evaluator (Gemini 2.5 Pro for GPT-4o outputs). Percent; higher is better.
Metric: Pass rate (%). Source: arxiv.org. Saturation forecast: Around May 2028. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 70.34 | #94 |
| 2 | GPT-5 | 61.6 | #105 |
| 3 | Claude Sonnet 4.5 | 60.83 | #148 |
| 4 | GPT-4o | 56.08 | #357 |
| 5 | GPT-5 Mini | 55.51 | #193 |
| 6 | Gemma 3 27B | 49.24 | #606 |
| 7 | Pixtral-12B | 44.68 | #805 |
| 8 | Gemma 3 12B | 42.97 | #674 |
| 9 | Gemma 3 4B | 30.79 | #1088 |
| 10 | Gemini 3 Flash | 17.96 | #103 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=viseditbench-correctness-repair · How It Works · Data refreshed daily, snapshot 2026-11-05.