Visual Question Answering on GQA (Accuracy and Average accuracy proportion w.r.t. vanilla)
64.2AccuracyVanilla
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| VanillaBase Model=LLaVA-NeXT-7B, Retained Tokens=2880, Reduction Ratio=0.0%2026.06 | 64.2 | 100 | |
| SCOPEBase Model=LLaVA-NeXT-7B, Retained Tokens=160, Reduction Ratio=94.4%2026.06 | 60 | 94.2 | |
| DCP-PruneBase Model=LLaVA-NeXT-7B, Retained Tokens=160, Reduction Ratio=94.4%2026.06 | 59.7 | 94.6 | |
| PRUNESIDBase Model=LLaVA-NeXT-7B, Retained Tokens=160, Reduction Ratio=94.4%2026.06 | 58.9 | 93.1 | |
| DCP-PruneBase Model=LLaVA-NeXT-7B, Retained Tokens=80, Reduction Ratio=97.2%2026.06 | 58.7 | 91.9 | |
| FlowCutBase Model=LLaVA-NeXT-7B, Retained Tokens=160, Reduction Ratio=94.4%2026.06 | 57.6 | 93.1 | |
| SCOPEBase Model=LLaVA-NeXT-7B, Retained Tokens=80, Reduction Ratio=97.2%2026.06 | 57.6 | 90.7 | |
| HoloVBase Model=LLaVA-NeXT-7B, Retained Tokens=160, Reduction Ratio=94.4%2026.06 | 57.2 | 91 | |
| PRUNESIDBase Model=LLaVA-NeXT-7B, Retained Tokens=80, Reduction Ratio=97.2%2026.06 | 56.3 | 88.9 | |
| DCP-PruneBase Model=LLaVA-NeXT-7B, Retained Tokens=40, Reduction Ratio=98.6%2026.06 | 56.1 | 88.5 | |
| VisionZipBase Model=LLaVA-NeXT-7B, Retained Tokens=160, Reduction Ratio=94.4%2026.06 | 55.5 | 92.3 | |
| FlowCutBase Model=LLaVA-NeXT-7B, Retained Tokens=80, Reduction Ratio=97.2%2026.06 | 54.9 | 89.6 | |
| SCOPEBase Model=LLaVA-NeXT-7B, Retained Tokens=40, Reduction Ratio=98.6%2026.06 | 54.3 | 85.5 | |
| VisionZipBase Model=LLaVA-NeXT-7B, Retained Tokens=80, Reduction Ratio=97.2%2026.06 | 52.3 | 87.2 | |
| DCP-PruneBase Model=LLaVA-NeXT-7B, Retained Tokens=20, Reduction Ratio=99.3%2026.06 | 51.9 | 80.7 | |
| FlowCutBase Model=LLaVA-NeXT-7B, Retained Tokens=40, Reduction Ratio=98.6%2026.06 | 50.7 | 83.8 | |
| PRUNESIDBase Model=LLaVA-NeXT-7B, Retained Tokens=40, Reduction Ratio=98.6%2026.06 | 49.4 | 78.5 | |
| HoloVBase Model=LLaVA-NeXT-7B, Retained Tokens=80, Reduction Ratio=97.2%2026.06 | 48.7 | 81.3 | |
| SCOPEBase Model=LLaVA-NeXT-7B, Retained Tokens=20, Reduction Ratio=99.3%2026.06 | 47.8 | 76 | |
| PRUNESIDBase Model=LLaVA-NeXT-7B, Retained Tokens=20, Reduction Ratio=99.3%2026.06 | 44.8 | 69.2 | |
| HoloVBase Model=LLaVA-NeXT-7B, Retained Tokens=40, Reduction Ratio=98.6%2026.06 | 44.3 | 74.6 | |
| VisionZipBase Model=LLaVA-NeXT-7B, Retained Tokens=40, Reduction Ratio=98.6%2026.06 | 44.1 | 74.7 | |
| VisionZipBase Model=LLaVA-NeXT-7B, Retained Tokens=20, Reduction Ratio=99.3%2026.06 | 42.3 | 73.9 | |
| HoloVBase Model=LLaVA-NeXT-7B, Retained Tokens=20, Reduction Ratio=99.3%2026.06 | 41.2 | 69 | |
| FlowCutBase Model=LLaVA-NeXT-7B, Retained Tokens=20, Reduction Ratio=99.3%2026.06 | 40.6 | 68 |