MinCU: leaderboard
MinCU (Mu et al., 2026) tests grounded minimal-change understanding: given two near-identical images that differ by one edit, a multimodal model must describe the change, box the changed region and name the changed entity. Grounded F1 in percent counts a prediction only when the box reaches IoU 0.5, the description reaches BERTScore 0.85 and the entity label is correct, over 2,000 image pairs scored zero-shot; higher is better.
Metric: Grounded F1 (%; overall on 2,000 image pairs, 500 per change type; a true positive needs the predicted box to reach IoU 0.5 with the changed region, the change description BERTScore 0.85 against the reference and the correct changed-entity label; zero-shot). Source: arxiv.org. Saturation forecast: Around June 2027. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 42.89 | #91 |
| 2 | Seed 1.8 | 42.43 | #239 |
| 3 | Gemini 2.5 Pro | 31.92 | #262 |
| 4 | GPT-5.4 | 19.58 | #121 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=mincu · How It Works · Data refreshed daily, snapshot 2026-11-12.