List your company
ResearchFormerly WizWand

Multi-discipline Long Video Understanding on MLVU

68.9ScoreQwen3-VL-8B-Instruct
42.48449.34256.263.058Oct 27, 2025Nov 28, 2025Dec 31, 2025Feb 2, 2026Mar 6, 2026Apr 8, 2026May 11, 2026
Updated 5mo ago

Evaluation results

MethodLinks
Qwen3-VL-8B-InstructMax Input Frames=64, Retention Ratio=100.0%2026.0368.9
Qwen3-VL-8B-InstructMax Input Frames=642026.0368.9
HoPEBackbone=Qwen3-VL-8B-Instruct2025.1068.81
VideoRoPEBackbone=Qwen3-VL-8B-Instruct2025.1068.31
MRoPE-IBackbone=Qwen3-VL-8B-Instruct2025.1067.52
MRoPEBackbone=Qwen3-VL-8B-Instruct2025.1067.34
MHRoPEBackbone=Qwen3-VL-8B-Instruct2025.1067.2
V-CASTMax Input Frames=64, Retention Ratio=35%2026.0367.1
CircleRoPEBackbone=Qwen3-VL-8B-Instruct2025.1066.61
FastVIDMax Input Frames=64, Retention Ratio=35%2026.0366.1
Vanilla RoPEBackbone=Qwen3-VL-8B-Instruct2025.1066.01
VidCom2Max Input Frames=64, Retention Ratio=35%2026.0365.5
V-CASTMax Input Frames=64, Retention Ratio=25%2026.0365.4
FastVIDMax Input Frames=64, Retention Ratio=25%2026.0365
V-CASTMax Input Frames=64, Retention Ratio=15%2026.0364.4
HoliTomMax Input Frames=64, Retention Ratio=35%2026.0364.4
VisionZipMax Input Frames=64, Retention Ratio=35%2026.0364.3
VidCom2Max Input Frames=64, Retention Ratio=25%2026.0364
Qwen3-VL-8B-InstructMax Input Frames=32, Retention Ratio=100.0%2026.0363.5
V-CASTMax Input Frames=32, Retention Ratio=25%2026.0363.5
Qwen3-VL-8B-InstructMax Input Frames=322026.0363.5
VisionZipMax Input Frames=64, Retention Ratio=25%2026.0363.4
HoliTomMax Input Frames=64, Retention Ratio=25%2026.0363.4
FastVIDMax Input Frames=64, Retention Ratio=15%2026.0363.3
V-CASTMax Input Frames=32, Retention Ratio=35%2026.0363.1
FlashVIDMax Input Frames=32, Retention Ratio=35%2026.0362.4
FastVIDMax Input Frames=32, Retention Ratio=35%2026.0362.3
VidCom2Max Input Frames=32, Retention Ratio=35%2026.0362
HoliTomMax Input Frames=32, Retention Ratio=35%2026.0362
HoliTomMax Input Frames=64, Retention Ratio=15%2026.0361.8
FlashVIDMax Input Frames=32, Retention Ratio=25%2026.0361.7
VisionZipMax Input Frames=64, Retention Ratio=15%2026.0361.4
VisionZipMax Input Frames=32, Retention Ratio=35%2026.0361.3
HoliTomMax Input Frames=32, Retention Ratio=25%2026.0361.2
VisionZipMax Input Frames=32, Retention Ratio=25%2026.0360.8
FastVIDMax Input Frames=32, Retention Ratio=25%2026.0360.7
VidCom2Max Input Frames=32, Retention Ratio=25%2026.0360.6
VidCom2Max Input Frames=64, Retention Ratio=15%2026.0360.3
V-CASTMax Input Frames=32, Retention Ratio=15%2026.0360.1
VisionZipMax Input Frames=32, Retention Ratio=15%2026.0360
FlashVIDMax Input Frames=32, Retention Ratio=15%2026.0360
HoliTomMax Input Frames=32, Retention Ratio=15%2026.0359.6
FastVIDMax Input Frames=32, Retention Ratio=15%2026.0359
VidCom2Max Input Frames=32, Retention Ratio=15%2026.0357.7
EchoPruneBackbone=Qwen2.5VL-3B, Frames / Pruning Configuration=F160 (5×) / 20%2026.0551.8
EchoPruneNumber of Frames (F)=320, Compression Ratio=20x, Token Budget=5% (ultra)2026.0551.4
EchoPruneBackbone=Qwen2.5VL-7B, Frames / Pruning Configuration=F160 (5×) / 20%2026.0551.4
RandomBackbone=Qwen2.5VL-7B, Frames / Pruning Configuration=F160 (5×) / 20%2026.0549.7
MI-PrunerBackbone=Qwen2.5VL-7B, Frames / Pruning Configuration=F160 (5×) / 20%2026.0549.6
FlashVIDBackbone=Qwen2.5VL-3B, Frames / Pruning Configuration=F160 (5×) / 20%2026.0548.9
FlashVIDBackbone=Qwen2.5VL-7B, Frames / Pruning Configuration=F160 (5×) / 20%2026.0548.6
MI-PrunerBackbone=Qwen2.5VL-3B, Frames / Pruning Configuration=F160 (5×) / 20%2026.0548.4
RandomBackbone=Qwen2.5VL-3B, Frames / Pruning Configuration=F160 (5×) / 20%2026.0548
Qwen2.5VL-7BBackbone=Qwen2.5VL-7B, Frames / Pruning Configuration=F32 (1×) / 100%2026.0544.6
Qwen2.5VL-3BBackbone=Qwen2.5VL-3B, Frames / Pruning Configuration=F32 (1×) / 100%2026.0543.5