COHERENCE: leaderboard

COHERENCE (2026) removes the images from 6,161 interleaved image-text documents (WikiHow, StoryBird, cooking recipes, science) and asks the model to put each candidate image back in its place. This board is the exact-match accuracy over all documents, the whole assignment right, in percent; higher is better.

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 February 2027. 19 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Gemini 3.1 Pro (Preview)71.82#99
2GPT-5.4High71.29#135
3Claude Sonnet 4.6Thinking65.72#173
4Qwen 3.5 397B A17B64.81#264
5Doubao-Seed-2.0-Pro-26021562.1#158
6Kimi K2.560.19#266
7Qwen 3.5 122B A10B59.57#346
8Qwen 3.5 35B A3B55.82#492
9Doubao-Seed-2.0-Lite-26021555.79#262
10GPT-5.252.46#185
11Doubao-Seed-2.0-Mini-26021550.24
12Intern-S1-Pro47.25#585
13Qwen 3 VL 235B A22B46.32#485
14Qwen 3 VL 30B A3B46.24#784
15Qwen 3.5 4B45.64#1241

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=coherence · How It Works · Data refreshed daily, snapshot 2026-11-21.