Visual Understanding on ScienceQA, TextVQA, and GQA
100Avg Relative AccuracyVanilla
Evaluation Results
| Method | Links | |
|---|---|---|
| VanillaModel=Qwen3-VL-4B, Token Retention Ratio=100%2026.04 | 100 | |
| VanillaModel=InternVL3.5-8B, Token Retention Ratio=100%2026.04 | 100 | |
| FastV w/ DSTPModel=Qwen3-VL-4B, Token Retention Ratio=33.3%2026.04 | 97.5 | |
| VisionZip w/ DSTPModel=Qwen3-VL-4B, Token Retention Ratio=33.3%2026.04 | 97.4 | |
| VisionZipModel=Qwen3-VL-4B, Token Retention Ratio=33.3%2026.04 | 96.5 | |
| DivPrune w/ DSTPModel=Qwen3-VL-4B, Token Retention Ratio=33.3%2026.04 | 95.5 | |
| FastV w/ DSTPModel=Qwen3-VL-4B, Token Retention Ratio=22.2%2026.04 | 95.5 | |
| DivPrune w/ DSTPModel=InternVL3.5-8B, Token Retention Ratio=33.3%2026.04 | 95.2 | |
| DivPrune w/ DSTPModel=Qwen3-VL-4B, Token Retention Ratio=22.2%2026.04 | 94.4 | |
| FastVModel=Qwen3-VL-4B, Token Retention Ratio=33.3%2026.04 | 94.3 | |
| VisionZip w/ DSTPModel=Qwen3-VL-4B, Token Retention Ratio=22.2%2026.04 | 94 | |
| VisionZipModel=Qwen3-VL-4B, Token Retention Ratio=22.2%2026.04 | 93 | |
| VisionZip w/ DSTPModel=InternVL3.5-8B, Token Retention Ratio=33.3%2026.04 | 91.8 | |
| FastV w/ DSTPModel=InternVL3.5-8B, Token Retention Ratio=33.3%2026.04 | 91.3 | |
| DivPruneModel=Qwen3-VL-4B, Token Retention Ratio=33.3%2026.04 | 91.2 | |
| DivPrune w/ DSTPModel=InternVL3.5-8B, Token Retention Ratio=22.2%2026.04 | 90.9 | |
| FastVModel=Qwen3-VL-4B, Token Retention Ratio=22.2%2026.04 | 90.2 | |
| DivPruneModel=InternVL3.5-8B, Token Retention Ratio=33.3%2026.04 | 89.2 | |
| DivPruneModel=Qwen3-VL-4B, Token Retention Ratio=22.2%2026.04 | 88.7 | |
| VisionZipModel=InternVL3.5-8B, Token Retention Ratio=33.3%2026.04 | 88.7 | |
| VisionZip w/ DSTPModel=InternVL3.5-8B, Token Retention Ratio=22.2%2026.04 | 88.7 | |
| FastV w/ DSTPModel=InternVL3.5-8B, Token Retention Ratio=22.2%2026.04 | 88.6 | |
| FastVModel=InternVL3.5-8B, Token Retention Ratio=33.3%2026.04 | 87.1 | |
| FastVModel=InternVL3.5-8B, Token Retention Ratio=22.2%2026.04 | 84.7 | |
| DivPruneModel=InternVL3.5-8B, Token Retention Ratio=22.2%2026.04 | 84.3 | |
| VisionZipModel=InternVL3.5-8B, Token Retention Ratio=22.2%2026.04 | 81.8 |