VS-Bench (VisualSwap) - MathVerse: leaderboard
VS-Bench (Shi et al., 2026) probe accuracy on the 200 image pairs built from vision-dominant plane and solid geometry problems from MathVerse: after reasoning about an edited image, the model is shown the original image mid-response with a self-reflection prompt and must answer for the image now in front of it. Accuracy in percent; higher is better.
Metric: Probe accuracy (%; 200 image pairs sourced from MathVerse: answer on the swapped image after a self-reflection prompt). Source: arxiv.org. Saturation forecast: Not forecast. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 VL 32B Instruct | 70.5 | #582 |
| 2 | Qwen 3 VL 235B A22B Instruct | 63.5 | #521 |
| 3 | Qwen 3 VL 30B A3B Instruct | 55.5 | #870 |
| 4 | Qwen 3 VL 8B Instruct | 44 | #982 |
| 5 | Qwen 2.5 VL 7B Instruct | 36 | #1842 |
| 6 | Qwen 3 VL 8B (Thinking) | 29.5 | #887 |
| 7 | Qwen 3 VL 32B (Thinking) | 24 | #557 |
| 8 | Qwen 3 VL 235B A22B (Thinking) | 22.5 | #451 |
| 9 | Qwen 3 VL 30B A3B (Thinking) | 10 | #710 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=vs-bench-visualswap-mathverse · How It Works · Data refreshed daily, snapshot 2026-11-12.