MechVQA - Anomaly Detection: leaderboard

MechVQA (2026) tests multimodal LLMs on understanding mechanical engineering drawings, with 20,778 questions on 3,281 drawings scored as accuracy by three LLM judges against reference answers. This board covers the Anomaly Detection subtask of the judging capability, as accuracy in percent on the test split; higher is better.

Metric: Accuracy (%) on the Anomaly Detection subtask (Judging capability), MechVQA test split (drawing-level 8:1:1 split of 20,778 questions on 3,281 mechanical drawings), answers judged against the reference by three LLM judges (GPT-OSS-120B, DeepSeek-V3.2, Kimi-k2); higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 15 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 3 Pro (Preview)78.37#119
2Qwen 3 VL 32B Instruct75.92#606
3GLM-4.6V74.29#710
4GPT-571.02#178
5Qwen 3 VL Plus67.76#512
6Claude Sonnet 4.564.9#261
7Qwen 3 VL 30B A3B Instruct64.08#895
8Qwen 3 VL 4B Instruct62.86#1332
9MiniCPM-V-4.560#1441
10MiMo-VL-7B-SFT59.59#1089
11InternVL3.5-8B53.47#1604
12GPT-4o53.06#751
13GPT-4o Mini35.1#1667
14Gemma 3 27B (IT)24.9#1324
15Llama 3.2 11B Instruct17.55#5174

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

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