MIOH - Counting: leaderboard
The counting task of MIOH (2026), a multi-image object hallucination benchmark for multimodal LLMs: questions about how many objects there are across several images, averaged over easy, hard-negative, hard-positive and eight-image conditions. Accuracy in percent; higher is better.
Metric: Accuracy (%; counting questions about how many objects there are across several images; multi-image questions built from COCO-ReM, PACO and Visual Genome scenes in three reasoning patterns (comprehensive, comparative, selective), averaged over the easy, hard-negative, hard-positive and eight-image conditions). Source: arxiv.org. Saturation forecast: Around February 2027. 29 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 57.5 | #266 |
| 2 | GPT-5 | 49.1 | #167 |
| 3 | MiniCPM-V-2.6 | 28.3 | #3063 |
| 4 | Qwen 2.5 VL 7B | 28 | #1787 |
| 5 | Qwen 2 VL 7B | 26 | #2261 |
| 6 | Phi-4 Multimodal Instruct | 24 | #3882 |
| 7 | Qwen 2 VL 2B | 21.2 | #6284 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=mioh-counting · How It Works · Data refreshed daily, snapshot 2026-11-13.