List your company

Science Question Answering on SQA IMG (Accuracy)

70.7AccuracyID-Selection
41.309648.939856.5764.2002Apr 7, 2026Apr 18, 2026Apr 30, 2026May 12, 2026May 24, 2026Jun 5, 2026Jun 17, 2026

Evaluation results

MethodLinks
ID-SelectionImportance Criterion=LLM second-layer cross-modal attention, Retain Tokens=162026.0470.7
LLaVA-Next-7BToken Budget=2880 (100%)2026.0470.1
SparseVLMBackbone=LLaVA-1.5-7B, Retained Tokens=642026.0669.7
LLaVA-1.5-7BRetain Tokens=5762026.0469.5
CDPrunerRetain Tokens=322026.0469.5
LLaVA-1.5-7BBackbone=LLaVA-1.5-7B, Retained Tokens=5762026.0669.5
DARTRetain Tokens=642026.0469.3
CDPrunerBackbone=LLaVA-1.5-7B, Retained Tokens=322026.0669.3
SPAREBackbone=LLaVA-1.5-7B, Retained Tokens=322026.0669.3
DARTBackbone=LLaVA-1.5-7B, Retained Tokens=1282026.0669.2
DARTBackbone=LLaVA-1.5-7B, Retained Tokens=322026.0669.2
VisPrunerToken Budget=160 (94.4%)2026.0469
CDPrunerBackbone=LLaVA-1.5-7B, Retained Tokens=1282026.0669
DARTBackbone=LLaVA-1.5-7B, Retained Tokens=642026.0668.8
SPAREBackbone=LLaVA-1.5-7B, Retained Tokens=1282026.0668.6
SparseVLMBackbone=LLaVA-1.5-7B, Retained Tokens=1282026.0668.5
VRCDDecoding Length=384, Forward Ratio (FR)=0.52026.0568.43
VRCDdecoding length (L)=384, forward ratio (FR)=0.52026.0568.43
PruMerge+Backbone=LLaVA-1.5-7B, Retained Tokens=1282026.0668.2
CDPrunerBackbone=LLaVA-1.5-7B, Retained Tokens=642026.0668.1
SPAREBackbone=LLaVA-1.5-7B, Retained Tokens=642026.0668.1
ID-Selection§Token Budget=160 (94.4%)2026.0468
PruMerge+Backbone=LLaVA-1.5-7B, Retained Tokens=642026.0668
IGdecoding length (L)=384, forward ratio (FR)=0.52026.0566.57
Confidencedecoding length (L)=384, forward ratio (FR)=0.52026.0566.06
Margindecoding length (L)=384, forward ratio (FR)=0.52026.0565.54
Margindecoding length (L)=384, forward ratio (FR)=0.252026.0564.98
VRCDdecoding length (L)=384, forward ratio (FR)=0.252026.0564.49
Entropydecoding length (L)=384, forward ratio (FR)=0.52026.0563.56
VRCDDecoding Length=192, Forward Ratio (FR)=0.252026.0563.38
Margindecoding length (L)=384, forward ratio (FR)=0.1252026.0563.38
Confidencedecoding length (L)=384, forward ratio (FR)=0.252026.0562.96
IGdecoding length (L)=384, forward ratio (FR)=0.252026.0562.3
VRCDdecoding length (L)=384, forward ratio (FR)=0.1252026.0556.62
ConfidenceDecoding Length=192, Forward Ratio (FR)=0.1252026.0555.68
Entropydecoding length (L)=384, forward ratio (FR)=0.252026.0554.47
IGdecoding length (L)=384, forward ratio (FR)=0.1252026.0554.12
Confidencedecoding length (L)=384, forward ratio (FR)=0.1252026.0553.93
MMaDAL=192, FR=0.52026.0549.22
MMaDA+VRCDL=384, FR=0.1252026.0548.29
MMaDA+VRCDL=192, FR=0.1252026.0547.51
MMaDA+VRCDL=384, FR=0.52026.0547.43
MMaDAL=384, FR=0.1252026.0547.08
MMaDA+VRCDL=192, FR=0.52026.0546.79
MMaDA+VRCDL=384, FR=0.252026.0546.47
MMaDAL=384, FR=0.52026.0545.51
MMaDA+VRCDL=192, FR=0.252026.0545.1
MMaDAL=192, FR=0.1252026.0544.75
MMaDAL=192, FR=0.252026.0543.86
MMaDAL=384, FR=0.252026.0543.1
Entropydecoding length (L)=384, forward ratio (FR)=0.1252026.0542.44