Visual Question Answering on GQA (accuracy, Retention %)
62.9AccuracyLLaVA-NeXT-7B (Upper Bound)
Evaluation results
| Method | Links | ||
|---|---|---|---|
| LLaVA-NeXT-7B (Upper Bound)Backbone=LLaVA-NeXT-7B, Tokens Kept=2880, Pruning Ratio=0.0%2025.12 | 62.9 | 100 | |
| FlashVLMBackbone=LLaVA-NeXT-7B, Tokens Kept=640, Pruning Ratio=77.8%2025.12 | 61.9 | 99.12 | |
| VisPrunerBackbone=LLaVA-NeXT-7B, Tokens Kept=640, Pruning Ratio=77.8%2025.12 | 61.6 | 98.6 | |
| FastVBackbone=LLaVA-NeXT-7B, Tokens Kept=640, Pruning Ratio=77.8%2025.12 | 60.4 | 97 | |
| VisionZipBackbone=LLaVA-NeXT-7B, Tokens Kept=640, Pruning Ratio=77.8%2025.12 | 60.1 | 96.7 | |
| SparseVLMBackbone=LLaVA-NeXT-7B, Tokens Kept=640, Pruning Ratio=77.8%2025.12 | 59.1 | 94.9 | |
| FlashVLMBackbone=LLaVA-NeXT-7B, Tokens Kept=320, Pruning Ratio=88.9%2025.12 | 58.9 | 94.61 | |
| VisPrunerBackbone=LLaVA-NeXT-7B, Tokens Kept=320, Pruning Ratio=88.9%2025.12 | 58.4 | 93.3 | |
| VisionZipBackbone=LLaVA-NeXT-7B, Tokens Kept=320, Pruning Ratio=88.9%2025.12 | 58.1 | 90.5 | |
| SparseVLMBackbone=LLaVA-NeXT-7B, Tokens Kept=320, Pruning Ratio=88.9%2025.12 | 56.5 | 87.9 | |
| FastVBackbone=LLaVA-NeXT-7B, Tokens Kept=320, Pruning Ratio=88.9%2025.12 | 55.9 | 87.7 | |
| FlashVLMBackbone=LLaVA-NeXT-7B, Tokens Kept=160, Pruning Ratio=94.4%2025.12 | 55.2 | 88.03 | |
| VisPrunerBackbone=LLaVA-NeXT-7B, Tokens Kept=160, Pruning Ratio=94.4%2025.12 | 54.7 | 86.7 | |
| VisionZipBackbone=LLaVA-NeXT-7B, Tokens Kept=160, Pruning Ratio=94.4%2025.12 | 54.3 | 82.3 | |
| SparseVLMBackbone=LLaVA-NeXT-7B, Tokens Kept=160, Pruning Ratio=94.4%2025.12 | 50.2 | 74.5 | |
| FastVBackbone=LLaVA-NeXT-7B, Tokens Kept=160, Pruning Ratio=94.4%2025.12 | 49.8 | 74.7 |