AMVICC - MMVP Pair Accuracy: leaderboard
Vision-language evaluation from AMVICC (Basappa, Goel et al., 2026) on the 300 original MMVP questions, graded by GPT-4, where a pair of questions about two near-identical images counts only if both are answered correctly. The score is the unweighted mean of pair accuracy (%) across MMVP's nine visual-pattern categories; higher is better.
Metric: Pair Accuracy (%). Source: arxiv.org. Saturation forecast: Not forecast. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Llama 3.2 90B Vision Instruct | 84.18 | #1757 |
| 2 | Llama 4 Maverick | 77.41 | #1105 |
| 3 | GPT-4o | 70.95 | #741 |
| 4 | Gemini 2.5 Pro | 70.42 | #262 |
| 5 | Claude Opus 4.1 | 69.32 | #268 |
| 6 | Llama 4 Scout | 68.48 | #1787 |
| 7 | Qwen 2.5 VL 72B Instruct | 61.6 | #845 |
| 8 | pixtral-large-2411 | 57.55 | #1314 |
| 9 | Gemma 3 27B | 53.81 | #1638 |
| 10 | Claude Sonnet 4 | 52.12 | #397 |
| 11 | Grok 4 | 39.67 | #306 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=amvicc-mmvp-pair-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-12.