Multimodal Reasoning on MMBench CN
82AccuracyInstruct
Evaluation Results
| Method | Links | |
|---|---|---|
| InstructModel Scale=14B, Base Model=InternVL32026.03 | 82 | |
| IPOModel Scale=14B, Base Model=InternVL32026.03 | 82 | |
| IPOModel Scale=8B, Base Model=InternVL32026.03 | 82 | |
| β-DPOModel Scale=14B, Base Model=InternVL32026.03 | 82 | |
| VanillaToken Budget=1296 Tokens (Upper Bound), Backbone=Qwen-2.5-VL-7B2026.02 | 81.8 | |
| DPOModel Scale=14B, Base Model=InternVL32026.03 | 81.6 | |
| InternVL2-8B-FTRP fine-tuning=None2024.08 | 81.5 | |
| ACPOModel Scale=8B, Base Model=InternVL32026.03 | 81.5 | |
| InternVL2-8B + RPRP fine-tuning=Object Replacement Data2024.08 | 81.4 | |
| InstructModel Scale=8B, Base Model=InternVL32026.03 | 81.2 | |
| DPO-ShiftModel Scale=14B, Base Model=InternVL32026.03 | 81 | |
| ACPOModel Scale=14B, Base Model=InternVL32026.03 | 81 | |
| PIO-FVLMToken Budget=Retain 33.30% Tokens, Venue=Ours, Backbone=Qwen-2.5-VL-7B2026.02 | 80.5 | |
| DARTToken Budget=Retain 33.30% Tokens, Venue=EMNLP'25, Backbone=Qwen-2.5-VL-7B2026.02 | 79 | |
| SimPOModel Scale=14B, Base Model=InternVL32026.03 | 79 | |
| PIO-FVLMToken Budget=Retain 22.2% Tokens, Venue=Ours, Backbone=Qwen-2.5-VL-7B2026.02 | 78.1 | |
| DPOModel Scale=8B, Base Model=InternVL32026.03 | 78 | |
| DPO-ShiftModel Scale=8B, Base Model=InternVL32026.03 | 78 | |
| DARTToken Budget=Retain 22.2% Tokens, Venue=EMNLP'25, Backbone=Qwen-2.5-VL-7B2026.02 | 76.5 | |
| Qwen3-VL-2BRatio=100%2026.05 | 75.93 | |
| F3ARatio=40%2026.05 | 75.58 | |
| DivPruneRatio=60%2026.05 | 75.51 | |
| PIO-FVLMToken Budget=Retain 11.1% Tokens, Venue=Ours, Backbone=Qwen-2.5-VL-7B2026.02 | 75.5 | |
| F3ARatio=60%2026.05 | 75.49 | |
| VisionZipRatio=60%2026.05 | 75.44 | |
| CDPrunerRatio=60%2026.05 | 75.33 | |
| SimPOModel Scale=8B, Base Model=InternVL32026.03 | 75 | |
| CDPrunerRatio=40%2026.05 | 74.79 | |
| β-DPOModel Scale=8B, Base Model=InternVL32026.03 | 74.5 | |
| VisionZipRatio=40%2026.05 | 74.33 | |
| FastVRatio=60%2026.05 | 74.19 | |
| DivPruneRatio=40%2026.05 | 73.94 | |
| F3ARatio=20%2026.05 | 73.87 | |
| FastVRatio=40%2026.05 | 73.2 | |
| CDPrunerRatio=20%2026.05 | 72.41 | |
| DARTToken Budget=Retain 11.1% Tokens, Venue=EMNLP'25, Backbone=Qwen-2.5-VL-7B2026.02 | 71.5 | |
| DivPruneRatio=20%2026.05 | 71.19 | |
| VisionZipRatio=20%2026.05 | 70.87 | |
| FastVRatio=20%2026.05 | 70.71 | |
| LLaVA-1.5-13B + MMFuserLLM=Vicuna-13B2024.10 | 63.8 | |
| LLaVA-1.5-13BLLM=Vicuna-13B2024.10 | 63.6 | |
| LLaVABackbone=LLaVA-NeXT-7B, Tokens Retained=2880, Retention Ratio=100%2025.08 | 60.6 | |
| VanillaVisual Token Budget=2880, Token Retention Ratio=100%2026.06 | 60.6 | |
| LLaVA-1.5-7B + MMFuserLLM=Vicuna-7B2024.10 | 60.1 | |
| LLaVA-1.5-7B + RPRP fine-tuning=Object Replacement Data2024.08 | 59.8 | |
| MGM-7B + RPRP fine-tuning=Object Replacement Data2024.08 | 59.6 | |
| HiPruneBackbone=LLaVA-NeXT-7B, Tokens Retained=640, Retention Ratio=22.2%2025.08 | 59.3 | |
| HiPrune++Backbone=LLaVA-NeXT-7B, Tokens Retained=640, Retention Ratio=22.2%2025.08 | 59.1 | |
| MGM-7BRP fine-tuning=None2024.08 | 58.9 | |
| SparseVLMBackbone=LLaVA-NeXT-7B, Tokens Retained=640, Retention Ratio=22.2%2025.08 | 58.6 | |
| LLaVA-1.5-7BLLM=Vicuna-7B2024.10 | 58.3 | |
| LLaVA-1.5-7BRP fine-tuning=None2024.08 | 58.3 | |
| VanillaBackbone=LLaVA-1.5-7B, Retained Tokens=576, Compression Ratio=100%2026.04 | 58.1 | |
| VisionZipBackbone=LLaVA-NeXT-7B, Tokens Retained=640, Retention Ratio=22.2%2025.08 | 58.1 | |
| FastVBackbone=LLaVA-1.5-7B, Retained Tokens=192, Compression Ratio=66.7%2026.04 | 57.5 | |
| DARTBackbone=LLaVA-1.5-7B, Retained Tokens=128, Compression Ratio=77.8%2026.04 | 57.5 | |
| VisionZipBackbone=LLaVA-1.5-7B, Retained Tokens=192, Compression Ratio=66.7%2026.04 | 57.4 | |
| HiPrune++Backbone=LLaVA-NeXT-7B, Tokens Retained=320, Retention Ratio=11.1%2025.08 | 57.4 | |
| EvoCompBackbone=LLaVA-1.5-7B, Retained Tokens=192, Compression Ratio=66.7%, Number of compressor layers (l)=22026.04 | 57.2 | |
| EvoCompBackbone=LLaVA-1.5-7B, Retained Tokens=128, Compression Ratio=77.8%, Number of compressor layers (l)=22026.04 | 57.1 | |
| DARTBackbone=LLaVA-1.5-7B, Retained Tokens=192, Compression Ratio=66.7%2026.04 | 57 | |
| HiREDBackbone=LLaVA-NeXT-7B, Tokens Retained=640, Retention Ratio=22.2%2025.08 | 57 | |
| HiPruneBackbone=LLaVA-NeXT-7B, Tokens Retained=320, Retention Ratio=11.1%2025.08 | 57 | |
| VisionZipBackbone=LLaVA-1.5-7B, Retained Tokens=128, Compression Ratio=77.8%2026.04 | 56.9 | |
| PyramidDropBackbone=LLaVA-1.5-7B, Retained Tokens=192, Compression Ratio=66.7%2026.04 | 56.8 | |
| Qwen-VL-ChatLLM=Qwen-7B2024.10 | 56.7 | |
| SparseVLMBackbone=LLaVA-NeXT-7B, Tokens Retained=320, Retention Ratio=11.1%2025.08 | 56.7 | |
| FastVBackbone=LLaVA-1.5-7B, Retained Tokens=128, Compression Ratio=77.8%2026.04 | 56.6 | |
| PyramidDropBackbone=LLaVA-1.5-7B, Retained Tokens=128, Compression Ratio=77.8%2026.04 | 56.6 | |
| DivPruneBackbone=LLaVA-NeXT-7B, Tokens Retained=640, Retention Ratio=22.2%2025.08 | 56.4 | |
| HiREDBackbone=LLaVA-NeXT-7B, Tokens Retained=320, Retention Ratio=11.1%2025.08 | 56.4 | |
| PDropVisual Token Budget=320, Token Retention Ratio=11.1%2026.06 | 56.2 | |
| EvoCompBackbone=LLaVA-1.5-7B, Retained Tokens=192, Compression Ratio=66.7%, Number of compressor layers (l)=02026.04 | 56.1 | |
| EvoCompBackbone=LLaVA-1.5-7B, Retained Tokens=128, Compression Ratio=77.8%, Number of compressor layers (l)=02026.04 | 56.1 | |
| VisPrunerBackbone=LLaVA-NeXT-7B, Tokens Retained=640, Retention Ratio=22.2%2025.08 | 56 | |
| MustDropBackbone=LLaVA-1.5-7B, Retained Tokens=192, Compression Ratio=66.7%2026.04 | 55.8 | |
| TRIMBackbone=LLaVA-NeXT-7B, Tokens Retained=640, Retention Ratio=22.2%2025.08 | 55.8 | |
| VisionZipBackbone=LLaVA-NeXT-7B, Tokens Retained=320, Retention Ratio=11.1%2025.08 | 55.6 | |
| VisPrunerBackbone=LLaVA-NeXT-7B, Tokens Retained=320, Retention Ratio=11.1%2025.08 | 55.4 | |
| VisionZipBackbone=LLaVA-1.5-7B, Retained Tokens=64, Compression Ratio=88.9%2026.04 | 55.3 | |
| PruMergeVisual Token Budget=320, Token Retention Ratio=11.1%2026.06 | 55.3 | |
| MustDropBackbone=LLaVA-1.5-7B, Retained Tokens=128, Compression Ratio=77.8%2026.04 | 55.2 | |
| PDropBackbone=LLaVA-NeXT-7B, Tokens Retained=640, Retention Ratio=22.2%2025.08 | 55.2 | |
| DivPruneBackbone=LLaVA-NeXT-7B, Tokens Retained=320, Retention Ratio=11.1%2025.08 | 55.2 | |
| EvoCompBackbone=LLaVA-1.5-7B, Retained Tokens=64, Compression Ratio=88.9%, Number of compressor layers (l)=02026.04 | 55.1 | |
| HiREDBackbone=LLaVA-1.5-7B, Retained Tokens=192, Compression Ratio=66.7%2026.04 | 54.7 | |
| SparseVLMVisual Token Budget=320, Token Retention Ratio=11.1%2026.06 | 54.5 | |
| PriorTRVisual Token Budget=320, Token Retention Ratio=11.1%2026.06 | 54.4 | |
| SparseVLMBackbone=LLaVA-1.5-7B, Retained Tokens=192, Compression Ratio=66.7%2026.04 | 53.7 | |
| EvoCompBackbone=LLaVA-1.5-7B, Retained Tokens=64, Compression Ratio=88.9%, Number of compressor layers (l)=22026.04 | 53.7 | |
| DARTBackbone=LLaVA-1.5-7B, Retained Tokens=192, Compression Ratio=66.7%, Number of compressor layers (l)=02026.04 | 53.6 | |
| HiREDBackbone=LLaVA-1.5-7B, Retained Tokens=128, Compression Ratio=77.8%2026.04 | 53.6 | |
| FastVBackbone=LLaVA-NeXT-7B, Tokens Retained=640, Retention Ratio=22.2%2025.08 | 53.5 | |
| DARTBackbone=LLaVA-1.5-7B, Retained Tokens=64, Compression Ratio=88.9%2026.04 | 53.2 | |
| MustDropBackbone=LLaVA-1.5-7B, Retained Tokens=64, Compression Ratio=88.9%2026.04 | 53.1 | |
| PruMergeBackbone=LLaVA-1.5-7B, Retained Tokens=192, Compression Ratio=66.7%2026.04 | 52.9 | |
| FastVBackbone=LLaVA-1.5-7B, Retained Tokens=64, Compression Ratio=88.9%2026.04 | 52.8 | |
| DivPruneBackbone=LLaVA-NeXT-7B, Tokens Retained=160, Retention Ratio=5.6%2025.08 | 52.3 | |
| DARTBackbone=LLaVA-1.5-7B, Retained Tokens=128, Compression Ratio=77.8%, Number of compressor layers (l)=02026.04 | 51.9 | |
| FastVVisual Token Budget=320, Token Retention Ratio=11.1%2026.06 | 51.9 |