OpenVLM A-Bench - Contradiction Overcome: leaderboard
A-Bench Contradiction Overcome accuracy from the OpenCompass OpenVLM Leaderboard (VLMEvalKit): multiple-choice questions on AI-generated images, on images generated from prompts that contradict common sense, answered from what the image shows. A dimension of A-Bench-P1 (high-level semantic understanding). Accuracy in percent, higher is better.
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 160 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | InternVL3-8B | 95.8 | #624 |
| 2 | Grok 2 (1212) | 91.7 | #502 |
| 3 | Qwen 2 VL 7B | 87.5 | #743 |
| 4 | Qwen 2 VL 72B | 87.5 | #450 |
| 5 | Qwen 2 VL 2B | 87.5 | #1191 |
| 6 | InternVL2-8B | 83.3 | #813 |
| 7 | Gemma 3 27B | 83.3 | #606 |
| 8 | Gemma 3 4B | 83.3 | #1088 |
| 9 | InternVL3-38B | 79.2 | #413 |
| 10 | InternVL3-14B | 79.2 | #506 |
| 11 | Qwen 2.5 VL 7B | 79.2 | #641 |
| 12 | Gemma 3 12B | 79.2 | #674 |
| 13 | Gemini 1.5 Flash | 79.2 | #620 |
| 14 | InternVL3-78B | 75 | #361 |
| 15 | Gemini 1.5 Pro | 75 | #454 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=openvlm-a-bench-contradiction-overcome · How It Works · Data refreshed daily, snapshot 2026-11-05.