Multimodal Benchmark on MMBench (Language Specific & Relative Scores)
83.2MMB-EN ScoreQwen2.5-VL-7B
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Qwen2.5-VL-7BRetained Tokens=1225, Token Retention Ratio=100%2026.03 | 83.2 | 83.2 | 100 | |
| DivPruneRetained Tokens=512, Token Retention Ratio=60.5% Reduction2026.03 | 81.8 | 81 | 96.8 | |
| AgilePrunerRetained Tokens=512, Token Retention Ratio=60.5% Reduction2026.03 | 81.3 | 80.8 | 96.9 | |
| AgilePrunerRetained Tokens=256, Token Retention Ratio=79.1% Reduction2026.03 | 80.9 | 80.3 | 92.1 | |
| DivPruneRetained Tokens=256, Token Retention Ratio=79.1% Reduction2026.03 | 80.3 | 79.8 | 90.7 | |
| Full-FinetuneTarget Model=LLaVA-1.5-13B, Data Budget=100%2026.05 | 68.5 | 61.5 | — | |
| MAGICTarget Model=LLaVA-1.5-13B, Data Budget=20%2026.05 | 67.4 | 61.3 | — | |
| D2-PruningTarget Model=LLaVA-1.5-13B, Data Budget=20%2026.05 | 67.3 | 59.9 | — | |
| COINCIDETarget Model=LLaVA-1.5-13B, Data Budget=20%2026.05 | 64 | 57.7 | — | |
| Self-SupTarget Model=LLaVA-1.5-13B, Data Budget=20%2026.05 | 63.7 | 57.6 | — | |
| RandomTarget Model=LLaVA-1.5-13B, Data Budget=20%2026.05 | 62.8 | 55.2 | — | |
| CLIP-ScoreTarget Model=LLaVA-1.5-13B, Data Budget=20%2026.05 | 60.4 | 54 | — | |
| EL2NTarget Model=LLaVA-1.5-13B, Data Budget=20%2026.05 | 59.3 | 52.3 | — | |
| SemDedupTarget Model=LLaVA-1.5-13B, Data Budget=20%2026.05 | 59 | 51.1 | — | |
| Self-FilterTarget Model=LLaVA-1.5-13B, Data Budget=20%2026.05 | 58.8 | 51.8 | — | |
| PerplexityTarget Model=LLaVA-1.5-13B, Data Budget=20%2026.05 | 57.5 | 50.3 | — |