Visual Question Answering on COCO FuzzingRL probes (held-out)
100AccuracyHuman
Evaluation results
| Method | Links | |
|---|---|---|
| Human2026.02 | 100 | |
| GPT-4o2026.02 | 83.86 | |
| Gemini-1.5-Flash (V)2026.02 | 78.71 | |
| Qwen3-VL-32B2026.02 | 73.78 | |
| Llama-3.2-Vision-11B2026.02 | 69.58 | |
| LLaVA-OneVision-1.5-8B2026.02 | 67.45 |