Science Question Answering on SQA IMG (Accuracy)
70.7AccuracyID-Selection
Evaluation Results
| Method | Links | |
|---|---|---|
| ID-SelectionImportance Criterion=LLM second-layer cross-modal attention, Retain Tokens=162026.04 | 70.7 | |
| LLaVA-Next-7BToken Budget=2880 (100%)2026.04 | 70.1 | |
| SparseVLMBackbone=LLaVA-1.5-7B, Retained Tokens=642026.06 | 69.7 | |
| LLaVA-1.5-7BRetain Tokens=5762026.04 | 69.5 | |
| CDPrunerRetain Tokens=322026.04 | 69.5 | |
| LLaVA-1.5-7BBackbone=LLaVA-1.5-7B, Retained Tokens=5762026.06 | 69.5 | |
| DARTRetain Tokens=642026.04 | 69.3 | |
| CDPrunerBackbone=LLaVA-1.5-7B, Retained Tokens=322026.06 | 69.3 | |
| SPAREBackbone=LLaVA-1.5-7B, Retained Tokens=322026.06 | 69.3 | |
| DARTBackbone=LLaVA-1.5-7B, Retained Tokens=1282026.06 | 69.2 | |
| DARTBackbone=LLaVA-1.5-7B, Retained Tokens=322026.06 | 69.2 | |
| VisPrunerToken Budget=160 (94.4%)2026.04 | 69 | |
| CDPrunerBackbone=LLaVA-1.5-7B, Retained Tokens=1282026.06 | 69 | |
| DARTBackbone=LLaVA-1.5-7B, Retained Tokens=642026.06 | 68.8 | |
| SPAREBackbone=LLaVA-1.5-7B, Retained Tokens=1282026.06 | 68.6 | |
| SparseVLMBackbone=LLaVA-1.5-7B, Retained Tokens=1282026.06 | 68.5 | |
| VRCDDecoding Length=384, Forward Ratio (FR)=0.52026.05 | 68.43 | |
| VRCDdecoding length (L)=384, forward ratio (FR)=0.52026.05 | 68.43 | |
| PruMerge+Backbone=LLaVA-1.5-7B, Retained Tokens=1282026.06 | 68.2 | |
| CDPrunerBackbone=LLaVA-1.5-7B, Retained Tokens=642026.06 | 68.1 | |
| SPAREBackbone=LLaVA-1.5-7B, Retained Tokens=642026.06 | 68.1 | |
| ID-Selection§Token Budget=160 (94.4%)2026.04 | 68 | |
| PruMerge+Backbone=LLaVA-1.5-7B, Retained Tokens=642026.06 | 68 | |
| IGdecoding length (L)=384, forward ratio (FR)=0.52026.05 | 66.57 | |
| Confidencedecoding length (L)=384, forward ratio (FR)=0.52026.05 | 66.06 | |
| Margindecoding length (L)=384, forward ratio (FR)=0.52026.05 | 65.54 | |
| Margindecoding length (L)=384, forward ratio (FR)=0.252026.05 | 64.98 | |
| VRCDdecoding length (L)=384, forward ratio (FR)=0.252026.05 | 64.49 | |
| Entropydecoding length (L)=384, forward ratio (FR)=0.52026.05 | 63.56 | |
| VRCDDecoding Length=192, Forward Ratio (FR)=0.252026.05 | 63.38 | |
| Margindecoding length (L)=384, forward ratio (FR)=0.1252026.05 | 63.38 | |
| Confidencedecoding length (L)=384, forward ratio (FR)=0.252026.05 | 62.96 | |
| IGdecoding length (L)=384, forward ratio (FR)=0.252026.05 | 62.3 | |
| VRCDdecoding length (L)=384, forward ratio (FR)=0.1252026.05 | 56.62 | |
| ConfidenceDecoding Length=192, Forward Ratio (FR)=0.1252026.05 | 55.68 | |
| Entropydecoding length (L)=384, forward ratio (FR)=0.252026.05 | 54.47 | |
| IGdecoding length (L)=384, forward ratio (FR)=0.1252026.05 | 54.12 | |
| Confidencedecoding length (L)=384, forward ratio (FR)=0.1252026.05 | 53.93 | |
| MMaDAL=192, FR=0.52026.05 | 49.22 | |
| MMaDA+VRCDL=384, FR=0.1252026.05 | 48.29 | |
| MMaDA+VRCDL=192, FR=0.1252026.05 | 47.51 | |
| MMaDA+VRCDL=384, FR=0.52026.05 | 47.43 | |
| MMaDAL=384, FR=0.1252026.05 | 47.08 | |
| MMaDA+VRCDL=192, FR=0.52026.05 | 46.79 | |
| MMaDA+VRCDL=384, FR=0.252026.05 | 46.47 | |
| MMaDAL=384, FR=0.52026.05 | 45.51 | |
| MMaDA+VRCDL=192, FR=0.252026.05 | 45.1 | |
| MMaDAL=192, FR=0.1252026.05 | 44.75 | |
| MMaDAL=192, FR=0.252026.05 | 43.86 | |
| MMaDAL=384, FR=0.252026.05 | 43.1 | |
| Entropydecoding length (L)=384, forward ratio (FR)=0.1252026.05 | 42.44 |