CoSimUE: leaderboard
Object-hallucination benchmark for vision-language models (CoSimUE, He et al., Sichuan University, 2026): the same yes/no question is asked about an MSCOCO image and about a FLUX-edited copy whose correct answer is flipped, either by an edit that adds or swaps objects that usually co-occur with the scene (co-occurrence) or by small random edits that create look-alike objects (similarity). The score is pair accuracy in percent over 512 question pairs, where a pair counts only when both answers are right; zero-shot at temperature 0, and higher is better.
Metric: Pair accuracy (%; both the original and the edited image answered correctly, 512 yes/no question pairs). Source: arxiv.org. Saturation forecast: Estimated already saturated. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Grok 4 | 73.44 | #306 |
| 2 | GPT-5 | 72.66 | #163 |
| 3 | Gemini 2.5 Pro | 67.38 | #262 |
| 4 | Claude Sonnet 4.5 | 60.16 | #244 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=cosimue · How It Works · Data refreshed daily, snapshot 2026-11-12.