Visual Question Answering on RIO-Bench Hard
65.92AccuracyQwen3-VL-8B
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Qwen3-VL-8BMethod Variant=Ours2026.07 | 65.92 | 1.01 | |
| Qwen3-VL-8BMethod Variant=Base2026.07 | 64.91 | — | |
| Qwen3-VL-30B-A3BMethod Variant=Ours2026.07 | 62.35 | 1.45 | |
| Qwen3-VL-30B-A3BMethod Variant=Base2026.07 | 60.9 | — | |
| Qwen3-VL-4BMethod Variant=Ours2026.07 | 55.53 | 0.63 | |
| Qwen3-VL-4BMethod Variant=Base2026.07 | 54.9 | — | |
| InternVL3.5-8BMethod Variant=Ours2026.07 | 50.28 | 0.19 | |
| InternVL3.5-8BMethod Variant=Base2026.07 | 50.09 | — | |
| Gemma3-12BMethod Variant=Ours2026.07 | 47.22 | 1.14 | |
| InternVL3.5-14BMethod Variant=Ours2026.07 | 46.87 | 0.25 | |
| InternVL3.5-14BMethod Variant=Base2026.07 | 46.62 | — | |
| Gemma3-12BMethod Variant=Base2026.07 | 46.08 | — | |
| Gemma3-4BMethod Variant=Ours2026.07 | 40.52 | 1.92 | |
| Gemma3-4BMethod Variant=Base2026.07 | 38.6 | — |