COHERENCE: leaderboard
Metric: Exact-match accuracy (%): the whole image-to-placeholder assignment is right, over all 6,161 interleaved image-text documents whose images were removed: the model assigns each candidate image to its placeholder; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 19 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 71.82 |
| 2 | GPT-5.4 (High) | 71.29 |
| 3 | Claude Sonnet 4.6 (Thinking) | 65.72 |
| 4 | Qwen 3.5 397B A17B | 64.81 |
| 5 | Doubao-Seed-2.0-Pro-260215 | 62.1 |
| 6 | Kimi K2.5 | 60.19 |
| 7 | Qwen 3.5 122B A10B | 59.57 |
| 8 | Qwen 3.5 35B A3B | 55.82 |
| 9 | GPT-5.2 | 52.46 |
| 10 | Qwen 3 VL 235B A22B | 46.32 |
| 11 | Qwen 3.5 4B | 45.64 |
| 12 | GLM-4.6V | 43.09 |
| 13 | Step3 VL 10B | 40.19 |
| 14 | Qwen 3 VL 8B Instruct | 35.45 |
| 15 | Qwen 3 VL 4B Instruct | 34.28 |
Interactive version: theaggregate.ai/benchmark?slug=coherence · How It Works · Data refreshed daily, snapshot 2026-10-07.