CArtBench - CuratorQA: leaderboard

Chinese art understanding on CuratorQA, the curator-style question task of CArtBench (2026): multiple-choice and true-or-false questions generated from Palace Museum catalog text and audited by an art expert, answered from the artwork image, over all 14,421 questions on 1,589 artworks. Exact-match accuracy in percent; higher is better.

Metric: Exact-match accuracy (%) on all 14,421 CuratorQA questions of CArtBench (Palace Museum artworks with catalog-grounded multiple-choice and true-or-false questions, answered from the artwork image, constrained decoding, temperature 0); higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 9 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 3 VL 235B A22B (Thinking)84#459
2Qwen 2.5 VL 72B81#781
3Qwen 2.5 VL 32B80#941
4Qwen 3 VL 30B A3B80#784
5GPT-5 Mini79#364
6Qwen 2.5 VL 7B76#1862
7GPT-5 Nano59#906
8GLM-4.5V57#754
9Gemini 2.5 Flash48#463

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

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