Visual Question Answering on RIO-Bench Overall Average
71.69AccuracyQwen3-VL-8B
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Qwen3-VL-8BMethod Variant=Ours2026.07 | 71.69 | 1.58 | |
| Qwen3-VL-8BMethod Variant=Base2026.07 | 70.11 | — | |
| Qwen3-VL-30B-A3BMethod Variant=Ours2026.07 | 66.11 | 1.13 | |
| Qwen3-VL-30B-A3BMethod Variant=Base2026.07 | 64.98 | — | |
| Qwen3-VL-4BMethod Variant=Ours2026.07 | 63.92 | 0.97 | |
| Qwen3-VL-4BMethod Variant=Base2026.07 | 62.95 | — | |
| InternVL3.5-8BMethod Variant=Ours2026.07 | 58.47 | 0.23 | |
| InternVL3.5-8BMethod Variant=Base2026.07 | 58.24 | — | |
| InternVL3.5-14BMethod Variant=Ours2026.07 | 53.08 | 0.01 | |
| InternVL3.5-14BMethod Variant=Base2026.07 | 53.07 | — | |
| Gemma3-12BMethod Variant=Ours2026.07 | 49.78 | 1.53 | |
| Gemma3-12BMethod Variant=Base2026.07 | 48.24 | — | |
| Gemma3-4BMethod Variant=Ours2026.07 | 46.07 | 1.73 | |
| Gemma3-4BMethod Variant=Base2026.07 | 44.35 | — |