SciVQR: leaderboard

Multimodal scientific reasoning on SciVQR (2026), 3,254 image-based scientific questions from competitions, entrance exams and graduate-level sources in English and Chinese, answered with zero-shot chain-of-thought, across mathematics, physics, chemistry, biology, geography and astronomy. Overall accuracy in percent, the mean of the six subject accuracies; higher is better.

Metric: Overall accuracy (%) with zero-shot chain-of-thought prompting on SciVQR (3,254 image-based scientific questions from competitions, entrance exams and graduate-level sources, 78 percent multiple-choice and 22 percent free-form, in English and Chinese, high-school to graduate level); higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 15 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1O4 Mini79.3#368
2Gemini 2.5 Flash (Preview 04-17)79.3#453
3O177.2#403
4MiMo-VL-7B-RL-250869.2#823
5Qwen 2.5 VL 72B Instruct61.8#841
6GPT-4o60.7#751
7Keye-VL-1.5-8BThinking59.7#1720
8InternVL3-78B-Instruct53.6#1415
9InternVL3-14B-Instruct49.9#1801
10Qwen 2.5 VL 7B Instruct45#1902
11InternVL3-8B-Instruct43.6#2872
12LLaVA-OneVision-7B32.5#2981
13Molmo-7B-D-092429.8#7207
14MiniCPM-o-2.628.8#2423
15LLaVA-NeXT-7B19.5#5032

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

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