SOCO - Visual Query: leaderboard

SOCO (2026) tests whether models can match corresponding object parts across images, here posed to vision-language models as four-choice questions over candidate keypoints in a target image, scored strictly so that a question counts only if answered correctly under all four answer orders. In this board the query keypoint is marked on a source image, testing cross-image visual matching; accuracy in percent, higher is better.

Metric: Accuracy (%) when the query keypoint is marked on the source image, four-choice semantic-correspondence questions on SOCO image pairs (target image with candidate keypoints A to D), CircularEval: correct only if the model picks the right keypoint under all four answer permutations (random 0.4, consistent-permutation baseline 25); higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 7 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 3 VL 8B34.2#917
2GPT-4o30.2#751
3InternVL3.5-8B24.9#1604
4Qwen 2.5 VL 7B Instruct19.4#1902
5Qwen 3 VL 4B8.6#1151
6Qwen 2.5 VL 3B Instruct5.2#2713
7LLaVA-OneVision-7B2.9#2981

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

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