Visual Question Answering on ScienceQA (test)
98.2AccuracyIVC-Prune
Evaluation results
| Method | Links | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| IVC-PruneModels=InternVL 2.5 8B, Average Tokens=50%2026.02 | 98.2 | — | — | — | — | — | — | — | — | 100 | — | — | |
| VanillaModels=InternVL 2.5 8B, Average Tokens=100%2026.02 | 98.1 | — | — | — | — | — | — | — | — | 100 | — | — | |
| PDropModels=InternVL 2.5 8B, Average Tokens=56%2026.02 | 97.9 | — | — | — | — | — | — | — | — | 99.1 | — | — | |
| FastVModels=InternVL 2.5 8B, Average Tokens=53%2026.02 | 97.4 | — | — | — | — | — | — | — | — | 98.6 | — | — | |
| VanillaModels=DeepSeek-VL2 Small-16B, Average Tokens=100%2026.02 | 96.9 | — | — | — | — | — | — | — | — | 100 | — | — | |
| PDropModels=DeepSeek-VL2 Small-16B, Average Tokens=57%2026.02 | 96.8 | — | — | — | — | — | — | — | — | 100 | — | — | |
| IVC-PruneModels=DeepSeek-VL2 Small-16B, Average Tokens=52%2026.02 | 96.8 | — | — | — | — | — | — | — | — | 100.1 | — | — | |
| FastVModels=DeepSeek-VL2 Small-16B, Average Tokens=54%2026.02 | 96.3 | — | — | — | — | — | — | — | — | 98.8 | — | — | |
| SMoLA-PaLI-X_FT (Specialist)Base model=PaLI-X, Per-task LoRA tuning=true, LoRA rank=42023.12 | 94.7 | — | — | — | — | — | — | — | — | — | — | — | |
| SOTA [58]2023.12 | 92.7 | — | — | — | — | — | — | — | — | — | — | — | |
| CogVLM-ChatLLM=Vicuna-7B, Trained during SFT stage=true2023.11 | 91.2 | — | — | — | — | — | — | — | — | — | — | — | |
| Dense BaselineModel=LLaVA-13B, Bits=FP, Tokens=2562026.04 | 91.09 | 90.99 | 94.94 | 88.18 | 90.03 | 88.45 | 90.73 | — | — | — | — | — | |
| Q-VLMModel=LLaVA-13B, Bitwidth=W8A82024.10 | 89.81 | — | — | — | — | — | — | — | — | — | 11.2 | — | |
| Dense BaselineModel=LLaVA-7B, Bits=FP, Tokens=2562026.04 | 89.6 | 89.43 | 95.61 | 85.09 | 88.61 | 87.41 | 88.15 | — | — | — | — | — | |
| Q-VLMModel=LLaVA-7B, Bitwidth=W8A82024.10 | 89.58 | — | — | — | — | — | — | — | — | — | 8.3 | — | |
| ZeroQuant-V2Model=LLaVA-13B, Bitwidth=W8A82024.10 | 89.13 | — | — | — | — | — | — | — | — | — | 12.6 | — | |
| ZeroQuant-V2Model=LLaVA-7B, Bitwidth=W8A82024.10 | 89.04 | — | — | — | — | — | — | — | — | — | 10.7 | — | |
| VanillaModels=Qwen2.5-VL 7B, Average Tokens=100%2026.02 | 88.7 | — | — | — | — | — | — | — | — | 100 | — | — | |
| Human2023.03 | 88.4 | 90.23 | 84.97 | 87.48 | 89.6 | 87.5 | 88.1 | 91.59 | 82.42 | — | — | — | |
| PDropModels=Qwen2.5-VL 7B, Average Tokens=61%2026.02 | 87.5 | — | — | — | — | — | — | — | — | 98 | — | — | |
| IVC-PruneModels=Qwen2.5-VL 7B, Average Tokens=50%2026.02 | 87 | — | — | — | — | — | — | — | — | 100.2 | — | — | |
| Unified-IO2LLM=UIO-2XXL, Trained during SFT stage=true2023.11 | 86.2 | — | — | — | — | — | — | — | — | — | — | — | |
| FastVModels=Qwen2.5-VL 7B, Average Tokens=54%2026.02 | 85.8 | — | — | — | — | — | — | — | — | 98.2 | — | — | |
| LLaMA-AdapterTuned Params=1.8M2023.03 | 85.19 | 84.37 | 88.3 | 84.36 | 83.72 | 80.32 | 86.9 | 85.83 | 84.05 | — | — | — | |
| MM-CoTTuned Params=223M2023.03 | 84.91 | 87.52 | 77.17 | 85.82 | 87.88 | 82.9 | 86.83 | 84.65 | 85.37 | — | — | — | |
| QAPrunerModel=LLaVA-13B, Bits=W4A4, Tokens=1282026.04 | 84.72 | 83.08 | 92.13 | 82.09 | 82.16 | 81.61 | 84.6 | — | — | — | — | — | |
| QVLMModel=LLaVA-13B, Bits=W4A4, Tokens=2562026.04 | 84.39 | 82.9 | 91.34 | 81.82 | 81.82 | 80.91 | 84.6 | — | — | — | — | — | |
| QAPrunerModel=LLaVA-13B, Bits=W4A4, Tokens=322026.04 | 84.27 | 82.99 | 90.33 | 82 | 82.26 | 80.66 | 84.6 | — | — | — | — | — | |
| With PruneModel=LLaVA-13B, Bits=W4A4, Tokens=1282026.04 | 84.06 | 82.95 | 89.43 | 82 | 82.06 | 80.22 | 84.6 | — | — | — | — | — | |
| GPT-4Tuned Params=0M, Chain-of-thought=true2023.03 | 83.99 | 85.48 | 72.44 | 90.27 | 82.65 | 71.49 | 92.89 | 86.66 | 79.04 | — | — | — | |
| With PruneModel=LLaVA-13B, Bits=W4A4, Tokens=322026.04 | 82.69 | 82.19 | 84.93 | 81.91 | 81.62 | 77.34 | 84.6 | — | — | — | — | — | |
| QAPrunerModel=LLaVA-7B, Bits=W4A4, Tokens=1282026.04 | 80.95 | 78.69 | 90.55 | 77.82 | 77.22 | 78.98 | 80.77 | — | — | — | — | — | |
| QAPrunerModel=LLaVA-7B, Bits=W4A4, Tokens=322026.04 | 80.85 | 79.4 | 88.08 | 78 | 78.25 | 78.78 | 80.77 | — | — | — | — | — | |
| Q-VLMModel=LLaVA-13B, Bitwidth=W4A42024.10 | 80.78 | — | — | — | — | — | — | — | — | — | 8.9 | — | |
| QVLMModel=LLaVA-7B, Bits=W4A4, Tokens=2562026.04 | 80.36 | 77.62 | 89.99 | 78.18 | 75.95 | 77.74 | 80.77 | — | — | — | — | — | |
| Q-VLMModel=LLaVA-7B, Bitwidth=W4A42024.10 | 79.79 | — | — | — | — | — | — | — | — | — | 6.1 | — | |
| With PruneModel=LLaVA-7B, Bits=W4A4, Tokens=1282026.04 | 79.75 | 77.62 | 87.63 | 77.73 | 75.9 | 76.45 | 80.77 | — | — | — | — | — | |
| Fine-Tuning (Clean)Target Model=LLaVA2026.05 | 79.5 | — | — | — | — | — | — | — | — | — | — | 0 | |
| ZeroQuant-V2Model=LLaVA-13B, Bitwidth=W4A42024.10 | 78.81 | — | — | — | — | — | — | — | — | — | 9.7 | — | |
| With PruneModel=LLaVA-7B, Bits=W4A4, Tokens=322026.04 | 78.61 | 77.58 | 82.34 | 77.73 | 75.76 | 74.07 | 80.77 | — | — | — | — | — | |
| ChatGPTTuned Params=0M, Chain-of-thought=true2023.03 | 78.31 | 78.82 | 70.98 | 83.18 | 77.37 | 67.92 | 86.13 | 80.72 | 74.03 | — | — | — | |
| LLaMA-AdapterTuned Params=1.2M, Text-only=true2023.03 | 78.31 | 79 | 73.79 | 80.55 | 78.3 | 70.35 | 83.14 | 79.77 | 75.68 | — | — | — | |
| ZeroQuant-V2Model=LLaVA-7B, Bitwidth=W4A42024.10 | 78.08 | — | — | — | — | — | — | — | — | — | 7.3 | — | |
| LLaVA-FA-3BLLM=LLaMA-3-8B, #Sample=5M, #Param=∼3B2026.01 | 77 | — | — | — | — | — | — | — | — | — | — | — | |
| MiniCPM-V-2LLM=MiniCPM-2.4B, #Sample=570M, #Param=∼3B2026.01 | 76.3 | — | — | — | — | — | — | — | — | — | — | — | |
| LLaVA-FA-7BLLM=InternLM-2-20B, #Sample=5M, #Param=≥7B2026.01 | 76 | — | — | — | — | — | — | — | — | — | — | — | |
| MiniCPM-VLLM=MiniCPM-2.4B, #Sample=570M, #Param=∼3B2026.01 | 74.4 | — | — | — | — | — | — | — | — | — | — | — | |
| UnifiedQATuned Params=223M, Chain-of-thought=true2023.03 | 74.11 | 71 | 76.04 | 78.91 | 66.42 | 66.53 | 81.81 | 77.06 | 68.82 | — | — | — | |
| Deepseek-VL-7BLLM=DLLM-7B, #Sample=2000M, #Param=≥7B2026.01 | 74 | — | — | — | — | — | — | — | — | — | — | — | |
| LLaVA-NeXTLLM=Vicuna-1.5-13B, #Sample=1.3M, #Param=≥7B2026.01 | 73.6 | — | — | — | — | — | — | — | — | — | — | — | |
| MMGUARD-BPHTarget Model=LLaVA2026.05 | 73.5 | — | — | — | — | — | — | — | — | — | — | 6 | |
| Imp-3BLLM=Phi-2-2.7B, #Sample=1.6M, #Param=∼3B2026.01 | 72.8 | — | — | — | — | — | — | — | — | — | — | — | |
| LLaVA-1.5LLM=Vicuna-13B, Trained during SFT stage=false2023.11 | 71.6 | — | — | — | — | — | — | — | — | — | — | — | |
| Bunny-3BLLM=Phi-2-2.7B, #Sample=2.7M, #Param=∼3B2026.01 | 70.9 | — | — | — | — | — | — | — | — | — | — | — | |
| SPHINX-2kLLM=LLaMA2 13B, Trained during SFT stage=false2023.11 | 70.6 | — | — | — | — | — | — | — | — | — | — | — | |
| MM-CoTTuned Params=223M, Text-only=true2023.03 | 70.53 | 71.09 | 70.75 | 69.18 | 71.16 | 65.84 | 71.57 | 71 | 69.68 | — | — | — | |
| UnifiedQATuned Params=223M2023.03 | 70.12 | 68.16 | 69.18 | 74.91 | 63.78 | 61.38 | 77.84 | 72.98 | 65 | — | — | — | |
| MobileVLMv2LLM=MLLaMA-2.7B, #Sample=3.6M, #Param=∼3B2026.01 | 70 | — | — | — | — | — | — | — | — | — | — | — | |
| VILA-3BLLM=LLaMA-2.7B, #Sample=51M, #Param=∼3B2026.01 | 69 | — | — | — | — | — | — | — | — | — | — | — | |
| Qwen-VL-ChatLLM=Qwen-7B, Trained during SFT stage=false2023.11 | 68.8 | — | — | — | — | — | — | — | — | — | — | — | |
| mPLUG-Owl2LLM=LLaMA2-7B, Trained during SFT stage=false2023.11 | 68.7 | — | — | — | — | — | — | — | — | — | — | — | |
| MoE-LLaVA-3BLLM=Phi-2-2.7B, #Sample=2.2M, #Param=∼3B2026.01 | 68.5 | — | — | — | — | — | — | — | — | — | — | — | |
| VILA-7BLLM=LLaMA-7B, #Sample=50M, #Param=≥7B2026.01 | 68.2 | — | — | — | — | — | — | — | — | — | — | — | |
| Qwen-VL-ChatLLM=Qwen-7B, #Sample=1450M, #Param=≥7B2026.01 | 68.2 | — | — | — | — | — | — | — | — | — | — | — | |
| LLaVA-FA-2BLLM=Qwen-2.5-7B, #Sample=5M, #Param=∼2B2026.01 | 68.2 | — | — | — | — | — | — | — | — | — | — | — | |
| Dense BaselineModel=LLaVA-1.5, Bits=FP, Tokens=5762026.04 | 67.79 | 66.3 | 70.75 | 68.45 | 66.23 | 65.49 | 68.5 | — | — | — | — | — | |
| LLaVA-1.5LLM=Vicuna-7B, Trained during SFT stage=false2023.11 | 66.8 | — | — | — | — | — | — | — | — | — | — | — | |
| LLaVA-1.5-7BLLM=Vicuna-1.5-7B, #Sample=1.2M, #Param=≥7B2026.01 | 66.8 | — | — | — | — | — | — | — | — | — | — | — | |
| VanillaModels=LLaVA-v1.5 7B, Average Tokens=100%2026.02 | 66.4 | — | — | — | — | — | — | — | — | 100 | — | — | |
| IVC-PruneModels=LLaVA-v1.5 7B, Average Tokens=28%2026.02 | 66.4 | — | — | — | — | — | — | — | — | 101.3 | — | — | |
| FastVModels=LLaVA-v1.5 7B, Average Tokens=30%2026.02 | 66.3 | — | — | — | — | — | — | — | — | 99.7 | — | — | |
| PDropModels=LLaVA-v1.5 7B, Average Tokens=47%2026.02 | 66.3 | — | — | — | — | — | — | — | — | 99.8 | — | — | |
| Imp-2BLLM=Qwen-1.5-1.8B, #Sample=1.6M, #Param=∼2B2026.01 | 66.1 | — | — | — | — | — | — | — | — | — | — | — | |
| CogVLMLLM=Vicuna-7B, #Sample=1500M, #Param=≥7B2026.01 | 65.6 | — | — | — | — | — | — | — | — | — | — | — | |
| Bunny-2BLLM=Qwen-1.5-1.8B, #Sample=2.7M, #Param=∼2B2026.01 | 64.6 | — | — | — | — | — | — | — | — | — | — | — | |
| DeepSeek-VL-1.3BLLM=DLLM-1.3B, #Sample=2000M, #Param=∼2B2026.01 | 64.2 | — | — | — | — | — | — | — | — | — | — | — | |
| InstructBLIPLLM=Vicuna-13B, #Sample=130M, #Param=≥7B2026.01 | 63.1 | — | — | — | — | — | — | — | — | — | — | — | |
| Mini-Gemini-2BLLM=Gemma-2B, #Sample=2.7M, #Param=∼2B2026.01 | 63.1 | — | — | — | — | — | — | — | — | — | — | — | |
| MoE-LLaVA-2BLLM=Qwen-1.5-1.8B, #Sample=2.2M, #Param=∼2B2026.01 | 63.1 | — | — | — | — | — | — | — | — | — | — | — | |
| VisualBERTTuned Params=111M2023.03 | 61.87 | 59.33 | 69.18 | 61.18 | 62.71 | 62.17 | 58.54 | 62.96 | 59.92 | — | — | — | |
| IDEFICS-InstructLLM=LLaMA-65B, Trained during SFT stage=false2023.11 | 61.8 | — | — | — | — | — | — | — | — | — | — | — | |
| QAPrunerModel=LLaVA-1.5, Bits=W4A4, Tokens=322026.04 | 61.59 | 61.06 | 66.14 | 59 | 60.61 | 62.07 | 59.16 | — | — | — | — | — | |
| LLaVA-FA-1BLLM=Qwen-2.5-3B, #Sample=5M, #Param=∼1B2026.01 | 61.3 | — | — | — | — | — | — | — | — | — | — | — | |
| QAPrunerModel=LLaVA-1.5, Bits=W4A4, Tokens=1282026.04 | 61.24 | 61.06 | 64.12 | 59.27 | 60.56 | 61.33 | 59.16 | — | — | — | — | — | |
| With PruneModel=LLaVA-1.5, Bits=W4A4, Tokens=322026.04 | 61.05 | 60.57 | 64.57 | 59.18 | 60.07 | 60.93 | 59.16 | — | — | — | — | — | |
| BLIP-2LLM=Vicuna-13B, #Sample=129M, #Param=≥7B2026.01 | 61 | — | — | — | — | — | — | — | — | — | — | — | |
| MobileVLMLLM=MLLaMA-2.7B, #Sample=1.3M, #Param=∼3B2026.01 | 61 | — | — | — | — | — | — | — | — | — | — | — | |
| With PruneModel=LLaVA-1.5, Bits=W4A4, Tokens=1282026.04 | 60.65 | 60.44 | 62.99 | 59.18 | 59.43 | 60.09 | 59.16 | — | — | — | — | — | |
| QVLMModel=LLaVA-1.5, Bits=W4A4, Tokens=5762026.04 | 60.62 | 60.12 | 64 | 58.91 | 59.38 | 60.04 | 59.16 | — | — | — | — | — | |
| InstructBLIPLLM=Vicuna-7B, Trained during SFT stage=false2023.11 | 60.5 | — | — | — | — | — | — | — | — | — | — | — | |
| MCANTuned Params=95M2023.03 | 54.54 | 56.08 | 46.23 | 58.09 | 59.43 | 51.17 | 55.4 | 51.65 | 59.72 | — | — | — | |
| CLIPModule Configuration=MuRes(V+L)2025.12 | 51.85 | — | — | — | — | — | — | — | — | — | — | — | |
| ViLTModule Configuration=MuRes(V+L)2025.12 | 49.33 | — | — | — | — | — | — | — | — | — | — | — | |
| ViLTModule Configuration=With Residual2025.12 | 47.05 | — | — | — | — | — | — | — | — | — | — | — | |
| OpenFlamingoLLM=MPT-7B, Trained during SFT stage=false2023.11 | 44.8 | — | — | — | — | — | — | — | — | — | — | — | |
| ViLTModule Configuration=Without Residual2025.12 | 44.52 | — | — | — | — | — | — | — | — | — | — | — | |
| Dual-EncoderModule Configuration=MuRes(V+L)2025.12 | 43.57 | — | — | — | — | — | — | — | — | — | — | — | |
| ViLTModule Configuration=MuRes(V)2025.12 | 42.78 | — | — | — | — | — | — | — | — | — | — | — | |
| ViLTModule Configuration=MuRes(L)2025.12 | 42.58 | — | — | — | — | — | — | — | — | — | — | — | |
| CLIPModule Configuration=MuRes(V)2025.12 | 40.75 | — | — | — | — | — | — | — | — | — | — | — |