Multimodal Evaluation on MM-Vet
64ScoreUpper Bound
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| Upper BoundBackbone=InternVL2-26B2026.04 | 64 | — | — | — | |
| CLASPBackbone=InternVL2-26B, Token Retention=35%2026.04 | 63.2 | — | — | — | |
| InternVL2-8BModel=InternVL2-8B, Pruning Method=None2025.09 | 60 | — | — | — | |
| PTPModel=InternVL2-8B, Pruning Method=PTP, Pruning Ratio (r)=0.5, Balancing Weight (alpha)=0.52025.09 | 56.9 | — | — | — | |
| ToMeBackbone=InternVL2-26B, Token Retention=35%2026.04 | 52.5 | — | — | — | |
| Qwen-VL-Chat#Vision Tokens=2562026.03 | 47.3 | — | — | — | |
| VisionZipBackbone=LLaVA-NeXT-13B, Retained Tokens=640, Pruning Ratio=77.8%2026.02 | 46.8 | — | — | — | |
| VisPrunerBackbone=LLaVA-NeXT-13B, Retained Tokens=640, Pruning Ratio=77.8%2026.02 | 46.1 | — | — | — | |
| VisPrunerBackbone=LLaVA-NeXT-13B, Retained Tokens=960, Pruning Ratio=66.7%2026.02 | 45 | — | — | — | |
| FastVBackbone=InternVL2-26B, Token Retention=35%2026.04 | 45 | — | — | — | |
| FSRBackbone=LLaVA-NeXT-13B, Retained Tokens=960, Pruning Ratio=66.7%2026.02 | 44.7 | — | — | — | |
| FSRBackbone=LLaVA-NeXT-13B, Retained Tokens=640, Pruning Ratio=77.8%2026.02 | 44.1 | — | — | — | |
| FastVBackbone=LLaVA-NeXT-13B, Retained Tokens=640, Pruning Ratio=77.8%2026.02 | 43.8 | — | — | — | |
| HoloVBackbone=LLaVA-NeXT-13B, Retained Tokens=960, Pruning Ratio=66.7%2026.02 | 42.3 | — | — | — | |
| FSRBackbone=LLaVA-NeXT-13B, Retained Tokens=320, Pruning Ratio=88.9%2026.02 | 42.3 | — | — | — | |
| CDPrunerBackbone=LLaVA-NeXT-13B, Retained Tokens=320, Pruning Ratio=88.9%2026.02 | 42.1 | — | — | — | |
| FSRRetained Tokens=640, Reduction Ratio=↓ 77.8%2026.02 | 41.9 | — | — | — | |
| CDPrunerBackbone=LLaVA-NeXT-13B, Retained Tokens=960, Pruning Ratio=66.7%2026.02 | 41.7 | — | — | — | |
| CDPrunerBackbone=LLaVA-NeXT-13B, Retained Tokens=640, Pruning Ratio=77.8%2026.02 | 41.4 | — | — | — | |
| HoloVBackbone=LLaVA-NeXT-13B, Retained Tokens=640, Pruning Ratio=77.8%2026.02 | 41.3 | — | — | — | |
| VisionZipBackbone=LLaVA-NeXT-13B, Retained Tokens=320, Pruning Ratio=88.9%2026.02 | 41.1 | — | — | — | |
| FSRRetained Tokens=960, Reduction Ratio=↓ 66.7%2026.02 | 41.1 | — | — | — | |
| CDPrunerRetained Tokens=640, Reduction Ratio=↓ 77.8%2026.02 | 41 | — | — | — | |
| HoloVBackbone=LLaVA-NeXT-13B, Retained Tokens=320, Pruning Ratio=88.9%2026.02 | 40.7 | — | — | — | |
| VisPrunerBackbone=LLaVA-NeXT-13B, Retained Tokens=320, Pruning Ratio=88.9%2026.02 | 39.7 | — | — | — | |
| IDEFICS-80B#Vision Tokens=642026.03 | 39.7 | — | — | — | |
| InternVL2-2BModel=InternVL2-2B, Pruning Method=None2025.09 | 39.6 | — | — | — | |
| FastVRetained Tokens=640, Reduction Ratio=↓ 77.8%2026.02 | 39.5 | — | — | — | |
| FSRRetained Tokens=320, Reduction Ratio=↓ 88.9%2026.02 | 39.3 | — | — | — | |
| DivPruneBackbone=LLaVA-NeXT-13B, Retained Tokens=320, Pruning Ratio=88.9%2026.02 | 39.2 | — | — | — | |
| LLaVA-NeXT-7BRetained Tokens=2880, Reduction Ratio=100%2026.02 | 39.2 | — | — | — | |
| DivPruneBackbone=LLaVA-NeXT-13B, Retained Tokens=640, Pruning Ratio=77.8%2026.02 | 39 | — | — | — | |
| CDPrunerRetained Tokens=960, Reduction Ratio=↓ 66.7%2026.02 | 39 | — | — | — | |
| HoloVRetained Tokens=640, Reduction Ratio=↓ 77.8%2026.02 | 38.9 | — | — | — | |
| FSRBackbone=LLaVA-1.5-13B, Retained Tokens=128, Pruning Ratio=77.8%2026.02 | 38.8 | — | — | — | |
| PTPModel=InternVL2-2B, Pruning Method=PTP, Pruning Ratio (r)=0.5, Balancing Weight (alpha)=0.52025.09 | 38.6 | — | — | — | |
| VisPrunerRetained Tokens=960, Reduction Ratio=↓ 66.7%2026.02 | 38.5 | — | — | — | |
| VisPrunerRetained Tokens=640, Reduction Ratio=↓ 77.8%2026.02 | 38.5 | — | — | — | |
| DivPruneRetained Tokens=640, Reduction Ratio=↓ 77.8%2026.02 | 38 | — | — | — | |
| FSRBackbone=LLaVA-1.5-13B, Retained Tokens=192, Pruning Ratio=66.7%2026.02 | 37.4 | — | — | — | |
| CDPrunerBackbone=LLaVA-1.5-13B, Retained Tokens=128, Pruning Ratio=77.8%2026.02 | 37.3 | — | — | — | |
| LeoLanguage Model Scale=7B, Training Data Scale=Standard2025.01 | 37.2 | — | — | — | |
| ERABackbone=LLaVA-1.5-13B, Retained Tokens=1282026.06 | 37.2 | — | — | — | |
| ERABackbone=LLaVA-1.5-13B, Retained Tokens=642026.06 | 37.1 | — | — | — | |
| VisPrunerBackbone=LLaVA-1.5-13B, Retained Tokens=192, Pruning Ratio=66.7%2026.02 | 37 | — | — | — | |
| VisPrunerBackbone=LLaVA-1.5-13B, Retained Tokens=128, Pruning Ratio=77.8%2026.02 | 36.9 | — | — | — | |
| VisPrunerBackbone=LLaVA-1.5-13B, Retained Tokens=1282026.06 | 36.9 | — | — | — | |
| LLaVA-1.5-13BBackbone=LLaVA-1.5-13B, Retained Tokens=576, Pruning Ratio=0%2026.02 | 36.7 | — | — | — | |
| FSRBackbone=LLaVA-1.5-13B, Retained Tokens=64, Pruning Ratio=88.9%2026.02 | 36.7 | — | — | — | |
| CDPrunerRetained Tokens=320, Reduction Ratio=↓ 88.9%2026.02 | 36.7 | — | — | — | |
| LLaVA-NeXT-13BBackbone=LLaVA-NeXT-13B, Retained Tokens=2880, Pruning Ratio=0%2026.02 | 36.6 | — | — | — | |
| HoloVRetained Tokens=320, Reduction Ratio=↓ 88.9%2026.02 | 36.5 | — | — | — | |
| VisPrunerRetained Tokens=320, Reduction Ratio=↓ 88.9%2026.02 | 36.3 | — | — | — | |
| MAGICTarget Model=LLaVA-1.5-13B, Data Budget=20%2026.05 | 36.3 | — | — | — | |
| mPLUG-OwL2#Vision Tokens=10242026.03 | 36.2 | — | — | — | |
| LLaVA-1.5-13B#Vision Tokens=576, Backbone=LLaVA-1.5-13B2026.03 | 36.2 | — | — | — | |
| CDPruner#Vision Tokens=128, Backbone=LLaVA-1.5-13B2026.03 | 36.2 | — | — | — | |
| LLaVA-1.5-13BBackbone=LLaVA-1.5-13B, Retained Tokens=576 (Upper Bound)2026.06 | 36.2 | — | — | — | |
| CDPrunerBackbone=LLaVA-1.5-13B, Retained Tokens=1282026.06 | 36.2 | — | — | — | |
| VisPruner#Vision Tokens=128, Backbone=LLaVA-1.5-13B2026.03 | 36.1 | — | — | — | |
| Full-FinetuneTarget Model=LLaVA-1.5-13B, Data Budget=100%2026.05 | 36.1 | — | — | — | |
| VisionZipBackbone=LLaVA-1.5-13B, Retained Tokens=128, Pruning Ratio=77.8%2026.02 | 36 | — | — | — | |
| SPHINX#Vision Tokens=2892026.03 | 36 | — | — | — | |
| VisionZip#Vision Tokens=128, Backbone=LLaVA-1.5-13B2026.03 | 36 | — | — | — | |
| VisionZipBackbone=LLaVA-1.5-13B, Retained Tokens=1282026.06 | 36 | — | — | — | |
| CDPrunerBackbone=LLaVA-1.5-13B, Retained Tokens=64, Pruning Ratio=88.9%2026.02 | 35.8 | — | — | — | |
| CDPrunerBackbone=LLaVA-1.5-13B, Retained Tokens=192, Pruning Ratio=66.7%2026.02 | 35.7 | — | — | — | |
| VanillaBackbone=LLaVA-1.5-13B2025.10 | 35.6 | — | — | — | |
| ETABackbone=LLaVA-1.5-13B2025.10 | 35.6 | — | — | — | |
| HoloVBackbone=LLaVA-1.5-13B, Retained Tokens=192, Pruning Ratio=66.7%2026.02 | 35.5 | — | — | — | |
| ECSOBackbone=LLaVA-1.5-13B2025.10 | 35.5 | — | — | — | |
| AutoSteerBackbone=LLaVA-1.5-13B2025.10 | 35.5 | — | — | — | |
| RandomTarget Model=LLaVA-1.5-13B, Data Budget=20%2026.05 | 35.5 | — | — | — | |
| HoloVBackbone=LLaVA-1.5-13B, Retained Tokens=128, Pruning Ratio=77.8%2026.02 | 35.4 | — | — | — | |
| FMVR-LLaVA#Vision Tokens=576, Backbone=LLaVA-1.5-13B2026.03 | 35.4 | — | — | — | |
| LLaVA-1.5LLM=LLaMA-13B [59], Model Type=Dense Model2025.07 | 35.4 | — | — | 64.2 | |
| SparseVLM#Vision Tokens=128, Backbone=LLaVA-1.5-13B2026.03 | 35.2 | — | — | — | |
| SparseVLMBackbone=LLaVA-1.5-13B, Retained Tokens=1282026.06 | 35.2 | — | — | — | |
| CDPrunerBackbone=LLaVA-1.5-13B, Retained Tokens=642026.06 | 35.2 | — | — | — | |
| TRIM#Vision Tokens=128, Backbone=LLaVA-1.5-13B2026.03 | 35.1 | — | — | — | |
| MoRASBackbone=LLaVA-1.5-13B2025.10 | 35.1 | — | — | — | |
| TRIMBackbone=LLaVA-1.5-13B, Retained Tokens=1282026.06 | 35.1 | — | — | — | |
| FSRRetained Tokens=128, Reduction Ratio=77.8%2026.02 | 34.9 | — | — | 98.3 | |
| VILALanguage Model Scale=7B, Training Data Scale=Standard2025.01 | 34.9 | — | — | — | |
| Our MethodLLM=Phi2-2.7B [62], Model Type=Sparse Model, Expert Selection (Top-k)=Top42025.07 | 34.9 | — | — | 62.3 | |
| DivPruneRetained Tokens=320, Reduction Ratio=↓ 88.9%2026.02 | 34.8 | — | — | — | |
| DART#Vision Tokens=128, Backbone=LLaVA-1.5-13B2026.03 | 34.8 | — | — | — | |
| ASTRABackbone=LLaVA-1.5-13B2025.10 | 34.8 | — | — | — | |
| DARTBackbone=LLaVA-1.5-13B, Retained Tokens=1282026.06 | 34.8 | — | — | — | |
| DivPruneBackbone=LLaVA-1.5-13B, Retained Tokens=128, Pruning Ratio=77.8%2026.02 | 34.4 | — | — | — | |
| HoloVRetained Tokens=960, Reduction Ratio=↓ 66.7%2026.02 | 34.4 | — | — | — | |
| DivPrune#Vision Tokens=128, Backbone=LLaVA-1.5-13B2026.03 | 34.4 | — | — | — | |
| DivPruneBackbone=LLaVA-1.5-13B, Retained Tokens=1282026.06 | 34.4 | — | — | — | |
| FMVR-LLaVA#Vision Tokens=5762026.03 | 34.3 | — | — | — | |
| MoE-LLaVA-4Top2LLM=Phi2-2.7B [62], Model Type=Sparse Model, Expert Selection (Top-k)=4Top22025.07 | 34.3 | — | — | 61.1 | |
| FMVR-LLaVA#Vision Tokens=144, Backbone=LLaVA-1.5-13B2026.03 | 34.2 | — | — | — | |
| FlashVLMToken Budget=128 Tokens, Pruning Rate=77.8%2025.12 | 34.1 | — | — | — | |
| HoloVBackbone=LLaVA-1.5-13B, Retained Tokens=64, Pruning Ratio=88.9%2026.02 | 33.9 | — | — | — | |
| CDPrunerRetained Tokens=192, Reduction Ratio=66.7%2026.02 | 33.9 | — | — | 98.5 | |
| FSRRetained Tokens=192, Reduction Ratio=66.7%2026.02 | 33.9 | — | — | 99.1 |