List your company

Video Understanding on LongVideoBench, MLVU, and VideoMME Aggregate

65.7Average ScoreInternVL3-8B + LiteFrame
37.51644.83352.1559.467Dec 7, 2025Jan 2, 2026Jan 29, 2026Feb 25, 2026Mar 24, 2026Apr 20, 2026May 17, 2026

Evaluation results

MethodLinks
InternVL3-8B + LiteFrameFrames=256 (8×)2026.0565.7—
InternVL3-8BFrames=322026.0565.3—
InternVL3-8B + LiteFrameFrames=128 (8×)2026.0563.7—
InternVL3-8B + FastVIDFrames=128 (4×)2026.0562.9—
InternVL3-8BFrames=162026.0562.2—
InternVL3-8B + LiteFrameFrames=64 (8×)2026.0561.1—
InternVL3-8B + FastVIDFrames=32 (4×)2026.0559.5—
InternVL3-8B + FastVIDFrames=64 (4×)2026.0559.5—
InternVL3-8BFrames=82026.0559—
LLaVA-OneVision-7BBase Model=LLaVA-OneVision-7B, Equivalent Retention Ratio=100%2026.0256.4100
CaCoVIDBase Model=LLaVA-OneVision-7B, Equivalent Retention Ratio=25%2026.0255.898.9
Qwen2.5-VL-3BBase Model=Qwen2.5-VL-3B, Equivalent Retention Ratio=100%2026.0255.6100
CaCoVIDBase Model=LLaVA-OneVision-7B, Equivalent Retention Ratio=20%2026.0255.398.1
PruneVIDBase Model=LLaVA-OneVision-7B, Equivalent Retention Ratio=25%2026.0255.197.7
DivPruneBase Model=LLaVA-OneVision-7B, Equivalent Retention Ratio=25%2026.025597.5
CaCoVIDBase Model=LLaVA-OneVision-7B, Equivalent Retention Ratio=15%2026.025597.5
VisionZipBase Model=LLaVA-OneVision-7B, Equivalent Retention Ratio=20%2026.0254.997.2
DivPruneBase Model=LLaVA-OneVision-7B, Equivalent Retention Ratio=20%2026.0254.796.9
PruneVIDBase Model=LLaVA-OneVision-7B, Equivalent Retention Ratio=20%2026.0254.796.9
VisionZipBase Model=LLaVA-OneVision-7B, Equivalent Retention Ratio=25%2026.0254.696.8
FrameFusionBase Model=LLaVA-OneVision-7B, Equivalent Retention Ratio=25%2026.0254.696.7
DyCoke + DyToK#Retained Tokens=4704, Token Pruning Ratio=25%2025.1254.4100.6
DivPruneBase Model=LLaVA-OneVision-7B, Equivalent Retention Ratio=15%2026.0254.396.3
DyCoke + DyToK#Retained Tokens=3136, Token Pruning Ratio=50%2025.1254.3100.4
CaCoVIDBase Model=LLaVA-OneVision-7B, Equivalent Retention Ratio=10%2026.0254.296.1
CaCoVIDBase Model=Qwen2.5-VL-3B, Equivalent Retention Ratio=25%2026.0254.297.5
FrameFusionBase Model=LLaVA-OneVision-7B, Equivalent Retention Ratio=20%2026.0254.195.8
Vanilla#Retained Tokens=62722025.1254.1100
PruneVIDBase Model=LLaVA-OneVision-7B, Equivalent Retention Ratio=15%2026.0253.995.5
DivPruneBase Model=Qwen2.5-VL-3B, Equivalent Retention Ratio=25%2026.0253.997
CaCoVIDBase Model=Qwen2.5-VL-3B, Equivalent Retention Ratio=20%2026.0253.996.9
VisionZip† + DyToK#Retained Tokens=3136, Token Pruning Ratio=50%2025.1253.999.6
VisionZip† + DyToK#Retained Tokens=4704, Token Pruning Ratio=25%2025.1253.899.4
DyCoke#Retained Tokens=4704, Token Pruning Ratio=25%2025.1253.899.4
FrameFusionBase Model=Qwen2.5-VL-3B, Equivalent Retention Ratio=25%2026.0253.796.5
VisionZip†#Retained Tokens=4704, Token Pruning Ratio=25%2025.1253.598.9
DyCoke#Retained Tokens=3136, Token Pruning Ratio=50%2025.1253.598.9
VisionZip†#Retained Tokens=3136, Token Pruning Ratio=50%2025.1253.498.7
DivPruneBase Model=LLaVA-OneVision-7B, Equivalent Retention Ratio=10%2026.0253.394.5
VisionZip† + DyToK#Retained Tokens=1568, Token Pruning Ratio=75%2025.1253.398.5
VisionZipBase Model=LLaVA-OneVision-7B, Equivalent Retention Ratio=15%2026.0253.294.3
DivPruneBase Model=Qwen2.5-VL-3B, Equivalent Retention Ratio=20%2026.0253.195.5
VisionZip† + DyToK#Retained Tokens=1120, Token Pruning Ratio=80%2025.125398
PruneVIDBase Model=LLaVA-OneVision-7B, Equivalent Retention Ratio=10%2026.0252.993.8
VisionZipBase Model=Qwen2.5-VL-3B, Equivalent Retention Ratio=25%2026.0252.995.1
VisionZip†#Retained Tokens=1568, Token Pruning Ratio=75%2025.1252.997.8
VisionZip#Retained Tokens=4704, Token Pruning Ratio=25%2025.1252.797.4
PruneVIDBase Model=Qwen2.5-VL-3B, Equivalent Retention Ratio=25%2026.0252.694.6
CaCoVIDBase Model=Qwen2.5-VL-3B, Equivalent Retention Ratio=15%2026.0252.594.4
DyCoke + DyToK#Retained Tokens=1568, Token Pruning Ratio=75%2025.1252.597
VisionZipBase Model=Qwen2.5-VL-3B, Equivalent Retention Ratio=20%2026.0252.494.2
VisionZipBase Model=Qwen2.5-VL-3B, Equivalent Retention Ratio=15%2026.0252.494.3
FastVBase Model=LLaVA-OneVision-7B, Equivalent Retention Ratio=25%2026.0252.392.7
FrameFusionBase Model=Qwen2.5-VL-3B, Equivalent Retention Ratio=20%2026.0252.394.1
PruneVIDBase Model=Qwen2.5-VL-3B, Equivalent Retention Ratio=20%2026.0252.293.8
FrameFusionBase Model=LLaVA-OneVision-7B, Equivalent Retention Ratio=15%2026.0252.192.3
CaCoVIDBase Model=Qwen2.5-VL-3B, Equivalent Retention Ratio=10%2026.025293.4
FastVBase Model=LLaVA-OneVision-7B, Equivalent Retention Ratio=20%2026.0251.891.8
DyCoke#Retained Tokens=1568, Token Pruning Ratio=75%2025.1251.895.7
DivPruneBase Model=Qwen2.5-VL-3B, Equivalent Retention Ratio=15%2026.0251.792.9
FastVBase Model=Qwen2.5-VL-3B, Equivalent Retention Ratio=25%2026.0251.692.7
VisionZip#Retained Tokens=3136, Token Pruning Ratio=50%2025.1251.695.4
VisionZip†#Retained Tokens=1120, Token Pruning Ratio=80%2025.1251.695.4
PruneVIDBase Model=Qwen2.5-VL-3B, Equivalent Retention Ratio=15%2026.0251.592.5
VisionZipBase Model=Qwen2.5-VL-3B, Equivalent Retention Ratio=10%2026.0251.392.3
DyCoke#Retained Tokens=1120, Token Pruning Ratio=80%2025.1251.194.5
DyCoke + DyToK#Retained Tokens=1120, Token Pruning Ratio=80%2025.1251.194.5
FrameFusionBase Model=Qwen2.5-VL-3B, Equivalent Retention Ratio=15%2026.025191.6
DivPruneBase Model=Qwen2.5-VL-3B, Equivalent Retention Ratio=10%2026.0250.590.9
PruneVIDBase Model=Qwen2.5-VL-3B, Equivalent Retention Ratio=10%2026.0250.290.2
VisionZipBase Model=LLaVA-OneVision-7B, Equivalent Retention Ratio=10%2026.0250.188.8
FastVBase Model=Qwen2.5-VL-3B, Equivalent Retention Ratio=20%2026.0250.190.1
FastVBase Model=LLaVA-OneVision-7B, Equivalent Retention Ratio=15%2026.0249.888.3
FrameFusionBase Model=LLaVA-OneVision-7B, Equivalent Retention Ratio=10%2026.0249.287.2
DyCoke + DyToK#Retained Tokens=448, Token Pruning Ratio=90%2025.1249.190.8
VisionZip#Retained Tokens=1568, Token Pruning Ratio=75%2025.1248.990.4
VisionZip† + DyToK#Retained Tokens=448, Token Pruning Ratio=90%2025.1248.890.2
DyCoke#Retained Tokens=448, Token Pruning Ratio=90%2025.1248.489.5
FastVBase Model=Qwen2.5-VL-3B, Equivalent Retention Ratio=15%2026.0248.286.7
FastVBase Model=LLaVA-OneVision-7B, Equivalent Retention Ratio=10%2026.0246.181.6
VisionZip#Retained Tokens=1120, Token Pruning Ratio=80%2025.1245.584.1
VisionZip†#Retained Tokens=448, Token Pruning Ratio=90%2025.1245.183.4
FastVBase Model=Qwen2.5-VL-3B, Equivalent Retention Ratio=10%2026.0244.479.9
VisionZip#Retained Tokens=448, Token Pruning Ratio=90%2025.1238.671.3