Multi-modal Understanding on LLM Evaluation Suite (GQA, MMB, MME, POPE, SQA, VQAv2, VQA^Text, VizWiz)
69.2GQA ScoreVanilla
Evaluation results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| VanillaModel Backbone=LLaVA-NeXT-7B, Token Budget=Upper Bound (100%)2026.06 | 69.2 | 67.9 | 1,842 | 86.4 | 70.2 | 80.1 | 61.3 | 55.2 | 100 | |
| STSModel Backbone=LLaVA-NeXT-7B, Token Budget=Retain 640 Tokens (↓ 77.8%)2026.06 | 63.4 | 65.9 | 1,801 | 87.6 | 67.7 | 79.9 | 58.9 | 55.7 | 97.6 | |
| Zoo-PruneModel Backbone=LLaVA-NeXT-7B, Token Budget=Retain 640 Tokens (↓ 77.8%)2026.06 | 62.2 | 65.2 | 1,816 | 86.8 | 68 | 79.6 | 58 | — | 96.6 | |
| AgilePruneModel Backbone=LLaVA-NeXT-7B, Token Budget=Retain 640 Tokens (↓ 77.8%)2026.06 | 62 | 65.9 | — | 86.1 | 67.8 | 79.3 | 59 | 56 | 97.1 | |
| DivPruneModel Backbone=LLaVA-NeXT-7B, Token Budget=Retain 640 Tokens (↓ 77.8%)2026.06 | 61.6 | 65.4 | 1,773 | 85.5 | 67.8 | 78.9 | 55.4 | 55.7 | 95.9 | |
| STSModel Backbone=LLaVA-NeXT-7B, Token Budget=Retain 320 Tokens (↓ 88.9%)2026.06 | 61.5 | 64.5 | 1,791 | 87.5 | 67.7 | 78.8 | 58.6 | 56.3 | 96.8 | |
| VisionZipModel Backbone=LLaVA-NeXT-7B, Token Budget=Retain 640 Tokens (↓ 77.8%)2026.06 | 61.3 | 65.8 | 1,787 | 86.3 | 68.1 | 79.1 | 60.2 | 57.1 | 97.5 | |
| DARTModel Backbone=LLaVA-NeXT-7B, Token Budget=Retain 640 Tokens (↓ 77.8%)2026.06 | 61.3 | 64.9 | 1,781 | 85 | 68.2 | 78.3 | 59.5 | 57 | 96.8 | |
| STSModel Backbone=LLaVA-NeXT-7B, Token Budget=Retain 160 Tokens (↓ 94.4%)2026.06 | 61.1 | 63.7 | 1,765 | 87.1 | 67.8 | 76.9 | 55.4 | 55.8 | 95.3 | |
| Zoo-PruneModel Backbone=LLaVA-NeXT-7B, Token Budget=Retain 320 Tokens (↓ 88.9%)2026.06 | 60.9 | 64.9 | 1,787 | 85.5 | 67.8 | 78.1 | 57.3 | — | 95.3 | |
| SparseVLMModel Backbone=LLaVA-NeXT-7B, Token Budget=Retain 640 Tokens (↓ 77.8%)2026.06 | 60.3 | 65.7 | 1,772 | 85.2 | 67.7 | 77.1 | 57.8 | 53.6 | 95.3 | |
| AgilePruneModel Backbone=LLaVA-NeXT-7B, Token Budget=Retain 320 Tokens (↓ 88.9%)2026.06 | 60.1 | 64.5 | — | 84 | 67.3 | 77.8 | 58.9 | 55.8 | 95.6 | |
| Zoo-PruneModel Backbone=LLaVA-NeXT-7B, Token Budget=Retain 160 Tokens (↓ 94.4%)2026.06 | 59.9 | 64.2 | 1,738 | 83.1 | 68.4 | 76.1 | 55.4 | — | 93.5 | |
| DivPruneModel Backbone=LLaVA-NeXT-7B, Token Budget=Retain 320 Tokens (↓ 88.9%)2026.06 | 59.6 | 63.7 | 1,731 | 83.5 | 67.8 | 76.6 | 53.9 | 55.6 | 93.9 | |
| DARTModel Backbone=LLaVA-NeXT-7B, Token Budget=Retain 320 Tokens (↓ 88.9%)2026.06 | 59.5 | 64.2 | 1,743 | 81 | 67.5 | 75.7 | 57.6 | 56.8 | 94.5 | |
| VisionZipModel Backbone=LLaVA-NeXT-7B, Token Budget=Retain 320 Tokens (↓ 88.9%)2026.06 | 59.3 | 63.1 | 1,702 | 82.1 | 67.3 | 76.2 | 58.9 | 56.2 | 94.4 | |
| DivPruneModel Backbone=LLaVA-NeXT-7B, Token Budget=Retain 160 Tokens (↓ 94.4%)2026.06 | 57.8 | 62 | 1,658 | 79.4 | 68 | 73.9 | 52.4 | 56.1 | 91.6 | |
| SparseVLMModel Backbone=LLaVA-NeXT-7B, Token Budget=Retain 320 Tokens (↓ 88.9%)2026.06 | 57.7 | 64.3 | 1,684 | 78.6 | 67.3 | 73.4 | 55.9 | 54.2 | 92.2 | |
| VisionZipModel Backbone=LLaVA-NeXT-7B, Token Budget=Retain 160 Tokens (↓ 94.4%)2026.06 | 55.5 | 60.1 | 1,630 | 74.8 | 68.3 | 71.4 | 56.2 | 55.5 | 90.3 | |
| SparseVLMModel Backbone=LLaVA-NeXT-7B, Token Budget=Retain 160 Tokens (↓ 94.4%)2026.06 | 51.2 | 63.1 | 1,542 | 77.3 | 67.5 | 66.3 | 46.4 | — | 85 |