Multimodal Question Answering on MMBench EN
86.3AccuracyQwen3-VL-8B
Evaluation results
| Method | Links | |
|---|---|---|
| Qwen3-VL-8BRetained Tokens=1024 Tokens (100%)2026.07 | 86.3 | |
| DivPruneRetained Tokens=512 Tokens (↓ 50.0%)2026.07 | 82.1 | |
| EADPRetained Tokens=512 Tokens (↓ 50.0%)2026.07 | 81.6 | |
| CDPrunerRetained Tokens=512 Tokens (↓ 50.0%)2026.07 | 81 | |
| EADPRetained Tokens=256 Tokens (↓ 75.0%)2026.07 | 80.5 | |
| DivPruneRetained Tokens=256 Tokens (↓ 75.0%)2026.07 | 80 | |
| EADPRetained Tokens=128 Tokens (↓ 87.5%)2026.07 | 79.9 | |
| CDPrunerRetained Tokens=256 Tokens (↓ 75.0%)2026.07 | 79.8 | |
| CDPrunerRetained Tokens=128 Tokens (↓ 87.5%)2026.07 | 77.9 | |
| DivPruneRetained Tokens=128 Tokens (↓ 87.5%)2026.07 | 77 | |
| QMoSLoRAModel=InternLM2 + ViT, Initialization=Kaiming uniform, Quantization=4-bit2024.06 | 73.8 | |
| QMoSLoRAModel=InternLM2 + ViT, Initialization=Orthogonal, Quantization=4-bit2024.06 | 73.5 | |
| MoSLoRAModel=LLaMA-3 + ViT, Initialization=Orthogonal2024.06 | 73 | |
| MoSLoRAModel=LLaMA-3 + ViT, Initialization=Kaiming uniform2024.06 | 72.5 | |
| LoRAModel=LLaMA-3 + ViT2024.06 | 72 | |
| QLoRAModel=InternLM2 + ViT, Quantization=4-bit2024.06 | 70.8 | |
| LLaVA-1.5-7BBackbone=LLaVA-1.5-7B, Retained Tokens=5762026.06 | 64.7 | |
| SparseVLMBackbone=LLaVA-1.5-7B, Retained Tokens=1282026.06 | 64.3 | |
| DARTBackbone=LLaVA-1.5-7B, Retained Tokens=1282026.06 | 63.1 | |
| CDPrunerBackbone=LLaVA-1.5-7B, Retained Tokens=1282026.06 | 63 | |
| SPAREBackbone=LLaVA-1.5-7B, Retained Tokens=1282026.06 | 62.5 | |
| DARTBackbone=LLaVA-1.5-7B, Retained Tokens=642026.06 | 61.6 | |
| SPAREBackbone=LLaVA-1.5-7B, Retained Tokens=642026.06 | 61.5 | |
| CDPrunerBackbone=LLaVA-1.5-7B, Retained Tokens=642026.06 | 61 | |
| PruMerge+Backbone=LLaVA-1.5-7B, Retained Tokens=1282026.06 | 60.7 | |
| SPAREBackbone=LLaVA-1.5-7B, Retained Tokens=322026.06 | 60.6 | |
| SparseVLMBackbone=LLaVA-1.5-7B, Retained Tokens=642026.06 | 60.2 | |
| CDPrunerBackbone=LLaVA-1.5-7B, Retained Tokens=322026.06 | 59.5 | |
| PruMerge+Backbone=LLaVA-1.5-7B, Retained Tokens=642026.06 | 59.2 | |
| DARTBackbone=LLaVA-1.5-7B, Retained Tokens=322026.06 | 58.8 |