Text-based Visual Question Answering on TextVQA (Score & Avg. Rel.)
85.2ScoreFull Model
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Full ModelRatio=1.02026.04 | 85.2 | 100 | |
| HAWKRatio=0.62026.04 | 85 | 99.1 | |
| HAWKRatio=0.82026.04 | 83 | 95.8 | |
| VanillaToken Reduction Ratio=Full Tokens (100%), Model=Qwen2-VL-72B, Quantization=4-bit2026.06 | 82 | — | |
| HAWKRatio=0.92026.04 | 79.8 | 88.5 | |
| FastVToken Reduction Ratio=↓ 66.7%, Model=Qwen2-VL-72B, Quantization=4-bit2026.06 | 79.1 | — | |
| PDropToken Reduction Ratio=↓ 66.7%, Model=Qwen2-VL-72B, Quantization=4-bit2026.06 | 78.9 | — | |
| CLSEToken Reduction Ratio=↓ 66.7%, Model=Qwen2-VL-72B, Quantization=4-bit2026.06 | 78.5 | — | |
| PDropToken Reduction Ratio=↓ 77.8%, Model=Qwen2-VL-72B, Quantization=4-bit2026.06 | 76.7 | — | |
| FastVToken Reduction Ratio=↓ 77.8%, Model=Qwen2-VL-72B, Quantization=4-bit2026.06 | 76.6 | — | |
| VanillaBackbone=InternVL-2.5-8B, Token Reduction Rate=100%2026.04 | 76.23 | — | |
| CLSEToken Reduction Ratio=↓ 77.8%, Model=Qwen2-VL-72B, Quantization=4-bit2026.06 | 76 | — | |
| VisionZipRatio=0.62026.04 | 75.3 | 96.1 | |
| FastVBackbone=InternVL-2.5-8B, Token Reduction Rate=25%2026.04 | 74.65 | — | |
| DARTRatio=0.62026.04 | 73.8 | 89.5 | |
| DASHBackbone=InternVL-2.5-8B, Token Reduction Rate=25%2026.04 | 73.28 | — | |
| CLSEToken Reduction Ratio=↓ 88.9%, Model=Qwen2-VL-72B, Quantization=4-bit2026.06 | 72.1 | — | |
| DivPruneBackbone=InternVL-2.5-8B, Token Reduction Rate=25%2026.04 | 70.43 | — | |
| VisionZipRatio=0.82026.04 | 70.1 | 89.2 | |
| PDropToken Reduction Ratio=↓ 88.9%, Model=Qwen2-VL-72B, Quantization=4-bit2026.06 | 70 | — | |
| FastVToken Reduction Ratio=↓ 88.9%, Model=Qwen2-VL-72B, Quantization=4-bit2026.06 | 68.8 | — | |
| DARTRatio=0.82026.04 | 67.2 | 80.4 | |
| DARTBackbone=InternVL-2.5-8B, Token Reduction Rate=25%2026.04 | 64.74 | — | |
| VisionZipRatio=0.92026.04 | 61.4 | 77.5 | |
| VanillaTokens=2880, Retaining Ratio=100%, Backbone=LLaVA-NeXT2025.11 | 61.3 | 100 | |
| OC-VTPTokens=640, Retaining Ratio=22.2%, Individual Training=true, Backbone=LLaVA-NeXT2025.11 | 61.1 | 99.7 | |
| VanillaModel=LLaVA-Next-7B, Token Budget=2880, Reduction Ratio=0%2026.05 | 60.3 | 100 | |
| VisionZipTokens=640, Retaining Ratio=22.2%, Backbone=LLaVA-NeXT2025.11 | 60.2 | 98.2 | |
| OC-VTPTokens=640, Retaining Ratio=22.2%, Backbone=LLaVA-NeXT2025.11 | 60.1 | 98 | |
| HiPruneTokens=640, Retaining Ratio=22.2%, Backbone=LLaVA-NeXT2025.11 | 60 | 97.9 | |
| DARTRatio=0.92026.04 | 59.8 | 72.2 | |
| SPprunerModel=LLaVA-Next-7B, Token Budget=320, Reduction Ratio=88.9%2026.05 | 59.3 | 97.7 | |
| VisionZipTokens=320, Retaining Ratio=11.1%, Backbone=LLaVA-NeXT2025.11 | 58.9 | 96.1 | |
| VisionZipModel=LLaVA-Next-7B, Token Budget=320, Reduction Ratio=88.9%2026.05 | 58.8 | 96.2 | |
| DARTModel=LLaVA-Next-7B, Token Budget=320, Reduction Ratio=88.9%2026.05 | 58.4 | 96.4 | |
| OC-VTPTokens=320, Retaining Ratio=11.1%, Individual Training=true, Backbone=LLaVA-NeXT2025.11 | 58.4 | 95.3 | |
| VanillaToken Budget (Retained Tokens)=576, Reduction Ratio (↓ %)=0%, Base Model (LLaVA-1.5-7B)=LLaVA-1.5-7B2026.05 | 58.2 | 100 | |
| OC-VTPTokens=320, Retaining Ratio=11.1%, Backbone=LLaVA-NeXT2025.11 | 58.2 | 94.9 | |
| HiPruneTokens=320, Retaining Ratio=11.1%, Backbone=LLaVA-NeXT2025.11 | 57.6 | 94 | |
| SPprunerToken Budget (Retained Tokens)=192, Reduction Ratio (↓ %)=66.7%, Base Model (LLaVA-1.5-7B)=LLaVA-1.5-7B2026.05 | 57.5 | 99.4 | |
| DARTToken Budget (Retained Tokens)=192, Reduction Ratio (↓ %)=66.7%, Base Model (LLaVA-1.5-7B)=LLaVA-1.5-7B2026.05 | 57.3 | 99 | |
| VisionZipToken Budget (Retained Tokens)=192, Reduction Ratio (↓ %)=66.7%, Base Model (LLaVA-1.5-7B)=LLaVA-1.5-7B2026.05 | 57.2 | 99 | |
| PyramidDropToken Budget (Retained Tokens)=192, Reduction Ratio (↓ %)=66.7%, Base Model (LLaVA-1.5-7B)=LLaVA-1.5-7B2026.05 | 56.9 | 98.7 | |
| SPprunerToken Budget (Retained Tokens)=128, Reduction Ratio (↓ %)=77.8%, Base Model (LLaVA-1.5-7B)=LLaVA-1.5-7B2026.05 | 56.9 | 98.5 | |
| PyramidDropModel=LLaVA-Next-7B, Token Budget=320, Reduction Ratio=88.9%2026.05 | 56.9 | 90.1 | |
| VisionZipToken Budget (Retained Tokens)=128, Reduction Ratio (↓ %)=77.8%, Base Model (LLaVA-1.5-7B)=LLaVA-1.5-7B2026.05 | 56.8 | 97.6 | |
| OC-VTPTokens=160, Retaining Ratio=5.6%, Individual Training=true, Backbone=LLaVA-NeXT2025.11 | 56.8 | 92.7 | |
| PyramidDropToken Budget (Retained Tokens)=128, Reduction Ratio (↓ %)=77.8%, Base Model (LLaVA-1.5-7B)=LLaVA-1.5-7B2026.05 | 56.7 | 96.5 | |
| DivPruneToken Budget (Retained Tokens)=192, Reduction Ratio (↓ %)=66.7%, Base Model (LLaVA-1.5-7B)=LLaVA-1.5-7B2026.05 | 56.6 | 98.8 | |
| DARTToken Budget (Retained Tokens)=128, Reduction Ratio (↓ %)=77.8%, Base Model (LLaVA-1.5-7B)=LLaVA-1.5-7B2026.05 | 56.6 | 97.2 | |
| DivPruneModel=LLaVA-Next-7B, Token Budget=320, Reduction Ratio=88.9%2026.05 | 56.2 | 96.1 | |
| DivPruneToken Budget (Retained Tokens)=128, Reduction Ratio (↓ %)=77.8%, Base Model (LLaVA-1.5-7B)=LLaVA-1.5-7B2026.05 | 56 | 97.4 | |
| FastVModel=LLaVA-Next-7B, Token Budget=320, Reduction Ratio=88.9%2026.05 | 55.7 | 86.8 | |
| OC-VTPTokens=160, Retaining Ratio=5.6%, Backbone=LLaVA-NeXT2025.11 | 55.7 | 90.9 | |
| VisionZipToken Budget (Retained Tokens)=64, Reduction Ratio (↓ %)=88.9%, Base Model (LLaVA-1.5-7B)=LLaVA-1.5-7B2026.05 | 55.6 | 90.4 | |
| PACTToken Budget (Retained Tokens)=192, Reduction Ratio (↓ %)=66.7%, Base Model (LLaVA-1.5-7B)=LLaVA-1.5-7B2026.05 | 55.5 | 98.1 | |
| VisionZipTokens=160, Retaining Ratio=5.6%, Backbone=LLaVA-NeXT2025.11 | 55.5 | 90.5 | |
| SPprunerToken Budget (Retained Tokens)=64, Reduction Ratio (↓ %)=88.9%, Base Model (LLaVA-1.5-7B)=LLaVA-1.5-7B2026.05 | 55.4 | 95.2 | |
| PACTToken Budget (Retained Tokens)=128, Reduction Ratio (↓ %)=77.8%, Base Model (LLaVA-1.5-7B)=LLaVA-1.5-7B2026.05 | 54.9 | 96.9 | |
| DivPruneToken Budget (Retained Tokens)=64, Reduction Ratio (↓ %)=88.9%, Base Model (LLaVA-1.5-7B)=LLaVA-1.5-7B2026.05 | 54.6 | 94.4 | |
| DARTToken Budget (Retained Tokens)=64, Reduction Ratio (↓ %)=88.9%, Base Model (LLaVA-1.5-7B)=LLaVA-1.5-7B2026.05 | 54.1 | 93.4 | |
| PyramidDropToken Budget (Retained Tokens)=64, Reduction Ratio (↓ %)=88.9%, Base Model (LLaVA-1.5-7B)=LLaVA-1.5-7B2026.05 | 53.7 | 88.6 | |
| PACTToken Budget (Retained Tokens)=64, Reduction Ratio (↓ %)=88.9%, Base Model (LLaVA-1.5-7B)=LLaVA-1.5-7B2026.05 | 53.2 | 94 | |
| FastVToken Budget (Retained Tokens)=192, Reduction Ratio (↓ %)=66.7%, Base Model (LLaVA-1.5-7B)=LLaVA-1.5-7B2026.05 | 52.5 | 90 | |
| HiPruneTokens=160, Retaining Ratio=5.6%, Backbone=LLaVA-NeXT2025.11 | 51.2 | 83.5 | |
| PRISMSample Count=57K2025.02 | 50.8 | — | |
| FastVToken Budget (Retained Tokens)=128, Reduction Ratio (↓ %)=77.8%, Base Model (LLaVA-1.5-7B)=LLaVA-1.5-7B2026.05 | 50.6 | 82.9 | |
| Full DataSample Count=186K2025.02 | 50.4 | — | |
| FastVToken Budget (Retained Tokens)=64, Reduction Ratio (↓ %)=88.9%, Base Model (LLaVA-1.5-7B)=LLaVA-1.5-7B2026.05 | 47.8 | 74 | |
| RandomSample Count=57K2025.02 | 46.9 | — |