MindEdit-Bench: leaderboard
Object-level spatial reasoning from three smartphone photos of real indoor scenes: 1,003 multiple-choice questions over six tasks, from cross-view object relations to what changes when an object is hypothetically moved. Pooled accuracy (%) over all questions, mean of three greedy answer-only runs, higher is better (MindEdit-Bench, Yu et al., ZODA, Zhejiang University and the University of Tokyo, 2026).
Metric: Pooled Accuracy (%). Source: arxiv.org. Saturation forecast: Around March 2028. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 34.6 | #91 |
| 2 | Qwen 3.6 Plus | 31.1 | #217 |
| 3 | Seed 2.0 Pro | 25.4 | #209 |
| 4 | Claude Opus 4.6 | 22.2 | #100 |
| 5 | GPT-5.4 | 17.3 | #121 |
| 6 | Qwen 3.5 35B A3B | 16.9 | #480 |
| 7 | Kimi K2.5 | 15.9 | #252 |
| 8 | InternVL3-8B | 12.5 | #1695 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=mindedit-bench · How It Works · Data refreshed daily, snapshot 2026-11-12.