Text-based Visual Question Answering on TextVQA (test)
83.8AccuracyVanilla
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| VanillaBackbone=InternVL3-38B, ViT token retention rate=100%, LLM token retention rate=100%2025.12 | 83.8 | 100 | 100 | — | — | — | |
| Upper BoundBackbone=Qwen2.5-VL-7B2026.04 | 77.8 | — | — | — | — | — | |
| CLASPBackbone=Qwen2.5-VL-7B, Token Pruning Rate=66.7%2026.04 | 74.7 | — | — | — | — | — | |
| CLASPBackbone=Qwen2.5-VL-7B, Token Pruning Rate=77.8%2026.04 | 73.1 | — | — | — | — | — | |
| SparseVLMBackbone=Qwen2.5-VL-7B, Token Pruning Rate=66.7%2026.04 | 71.5 | — | — | — | — | — | |
| IPCV+FastVBackbone=InternVL3-38B, ViT token retention rate=40%, LLM token retention rate=40%2025.12 | 70.3 | 89.9 | 64.4 | — | — | — | |
| SparseVLMBackbone=Qwen2.5-VL-7B, Token Pruning Rate=77.8%2026.04 | 68.7 | — | — | — | — | — | |
| CLASPBackbone=Qwen2.5-VL-7B, Token Pruning Rate=88.9%2026.04 | 68 | — | — | — | — | — | |
| Upper BoundBase Model=LLaVA-NeXT-13B2026.04 | 64.3 | — | — | — | — | — | |
| CLASPBase Model=LLaVA-NeXT-13B, Token Retention Budget=640 Tokens2026.04 | 61.5 | — | — | — | — | — | |
| SparseVLMBackbone=Qwen2.5-VL-7B, Token Pruning Rate=88.9%2026.04 | 61.2 | — | — | — | — | — | |
| CLASPBase Model=LLaVA-NeXT-13B, Token Retention Budget=320 Tokens2026.04 | 60.8 | — | — | — | — | — | |
| DART (ViT)Backbone=InternVL3-38B, ViT token retention rate=40%, LLM token retention rate=40%2025.12 | 59.7 | 88.1 | 63.9 | — | — | — | |
| SparseVLMBase Model=LLaVA-NeXT-13B, Token Retention Budget=640 Tokens2026.04 | 59.3 | — | — | — | — | — | |
| ALVTSTokens Retained=192, Reduction Ratio=67%2026.06 | 58.29 | — | — | — | — | — | |
| LLaVA-1.5-7BTokens Retained=576, Retention Ratio=100%2026.06 | 58.18 | — | — | — | — | — | |
| ALVTSTokens Retained=128, Reduction Ratio=78%2026.06 | 57.92 | — | — | — | — | — | |
| FastVTokens Retained=192, Reduction Ratio=67%2026.06 | 57.91 | — | — | — | — | — | |
| CLASPBase Model=LLaVA-NeXT-13B, Token Retention Budget=160 Tokens2026.04 | 57.8 | — | — | — | — | — | |
| ALVTSTokens Retained=64, Reduction Ratio=89%2026.06 | 57.41 | — | — | — | — | — | |
| DARTTokens Retained=192, Reduction Ratio=67%2026.06 | 57.4 | — | — | — | — | — | |
| SparseVLMBase Model=LLaVA-NeXT-13B, Token Retention Budget=320 Tokens2026.04 | 57.2 | — | — | — | — | — | |
| PDropTokens Retained=192, Reduction Ratio=67%2026.06 | 57.11 | — | — | — | — | — | |
| FastVTokens Retained=128, Reduction Ratio=78%2026.06 | 56.95 | — | — | — | — | — | |
| DARTTokens Retained=128, Reduction Ratio=78%2026.06 | 56.64 | — | — | — | — | — | |
| PDropTokens Retained=128, Reduction Ratio=78%2026.06 | 56.18 | — | — | — | — | — | |
| FastVTokens Retained=64, Reduction Ratio=89%2026.06 | 55.32 | — | — | — | — | — | |
| FastVBackbone=InternVL3-38B, ViT token retention rate=100%, LLM token retention rate=5%2025.12 | 55.1 | 78.6 | 70.7 | — | — | — | |
| DARTTokens Retained=64, Reduction Ratio=89%2026.06 | 54.53 | — | — | — | — | — | |
| PDropTokens Retained=64, Reduction Ratio=89%2026.06 | 53.36 | — | — | — | — | — | |
| SparseVLMBase Model=LLaVA-NeXT-13B, Token Retention Budget=160 Tokens2026.04 | 53.3 | — | — | — | — | — | |
| ToFuBackbone=InternVL3-38B, ViT token retention rate=40%, LLM token retention rate=40%2025.12 | 52.7 | 83.4 | 61.9 | — | — | — | |
| ToMeBackbone=InternVL3-38B, ViT token retention rate=40%, LLM token retention rate=40%2025.12 | 52.2 | 82.7 | 61.7 | — | — | — | |
| SparseVLMBackbone=InternVL3-38B, ViT token retention rate=100%, LLM token retention rate=5%2025.12 | 52.2 | 75.1 | 73.4 | — | — | — | |
| VTWTokens Retained=192, Reduction Ratio=67%2026.06 | 50.6 | — | — | — | — | — | |
| DARTBackbone=InternVL3-38B, ViT token retention rate=100%, LLM token retention rate=5%2025.12 | 46.4 | 73.6 | 71.5 | — | — | — | |
| VTWTokens Retained=128, Reduction Ratio=78%2026.06 | 44.26 | — | — | — | — | — | |
| VTWTokens Retained=64, Reduction Ratio=89%2026.06 | 42.85 | — | — | — | — | — | |
| V^2DropBackbone=InternVL3-38B, ViT token retention rate=100%, LLM token retention rate=5%2025.12 | 17.1 | 57.1 | 70.5 | — | — | — | |
| FGSMModel=LLaVA-1.5, Attack Type=White-box2026.02 | — | — | — | 47.3 | 37.2 | 10.1 | |
| FGSMModel=PrismVLM, Attack Type=White-box2026.02 | — | — | — | 56.9 | 41.3 | 15.6 | |
| MDCS-MIModel=LLaVA-1.5, Attack Type=White-box2026.02 | — | — | — | 47.3 | 20.9 | 26.4 | |
| MDCS-MIModel=PrismVLM, Attack Type=White-box2026.02 | — | — | — | 56.9 | 22.8 | 34.1 | |
| MEFModel=LLaVA-1.5, Attack Type=White-box2026.02 | — | — | — | 47.3 | 22.1 | 25.2 | |
| MEFModel=PrismVLM, Attack Type=White-box2026.02 | — | — | — | 56.9 | 25.8 | 31.1 | |
| MIModel=LLaVA-1.5, Attack Type=White-box2026.02 | — | — | — | 47.3 | 20.9 | 26.4 | |
| MIModel=PrismVLM, Attack Type=White-box2026.02 | — | — | — | 56.9 | 26 | 30.9 | |
| PGDModel=LLaVA-1.5, Attack Type=White-box2026.02 | — | — | — | 47.3 | 22.1 | 25.2 | |
| PGDModel=PrismVLM, Attack Type=White-box2026.02 | — | — | — | 56.9 | 26.1 | 30.8 |