KVBench (Brief Caption) - Math (Chinese): leaderboard

KVBench (2026) tests whether text-to-image models can draw correct senior high-school math content, such as diagrams and scientific illustrations, from Chinese prompts given as a brief caption, so the model must supply the textbook knowledge itself. Qwen2.5-VL-32B checks each image against a list of textbook-grounded constraints; the score is the percentage satisfied, higher is better.

Metric: Checklist score (%) on the math prompts in Chinese, each given as a brief caption, so the model must supply the textbook knowledge itself: KVBench knowledge-intensive text-to-image prompts from over 30 senior high-school textbooks (150 per subject and language, each with a textbook reference image); Qwen2.5-VL-32B-Instruct answers a checklist of four to six binary questions (key objects, attributes, spatial relations, reasoning outcomes) about each generated image, and the score is the mean share of satisfied items, in percent; higher is better. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 14 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Nano Banana Pro (Gemini 3 Pro Image)41.29
2OmniGen236.31#4579
3Seedream 4.035.6
4GPT-Image (KVBench checkpoint unspecified)29.33
5FLUX.2-dev28.27
6FLUX 2 Max24.89
7Show-o223.56
8FLUX.1-Krea-dev21.02
9FLUX.1-dev20.93
10SD3.5-Large (KVBench GGUF build, quantization unspecified)20.49
11Z-Image19.82
12Qwen-Image19.16
13BAGEL-7B-MoT14.09#4619
14Janus-Pro10.49

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