Multimodal Understanding on MLLM Evaluation Suite (LLaVA-NeXT Test Split)
64.2GQA AccuracyVanilla
Evaluation results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| VanillaRetained Tokens=2880, Compression Ratio=100%, Evaluation Protocol=Upper Bound2026.03 | 64.2 | 67.9 | 1,842 | 86.4 | 70.2 | 80.1 | 35.1 | 70.2 | 100 | |
| PRUNESIDRetained Tokens=640, Compression Ratio=77.8%2026.03 | 61.6 | 64.2 | 1,795 | 86.3 | 68.3 | 78.5 | 37.9 | 67.3 | 98.4 | |
| VisionZipRetained Tokens=640, Compression Ratio=77.8%2026.03 | 61.3 | 66.3 | 1,787 | 86.3 | 68.1 | 79.1 | 34.7 | 66.7 | 97.5 | |
| PRUNESIDRetained Tokens=320, Compression Ratio=88.9%2026.03 | 60.5 | 63 | 1,754 | 83.1 | 67.3 | 76.6 | 36.4 | 65 | 95.8 | |
| VisionZipRetained Tokens=320, Compression Ratio=88.9%2026.03 | 59.3 | 63.1 | 1,702 | 82.1 | 67.3 | 76.2 | 35.3 | 63.4 | 94.3 | |
| PRUNESIDRetained Tokens=160, Compression Ratio=94.4%2026.03 | 58.9 | 60.8 | 1,704 | 76.9 | 67.1 | 73.8 | 36.2 | 62.5 | 92.8 | |
| VisionZipRetained Tokens=160, Compression Ratio=94.4%2026.03 | 55.5 | 60.1 | 1,630 | 74.8 | 68.3 | 71.4 | 36.1 | 58.3 | 90.3 |