Text-based Visual Question Answering on TextVQA (accuracy, relative accuracy)
81.58AccuracyFP16
Evaluation results
| Method | Links | ||
|---|---|---|---|
| FP16Backbone=Qwen3-VL-8B2026.06 | 81.58 | — | |
| VanillaBackbone=Qwen2-VL-7B, Token Retention Ratio=100.0%2026.06 | 81.3 | — | |
| CoLTSize=8B, Backbone=Qwen3-VL-8B2026.06 | 81.3 | — | |
| SharQBackbone=Qwen3-VL-8B2026.06 | 80.57 | — | |
| InternVL3Size=8B2026.06 | 80.2 | — | |
| NVFP4Backbone=Qwen3-VL-8B2026.06 | 79.47 | — | |
| Qwen3-VL (Textual reasoning)Size=8B2026.06 | 75.2 | — | |
| Qwen3-VL (Direct answer)Size=8B2026.06 | 74.3 | — | |
| SIM-CoTSize=8B, Backbone=Qwen3-VL-8B2026.06 | 70.2 | — | |
| LVRSize=7B2026.06 | 70.2 | — | |
| SoftCoTSize=8B, Backbone=Qwen3-VL-8B2026.06 | 68.4 | — | |
| LaCoTSize=7B2026.06 | 67.6 | — | |
| CODISize=8B, Backbone=Qwen3-VL-8B2026.06 | 67.2 | — | |
| Multimodal CoTSize=8B, Backbone=Qwen3-VL-8B2026.06 | 66.1 | — | |
| ICoTSize=8B, Backbone=Qwen3-VL-8B2026.06 | 65.6 | — | |
| SCOPEBackbone=Qwen2-VL-7B, Token Retention Ratio=5%2026.06 | 65.2 | — | |
| DCP-PruneBackbone=Qwen2-VL-7B, Token Retention Ratio=5%2026.06 | 64.6 | — | |
| CCoTSize=8B, Backbone=Qwen3-VL-8B2026.06 | 63.1 | — | |
| FlowCutBackbone=Qwen2-VL-7B, Token Retention Ratio=5%2026.06 | 62.4 | — | |
| VanillaToken Retention Budget=2,880 Tokens (100%), Backbone=LLaVA-NeXT 7B2026.05 | 61.3 | 100 | |
| VanillaBase Model=LLaVA-NeXT-7B, Retained Tokens=2880, Reduction Ratio=0.0%2026.06 | 61.3 | 100 | |
| LLaVA-1.5-13BBackbone=LLaVA-1.5-13B, Retained Tokens=576 (Upper Bound)2026.06 | 61.2 | — | |
| PDropBackbone=LLaVA-1.5-13B, Retained Tokens=1282026.06 | 60.2 | — | |
| SparseVLMBackbone=LLaVA-1.5-13B, Retained Tokens=1282026.06 | 59.3 | — | |
| VisPrunerBackbone=LLaVA-1.5-13B, Retained Tokens=1282026.06 | 59.2 | — | |
| VisionZipToken Retention Budget=320 Tokens (↓ 88.9%), Backbone=LLaVA-NeXT 7B2026.05 | 58.9 | 94 | |
| VisionZipBackbone=LLaVA-1.5-13B, Retained Tokens=1282026.06 | 58.9 | — | |
| AgilePruneBackbone=LLaVA-1.5-13B, Retained Tokens=1282026.06 | 58.9 | — | |
| ZOO-PruneBackbone=LLaVA-1.5-13B, Retained Tokens=1282026.06 | 58.8 | — | |
| DARTBackbone=LLaVA-1.5-13B, Retained Tokens=1282026.06 | 58.7 | — | |
| FastVBackbone=LLaVA-1.5-13B, Retained Tokens=1282026.06 | 58.6 | — | |
| AgilePruneBackbone=LLaVA-1.5-13B, Retained Tokens=642026.06 | 58.6 | — | |
| ZOO-PruneBackbone=LLaVA-1.5-13B, Retained Tokens=642026.06 | 58.6 | — | |
| OccamTokenToken Retention Budget=320 Tokens (↓ 88.9%), Backbone=LLaVA-NeXT 7B2026.05 | 58.5 | 97.3 | |
| CDPrunerBackbone=LLaVA-1.5-13B, Retained Tokens=1282026.06 | 58.4 | — | |
| OC-VTPRetained Tokens=192, Pruning Ratio=33.3%, Training Protocol=Zero-shot2025.11 | 58.3 | 100.1 | |
| VanillaRetained Tokens=576, Pruning Ratio=100%, Training Protocol=Zero-shot2025.11 | 58.2 | 100 | |
| VanillaBackbone=LLaVA-1.5-7B, Visual Token Budget=5762026.06 | 58.2 | — | |
| ERABackbone=LLaVA-1.5-13B, Retained Tokens=1282026.06 | 58.2 | — | |
| DivPruneBackbone=LLaVA-1.5-13B, Retained Tokens=1282026.06 | 58 | — | |
| VisPrunerBackbone=LLaVA-1.5-13B, Retained Tokens=642026.06 | 57.9 | — | |
| OC-VTPRetained Tokens=192, Pruning Ratio=33.3%, Training Protocol=Benchmark-specific training2025.11 | 57.8 | 99.3 | |
| ERABackbone=LLaVA-1.5-13B, Retained Tokens=642026.06 | 57.7 | — | |
| HiPruneRetained Tokens=192, Pruning Ratio=33.3%, Training Protocol=Zero-shot2025.11 | 57.6 | 99 | |
| DARTToken Retention Budget=320 Tokens (↓ 88.9%), Backbone=LLaVA-NeXT 7B2026.05 | 57.6 | 93.8 | |
| EvoCutBackbone=LLaVA-1.5-7B, Visual Token Budget=1922026.06 | 57.6 | — | |
| CDPrunerBackbone=LLaVA-1.5-13B, Retained Tokens=642026.06 | 57.6 | — | |
| TwigVLMToken Retention Budget=320 Tokens (↓ 88.9%), Backbone=LLaVA-NeXT 7B, Params.=610M2026.05 | 57.4 | 96.3 | |
| OccamTokenToken Retention Budget=160 Tokens (↓ 94.4%), Backbone=LLaVA-NeXT 7B2026.05 | 57.4 | 95.9 | |
| VisionZipBackbone=LLaVA-1.5-13B, Retained Tokens=642026.06 | 57.4 | — | |
| DivPruneBackbone=LLaVA-1.5-13B, Retained Tokens=642026.06 | 57.4 | — | |
| VisionZipRetained Tokens=192, Pruning Ratio=33.3%, Training Protocol=Zero-shot2025.11 | 57.3 | 98.5 | |
| VisionZipBackbone=LLaVA-1.5-7B, Visual Token Budget=1922026.06 | 57.3 | — | |
| DARTBackbone=LLaVA-1.5-13B, Retained Tokens=642026.06 | 57.3 | — | |
| EvoCutBackbone=LLaVA-1.5-7B, Visual Token Budget=1282026.06 | 57.1 | — | |
| SparseVLMBackbone=LLaVA-1.5-13B, Retained Tokens=642026.06 | 57.1 | — | |
| OC-VTPRetained Tokens=128, Pruning Ratio=22.2%, Training Protocol=Benchmark-specific training2025.11 | 56.9 | 97.8 | |
| VisionZipRetained Tokens=128, Pruning Ratio=22.2%, Training Protocol=Zero-shot2025.11 | 56.8 | 97.6 | |
| OC-VTPRetained Tokens=128, Pruning Ratio=22.2%, Training Protocol=Zero-shot2025.11 | 56.8 | 97.6 | |
| VisionZipBackbone=LLaVA-1.5-7B, Visual Token Budget=1282026.06 | 56.8 | — | |
| PyramidDropRetained Tokens=128, Pruning Ratio=22.2%, Training Protocol=Zero-shot2025.11 | 56.6 | 97.3 | |
| HiPruneRetained Tokens=128, Pruning Ratio=22.2%, Training Protocol=Zero-shot2025.11 | 56.6 | 97.3 | |
| PyramidDropRetained Tokens=192, Pruning Ratio=33.3%, Training Protocol=Zero-shot2025.11 | 56.5 | 97.1 | |
| DivPruneToken Retention Budget=320 Tokens (↓ 88.9%), Backbone=LLaVA-NeXT 7B2026.05 | 56.2 | 95 | |
| VisionZipToken Retention Budget=160 Tokens (↓ 94.4%), Backbone=LLaVA-NeXT 7B2026.05 | 56.2 | 89.4 | |
| SparseVLMRetained Tokens=192, Pruning Ratio=33.3%, Training Protocol=Zero-shot2025.11 | 56.1 | 96.4 | |
| SparseVLMBackbone=LLaVA-1.5-7B, Visual Token Budget=1922026.06 | 56.1 | — | |
| PDropBackbone=LLaVA-1.5-7B, Visual Token Budget=1922026.06 | 56.1 | — | |
| PruMerge+Backbone=LLaVA-1.5-13B, Retained Tokens=1282026.06 | 56.1 | — | |
| EvoCutBackbone=LLaVA-1.5-7B, Visual Token Budget=642026.06 | 55.7 | — | |
| V2DropBackbone=LLaVA-1.5-7B, Visual Token Budget=1922026.06 | 55.6 | — | |
| VisionZipRetained Tokens=64, Pruning Ratio=11.1%, Training Protocol=Zero-shot2025.11 | 55.5 | 95.4 | |
| VisionZipBackbone=LLaVA-1.5-7B, Visual Token Budget=642026.06 | 55.5 | — | |
| OC-VTPRetained Tokens=64, Pruning Ratio=11.1%, Training Protocol=Benchmark-specific training2025.11 | 55.4 | 95.2 | |
| PDropBackbone=LLaVA-1.5-13B, Retained Tokens=642026.06 | 55.3 | — | |
| OC-VTPRetained Tokens=64, Pruning Ratio=11.1%, Training Protocol=Zero-shot2025.11 | 55 | 94.5 | |
| TRIMBackbone=LLaVA-1.5-13B, Retained Tokens=1282026.06 | 55 | — | |
| SparseVLMRetained Tokens=128, Pruning Ratio=22.2%, Training Protocol=Zero-shot2025.11 | 54.9 | 94.3 | |
| HiPruneRetained Tokens=64, Pruning Ratio=11.1%, Training Protocol=Zero-shot2025.11 | 54.9 | 94.3 | |
| DARTToken Retention Budget=160 Tokens (↓ 94.4%), Backbone=LLaVA-NeXT 7B2026.05 | 54.9 | 90.6 | |
| SparseVLMBackbone=LLaVA-1.5-7B, Visual Token Budget=1282026.06 | 54.9 | — | |
| PDropBackbone=LLaVA-1.5-7B, Visual Token Budget=1282026.06 | 54.8 | — | |
| ApETBackbone=LLaVA-1.5-7B, Visual Token Budget=1922026.06 | 54.4 | — | |
| PruMerge+Backbone=LLaVA-1.5-13B, Retained Tokens=642026.06 | 54.4 | — | |
| DivPruneToken Retention Budget=160 Tokens (↓ 94.4%), Backbone=LLaVA-NeXT 7B2026.05 | 54.1 | 91.7 | |
| ApETBackbone=LLaVA-1.5-7B, Visual Token Budget=1282026.06 | 53.9 | — | |
| V2DropBackbone=LLaVA-1.5-7B, Visual Token Budget=1282026.06 | 53.8 | — | |
| FastVBackbone=LLaVA-1.5-13B, Retained Tokens=642026.06 | 53.4 | — | |
| ApETBackbone=LLaVA-1.5-7B, Visual Token Budget=642026.06 | 53 | — | |
| FastVECRetained Tokens=192, Pruning Ratio=33.3%, Training Protocol=Zero-shot2025.11 | 52.5 | 90.2 | |
| FastVBackbone=LLaVA-1.5-7B, Visual Token Budget=1922026.06 | 52.5 | — | |
| TRIMBackbone=LLaVA-1.5-13B, Retained Tokens=642026.06 | 52 | — | |
| SparseVLMRetained Tokens=64, Pruning Ratio=11.1%, Training Protocol=Zero-shot2025.11 | 51.8 | 89 | |
| OccamTokenToken Retention Budget=40 Tokens (↓ 98.6%), Backbone=LLaVA-NeXT 7B2026.05 | 51.8 | 93.8 | |
| SparseVLMBackbone=LLaVA-1.5-7B, Visual Token Budget=642026.06 | 51.8 | — | |
| V2DropBackbone=LLaVA-1.5-7B, Visual Token Budget=642026.06 | 51.8 | — | |
| OC-VTPRetained Tokens=50.7, Pruning Ratio=8.8%, Training Protocol=Zero-shot2025.11 | 51.5 | 88.5 | |
| FastVECRetained Tokens=128, Pruning Ratio=22.2%, Training Protocol=Zero-shot2025.11 | 50.6 | 86.9 | |
| PyramidDropRetained Tokens=64, Pruning Ratio=11.1%, Training Protocol=Zero-shot2025.11 | 50.6 | 86.9 | |
| FastVBackbone=LLaVA-1.5-7B, Visual Token Budget=1282026.06 | 50.6 | — |