VeriOCRBench - Contextual Dimension (Proactive): leaderboard

Metric: Diagnosis accuracy (%; the 400 trap-injected tasks of the contextual verification dimension (domain mismatch and hallucinatory target: the provided context imposes a wrong domain on the image or refers to a target that does not exist in it); the model must reject the task and name the ground-truth trap mechanism; judged by GPT-4o at temperature 0; proactive paradigm: no verification instruction). Source: arxiv.org. Saturation forecast: Around February 2027. 15 models tracked.

Top models

#ModelScore
1Claude Sonnet 4.586.75
2Qwen 3.5 397B A17B60.75
3Qwen 3.5 35B A3B60.75
4Gemini 3.1 Pro (Preview)56.25
5Qwen 3.5 122B A10B55.25
6Llama 4 Scout46
7Llama 4 Maverick34.75
8InternVL3-38B28.75
9GPT-4o22.5
10GPT-518.5

Interactive version: theaggregate.ai/benchmark?slug=veriocrbench-contextual-dimension-proactive · How It Works · Data refreshed daily, snapshot 2026-09-26.