AMVICC - MMVP Pair Accuracy - Positional and Relational Context: leaderboard
AMVICC (Basappa, Goel et al., 2026) asks vision-language models the original MMVP questions about pairs of nearly identical images, grading free-form answers with GPT-4 against the authors' rubric. This board is pair accuracy in percent on the 16 image pairs of MMVP's Positional and Relational Context category, where a pair counts only when both of its questions are answered correctly; higher is better.
Metric: Pair accuracy (%; the 16 MMVP image pairs in the Positional and Relational Context category, a pair credited only when both of its questions are answered correctly; free-form answers graded by GPT-4 against the authors' rubric). Source: arxiv.org. Saturation forecast: Around 2029. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Llama 3.2 90B Vision Instruct | 81.25 | #1749 |
| 2 | Claude Opus 4.1 | 62.5 | #272 |
| 3 | GPT-4o | 56.25 | #741 |
| 4 | Gemini 2.5 Pro | 56.25 | #266 |
| 5 | Llama 4 Maverick | 56.25 | #1099 |
| 6 | Claude Sonnet 4 | 50 | #400 |
| 7 | Qwen 2.5 VL 72B Instruct | 50 | #846 |
| 8 | Llama 4 Scout | 50 | #1776 |
| 9 | Gemma 3 27B | 43.75 | #1631 |
| 10 | pixtral-large-2411 | 43.75 | #1309 |
| 11 | Grok 4 | 25 | #307 |
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-positional-and-relational-context · How It Works · Data refreshed daily, snapshot 2026-11-13.