List your company
ResearchFormerly WizWand

Real-world Visual Understanding on RealWorldQA (Score Metric)

72.29ScoreOurs-Mixed
-2.2405617.1087236.45855.80728Oct 27, 2025Nov 30, 2025Jan 3, 2026Feb 6, 2026Mar 12, 2026Apr 15, 2026May 20, 2026
Updated 5mo ago

Evaluation results

MethodLinks
Ours-MixedPruning Budget=10%2026.0372.29
Ours-MathPruning Budget=10%2026.0371.63
CKAPruning Budget=10%2026.0371.11
Ours-MathPruning Budget=25%2026.0370.46
Ours-NonMathPruning Budget=10%2026.0369.54
Ours-NonMathPruning Budget=25%2026.0368.76
DARTToken Retention Ratio=70%2026.0568.6
Ours-MixedPruning Budget=25%2026.0368.5
VanillaToken Retention Ratio=100%2026.0568.5
SPprunerToken Retention Ratio=70%2026.0568.4
SPprunerToken Retention Ratio=50%2026.0568
RandomPruning Budget=10%2026.0367.97
FastVToken Retention Ratio=50%2026.0567.4
DARTToken Retention Ratio=50%2026.0567.3
SPprunerToken Retention Ratio=30%2026.0566.8
FastVToken Retention Ratio=70%2026.0566.5
InterlacePruning Budget=10%2026.0364.97
FastVToken Retention Ratio=30%2026.0564.7
DARTToken Retention Ratio=30%2026.0563.9
Ours-MathPruning Budget=40%2026.0363.01
InterlacePruning Budget=25%2026.0362.88
InterlacePruning Budget=40%2026.0362.88
Ours-NonMathPruning Budget=40%2026.0361.7
Ours-MixedPruning Budget=40%2026.0360.92
MRoPE-IBackbone Model=Qwen3-VL-4B-Instruct, Positional Encoding Variant=MRoPE-I2025.1059.78
MHRoPEBackbone Model=Qwen3-VL-4B-Instruct, Positional Encoding Variant=MHRoPE2025.1058.86
RandomPruning Budget=25%2026.0358.43
CKAPruning Budget=25%2026.0357.91
HoPEBackbone Model=Qwen3-VL-4B-Instruct, Positional Encoding Variant=HoPE2025.1057.52
VideoRoPEBackbone Model=Qwen3-VL-4B-Instruct, Positional Encoding Variant=VideoRoPE2025.1057.25
CircleRoPEBackbone Model=Qwen3-VL-4B-Instruct, Positional Encoding Variant=CircleRoPE2025.1057.25
Vanilla RoPEBackbone Model=Qwen3-VL-4B-Instruct, Positional Encoding Variant=Vanilla RoPE2025.1057.12
RandomPruning Budget=40%2026.0356.73
MRoPEBackbone Model=Qwen3-VL-4B-Instruct, Positional Encoding Variant=MRoPE2025.1056.6
CKAPruning Budget=40%2026.0346.27
Pre-trainedTrainable Params=02026.030.664
IADA +LoRATrainable Params=7.7M2026.030.639
LoRA-onlyTrainable Params=7.6M2026.030.634
AttnRes+LoRATrainable Params=7.7M2026.030.626