MIOH - Position: leaderboard

Metric: Accuracy (%; position questions about object positions 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 July 2027. 29 models tracked.

Top models

#ModelScore
1GPT-567.1
2Gemini 2.5 Pro66.6
3Qwen 2 VL 7B61.1
4MiniCPM-V-2.659
5Qwen 2.5 VL 7B52.6
6Qwen 2 VL 2B45.3
7InternVL3.5-8B40.8
8Phi-4 Multimodal Instruct35.8

Interactive version: theaggregate.ai/benchmark?slug=mioh-position · How It Works · Data refreshed daily, snapshot 2026-09-29.