ContrAR (Single Frame) - Safety Inspection: leaderboard

ContrAR (2026) tests whether vision-language models detect contradictory virtual content attacks in augmented reality, answering yes or no on real AR videos of which half carry an overlay that contradicts the physical scene, given the middle frame of each video. Detection accuracy in percent on the safety inspection videos; higher is better.

Metric: Detection accuracy (%) for contradictory virtual content attacks in AR videos (42 safety inspection videos; half the videos carry a contradiction, a balanced yes/no decision), one middle frame of each video given (single-frame prompting); mean of three runs; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 11 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1GPT-4.190.48#505
2Gemini 2.5 Pro83.33#276
3GPT-580.95#178
4GPT-4o71.43#751
5Claude Sonnet 4.571.43#261
6Gemini 2.5 Flash69.05#463
7Claude Haiku 4.564.29#583
8Grok 464.29#319
9Qwen 2.5 VL 7B Instruct61.9#1902
10Grok 2 Vision61.9#1437
11Qwen 2.5 VL 72B Instruct50#841

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

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