Multimodal Science Question Answering on ScienceQA IMG
95.78AccuracyFP16
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| FP16Backbone=Molmo 7B-O2026.04 | 95.78 | — | — | — | |
| WSVD-noQBackbone=Molmo 7B-O, Singular-value ratio (ρ1)=90%2026.04 | 95.59 | — | — | — | |
| QSVD-noQBackbone=Molmo 7B-O, Singular-value ratio (ρ1)=90%2026.04 | 95.54 | — | — | — | |
| WSVD-noQBackbone=Molmo 7B-O, Singular-value ratio (ρ1)=80%2026.04 | 95.49 | — | — | — | |
| WSVD-noQBackbone=Molmo 7B-O, Singular-value ratio (ρ1)=70%2026.04 | 95.34 | — | — | — | |
| ASVDBackbone=Molmo 7B-O, Singular-value ratio (ρ1)=80%2026.04 | 95.19 | — | — | — | |
| SVD-LLMBackbone=Molmo 7B-O, Singular-value ratio (ρ1)=90%2026.04 | 95.09 | — | — | — | |
| SVD-LLMBackbone=Molmo 7B-O, Singular-value ratio (ρ1)=80%2026.04 | 95.09 | — | — | — | |
| SVD-LLMBackbone=Molmo 7B-O, Singular-value ratio (ρ1)=60%2026.04 | 95.09 | — | — | — | |
| WSVD-noQBackbone=Molmo 7B-O, Singular-value ratio (ρ1)=60%2026.04 | 95.09 | — | — | — | |
| ASVDBackbone=Molmo 7B-O, Singular-value ratio (ρ1)=90%2026.04 | 94.99 | — | — | — | |
| SVD-LLMBackbone=Molmo 7B-O, Singular-value ratio (ρ1)=70%2026.04 | 94.99 | — | — | — | |
| QSVD-noQBackbone=Molmo 7B-O, Singular-value ratio (ρ1)=80%2026.04 | 94.99 | — | — | — | |
| QSVD-noQBackbone=Molmo 7B-O, Singular-value ratio (ρ1)=70%2026.04 | 94.59 | — | — | — | |
| ASVDBackbone=Molmo 7B-O, Singular-value ratio (ρ1)=70%2026.04 | 94.1 | — | — | — | |
| QSVD-noQBackbone=Molmo 7B-O, Singular-value ratio (ρ1)=60%2026.04 | 93.85 | — | — | — | |
| ASVDBackbone=Molmo 7B-O, Singular-value ratio (ρ1)=60%2026.04 | 93.06 | — | — | — | |
| FP16 BaselineBackbone=SmolVLM 2B, Precision=FP162026.04 | 84.53 | 76.53 | — | — | |
| QSVD-noQBackbone=SmolVLM 2B, Parameter Ratio (ρ1)=90%, Precision=FP162026.04 | 77 | 55.64 | — | — | |
| WSVD-noQBackbone=SmolVLM 2B, Parameter Ratio (ρ1)=90%, Precision=FP162026.04 | 76.3 | 65.42 | — | — | |
| WSVD-noQBackbone=LLaVA-Next 13B, Parameter Ratio (ρ1)=70%, Precision=FP162026.04 | 73.57 | 72.17 | — | — | |
| WSVD-noQBackbone=LLaVA-Next 13B, Parameter Ratio (ρ1)=60%, Precision=FP162026.04 | 73.48 | 72.17 | — | — | |
| Ours (PSP & VRG)Model=LaViDa, Generation length L / step T=256/1282026.04 | 73.4 | — | — | — | |
| WSVD-noQBackbone=LLaVA-Next 13B, Parameter Ratio (ρ1)=50%, Precision=FP162026.04 | 73.28 | 72.17 | — | — | |
| FP16 BaselineBackbone=LLaVA-Next 13B, Precision=FP162026.04 | 73.23 | 72.27 | — | — | |
| FP16Model Variant=LLaVA-Next 13B2026.05 | 73.23 | — | — | — | |
| WSVDModel Variant=LLaVA-Next 13B2026.05 | 73.08 | — | — | — | |
| WSVD-noQBackbone=LLaVA-Next 13B, Parameter Ratio (ρ1)=80%, Precision=FP162026.04 | 72.98 | 72.17 | — | — | |
| WSVD-noQBackbone=LLaVA-Next 13B, Parameter Ratio (ρ1)=90%, Precision=FP162026.04 | 72.88 | 72.17 | — | — | |
| Ours (PSP & VRG)Model=LaViDa, Generation length L / step T=64/322026.04 | 72.8 | — | — | — | |
| Ours (PSP)Model=LaViDa, Generation length L / step T=256/1282026.04 | 72.7 | — | — | — | |
| Ours (PSP & VRG)Model=LaViDa, Generation length L / step T=128/642026.04 | 72.7 | — | — | — | |
| SVD-LLMBackbone=LLaVA-Next 13B, Parameter Ratio (ρ1)=90%, Precision=FP162026.04 | 72.53 | 70.94 | — | — | |
| Ours (PSP)Model=LaViDa, Generation length L / step T=128/642026.04 | 72.5 | — | — | — | |
| EntropyModel=LaViDa, Generation length L / step T=256/1282026.04 | 72.4 | — | — | — | |
| CCoTModel=LaViDa, Generation length L / step T=256/1282026.04 | 72.3 | — | — | — | |
| SVD-LLMBackbone=LLaVA-Next 13B, Parameter Ratio (ρ1)=80%, Precision=FP162026.04 | 72.24 | 70.94 | — | — | |
| Low-confModel=LaViDa, Generation length L / step T=256/1282026.04 | 72.2 | — | — | — | |
| QSVD-noQBackbone=LLaVA-Next 13B, Parameter Ratio (ρ1)=80%, Precision=FP162026.04 | 72.14 | 71.44 | — | — | |
| QSVD-noQBackbone=LLaVA-Next 13B, Parameter Ratio (ρ1)=60%, Precision=FP162026.04 | 72.14 | 71.44 | — | — | |
| LASERModel Variant=LLaVA-Next 13B2026.05 | 72.11 | — | — | — | |
| Ours (PSP)Model=LaViDa, Generation length L / step T=64/322026.04 | 72 | — | — | — | |
| QSVD-noQBackbone=LLaVA-Next 13B, Parameter Ratio (ρ1)=90%, Precision=FP162026.04 | 71.94 | 71.44 | — | — | |
| MarginModel=LaViDa, Generation length L / step T=256/1282026.04 | 71.8 | — | — | — | |
| QSVD-noQBackbone=LLaVA-Next 13B, Parameter Ratio (ρ1)=50%, Precision=FP162026.04 | 71.79 | 71.44 | — | — | |
| SVD-LLMBackbone=LLaVA-Next 13B, Parameter Ratio (ρ1)=70%, Precision=FP162026.04 | 71.74 | 70.94 | — | — | |
| QSVD-noQBackbone=LLaVA-Next 13B, Parameter Ratio (ρ1)=70%, Precision=FP162026.04 | 71.74 | 71.44 | — | — | |
| WSVD-noQBackbone=SmolVLM 2B, Parameter Ratio (ρ1)=80%, Precision=FP162026.04 | 71.74 | 65.42 | — | — | |
| DDCoTModel=LaViDa, Generation length L / step T=256/1282026.04 | 71.7 | — | — | — | |
| MarginModel=LaViDa, Generation length L / step T=128/642026.04 | 71.5 | — | — | — | |
| ASVDBackbone=LLaVA-Next 13B, Parameter Ratio (ρ1)=70%, Precision=FP162026.04 | 71.44 | 70.43 | — | — | |
| EntropyModel=LaViDa, Generation length L / step T=128/642026.04 | 71.4 | — | — | — | |
| ASVDBackbone=LLaVA-Next 13B, Parameter Ratio (ρ1)=60%, Precision=FP162026.04 | 71.38 | 70.43 | — | — | |
| ASVDBackbone=LLaVA-Next 13B, Parameter Ratio (ρ1)=90%, Precision=FP162026.04 | 71.24 | 70.43 | — | — | |
| CCoTModel=LaViDa, Generation length L / step T=64/322026.04 | 71.2 | — | — | — | |
| SVD-LLMBackbone=LLaVA-Next 13B, Parameter Ratio (ρ1)=60%, Precision=FP162026.04 | 71.15 | 70.94 | — | — | |
| DDCoTModel=LaViDa, Generation length L / step T=64/322026.04 | 71.1 | — | — | — | |
| DDCoTModel=LaViDa, Generation length L / step T=128/642026.04 | 71.1 | — | — | — | |
| CCoTModel=LaViDa, Generation length L / step T=128/642026.04 | 71.1 | — | — | — | |
| Low-confModel=LaViDa, Generation length L / step T=128/642026.04 | 71.1 | — | — | — | |
| MarginModel=LaViDa, Generation length L / step T=64/322026.04 | 71 | — | — | — | |
| Low-confModel=LaViDa, Generation length L / step T=64/322026.04 | 71 | — | — | — | |
| EntropyModel=LaViDa, Generation length L / step T=64/322026.04 | 70.9 | — | — | — | |
| ASVDBackbone=LLaVA-Next 13B, Parameter Ratio (ρ1)=80%, Precision=FP162026.04 | 70.6 | 70.43 | — | — | |
| SVD-LLMBackbone=LLaVA-Next 13B, Parameter Ratio (ρ1)=50%, Precision=FP162026.04 | 70.55 | 70.94 | — | — | |
| QSVDModel Variant=LLaVA-Next 13B2026.05 | 70.43 | — | — | — | |
| DuQuantModel Variant=LLaVA-Next 13B2026.05 | 70.2 | — | — | — | |
| TUR-DPO2026.04 | 69.9 | — | 61.3 | 74.1 | |
| ASVDBackbone=LLaVA-Next 13B, Parameter Ratio (ρ1)=50%, Precision=FP162026.04 | 69.81 | 70.43 | — | — | |
| FP16Model Variant=LLaVA-Next 7B2026.05 | 69.6 | — | — | — | |
| WSVD-noQBackbone=Qwen-VL 7B, Singular-value ratio (ρ1)=90%2026.04 | 68.77 | — | — | — | |
| LASERModel Variant=LLaVA-Next 7B2026.05 | 68.37 | — | — | — | |
| FP16Backbone=Qwen-VL 7B2026.04 | 68.32 | — | — | — | |
| QSVD-noQBackbone=LLaVA-v1.5 7B, Parameter Ratio (ρ1)=80%, Precision=FP162026.04 | 68.22 | 62.64 | — | — | |
| WSVD-noQBackbone=LLaVA-v1.5 7B, Parameter Ratio (ρ1)=90%, Precision=FP162026.04 | 68.17 | 63.54 | — | — | |
| WSVD-noQBackbone=Qwen-VL 7B, Singular-value ratio (ρ1)=80%2026.04 | 68.12 | — | — | — | |
| FP16Model Variant=LLaVA-v1.5 7B2026.05 | 68.1 | — | — | — | |
| FP16 BaselineBackbone=LLaVA-v1.5 7B, Precision=FP162026.04 | 68.01 | 64.1 | — | — | |
| QSVD-noQBackbone=LLaVA-v1.5 7B, Parameter Ratio (ρ1)=90%, Precision=FP162026.04 | 67.72 | 62.64 | — | — | |
| WSVD-noQBackbone=LLaVA-v1.5 7B, Parameter Ratio (ρ1)=80%, Precision=FP162026.04 | 67.72 | 63.54 | — | — | |
| WSVD-noQBackbone=LLaVA-v1.5 7B, Parameter Ratio (ρ1)=70%, Precision=FP162026.04 | 67.28 | 63.54 | — | — | |
| WSVD-noQBackbone=Qwen-VL 7B, Singular-value ratio (ρ1)=70%2026.04 | 67.23 | — | — | — | |
| QSVD-noQBackbone=LLaVA-v1.5 7B, Parameter Ratio (ρ1)=70%, Precision=FP162026.04 | 67.08 | 62.64 | — | — | |
| WSVDModel Variant=LLaVA-Next 7B2026.05 | 66.94 | — | — | — | |
| QSVD-noQBackbone=Qwen-VL 7B, Singular-value ratio (ρ1)=90%2026.04 | 66.78 | — | — | — | |
| QSVD-noQBackbone=Qwen-VL 7B, Singular-value ratio (ρ1)=70%2026.04 | 66.68 | — | — | — | |
| DuQuantModel Variant=LLaVA-Next 7B2026.05 | 66.34 | — | — | — | |
| DPO2026.04 | 66.3 | — | 56.4 | 69.2 | |
| QSVD-noQBackbone=Qwen-VL 7B, Singular-value ratio (ρ1)=80%2026.04 | 66.24 | — | — | — | |
| LASERModel Variant=LLaVA-v1.5 7B2026.05 | 66.14 | — | — | — | |
| QSVDModel Variant=LLaVA-Next 7B2026.05 | 66.1 | — | — | — | |
| WSVD-noQBackbone=LLaVA-v1.5 7B, Parameter Ratio (ρ1)=60%, Precision=FP162026.04 | 65.89 | 63.54 | — | — | |
| QSVDModel Variant=LLaVA-v1.5 7B2026.05 | 65.61 | — | — | — | |
| WSVD-noQBackbone=LLaVA-v1.5 7B, Parameter Ratio (ρ1)=50%, Precision=FP162026.04 | 65.49 | 63.54 | — | — | |
| WSVD-noQBackbone=Qwen-VL 7B, Singular-value ratio (ρ1)=60%2026.04 | 65.49 | — | — | — | |
| SVD-LLMBackbone=LLaVA-v1.5 7B, Parameter Ratio (ρ1)=90%, Precision=FP162026.04 | 65.44 | 58.47 | — | — | |
| SVD-LLMBackbone=Qwen-VL 7B, Singular-value ratio (ρ1)=90%2026.04 | 65.29 | — | — | — | |
| SVD-LLMBackbone=Qwen-VL 7B, Singular-value ratio (ρ1)=80%2026.04 | 65.29 | — | — | — | |
| SVD-LLMBackbone=Qwen-VL 7B, Singular-value ratio (ρ1)=60%2026.04 | 65.29 | — | — | — | |
| QVLMModel Variant=LLaVA-Next 13B2026.05 | 65.28 | — | — | — |