CSVQA: leaderboard

Each vision-language model answers all 1,378 Chinese high-school biology, chemistry, math and physics questions built around an image (1,278 multiple choice, some with more than one correct option, and 100 open) with step-by-step reasoning; the score is accuracy, with a GPT model reading each rationale to pick the answer and GPT-4o scoring open answers. Unlike English science image sets such as ScienceQA, the questions come from Chinese textbooks and exams and most need a calculation.

Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 27 models tracked. Updated 17 Nov 2026.

Top models

#ModelScoreOverall rank
1O1 (2024-12-17)49.6#314
2Qwen 2.5 VL 72B Instruct38.5#849
3InternVL3-78B37.4#755
4Claude 3.7 Sonnet (20250219)36.6#421
5QVQ-72B-Preview36.6#1173
6InternVL3-38B33.4#927
7InternVL3-14B31.9#1191
8Qwen 2.5 VL 32B Instruct28.9#1074
9InternVL2.5-78B28.4#1036
10InternVL2.5-38B25.2#1366
11Qwen 2 VL 72B Instruct24.1#1382
12Gemma 3 27B (IT)22.9#1339
13InternVL2.5-26B22.1#2191
14InternVL2.5-8B20.2#2401
15Qwen 2.5 VL 7B Instruct17.5#1893

No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.

Interactive version: theaggregate.ai/benchmark?slug=csvqa · How It Works · Data refreshed daily, snapshot 2026-11-17.