List your company
ResearchFormerly WizWand

Video Question Answering on LongVideoBench (test)

63.7Accuracy (Long)FOCUS (LLaVA-V-7B)
31.4639.8348.256.57Aug 5, 2025Sep 16, 2025Oct 29, 2025Dec 11, 2025Jan 23, 2026Mar 7, 2026Apr 19, 2026
Updated 6mo ago

Evaluation results

MethodLinks
FOCUS (LLaVA-V-7B)Frame=642026.0463.7—5972.3
GPT-4o (0513)Frame=2562026.0461.6—66.776.8
Qwen3-VL + Q-GateBackbone=Qwen3-VL, Strategy=Q-Gate, Frame=322026.0459.4—63.1170.59
LLaVA-OneVision-72BFrame=322026.0459.3—63.977.4
LLaVA-Video-72B-Qwen2Frame=1282026.0459.3—63.977.4
Qwen3-VL + Q-GateBackbone=Qwen3-VL, Strategy=Q-Gate, Frame=82026.0458.69—63.1168.24
Qwen3-VL + AKS*Backbone=Qwen3-VL, Strategy=AKS*, Frame=322026.0457.8—65.2967.06
Qwen3-VL + AKS*Backbone=Qwen3-VL, Strategy=AKS*, Frame=82026.0454.61—60.9270.59
Qwen3-VL + VSLSBackbone=Qwen3-VL, Strategy=VSLS, Frame=322026.0452.66—58.7465.88
Qwen3-VL + T*Backbone=Qwen3-VL, Strategy=T*, Frame=322026.0451.95—56.5569.41
Qwen3-VL + VSLSBackbone=Qwen3-VL, Strategy=VSLS, Frame=82026.0451.6—55.8369.41
Qwen3-VL (Uniform)Backbone=Qwen3-VL, Strategy=Uniform, Frame=322026.0451.24—56.5571.76
Qwen3-VL + T*Backbone=Qwen3-VL, Strategy=T*, Frame=82026.0450.89—54.1367.06
GPT-4o + Q-GateBackbone=GPT-4o, Strategy=Q-Gate, Frame=82026.0450.71—56.5565.41
Uniform SamplingBackbone Model=GPT-4o, Initial Keyframe Setting=Top 32, Avg. Frames=32.02025.0850.6—53.574
GPT-4o + Q-GateBackbone=GPT-4o, Strategy=Q-Gate, Frame=322026.0450.35—53.6463.75
GPT-4o + AKS*Backbone=GPT-4o, Strategy=AKS*, Frame=82026.0449.65—53.8859.49
AFP+GraphBackbone Model=GPT-4o, Initial Keyframe Setting=Top 32, Avg. Frames=4.12025.0849.1—53.584
GPT-4o + AKS*Backbone=GPT-4o, Strategy=AKS*, Frame=322026.0447.7—50.9761.88
AFP+GraphBackbone Model=GPT-4o, Initial Keyframe Setting=Top 8, Avg. Frames=2.12025.0847.3—53.178
Uniform SamplingBackbone Model=GPT-4o, Initial Keyframe Setting=Top 8, Avg. Frames=8.02025.0847.1—49.467.3
AKS*Backbone Model=GPT-4o, Initial Keyframe Setting=Top 8, Avg. Frames=8.02025.0847—46.266
AKS*Backbone Model=GPT-4o, Initial Keyframe Setting=Top 32, Avg. Frames=32.02025.0847—49.258
Qwen3-VL (Uniform)Backbone=Qwen3-VL, Strategy=Uniform, Frame=82026.0446.63—51.4671.76
GPT-4o + VSLSBackbone=GPT-4o, Strategy=VSLS, Frame=82026.0445.21—49.7656.25
AFP+GraphBackbone Model=LLaVA-Video-7B, Initial Keyframe Setting=Top 32, Avg. Frames=4.12025.0845.2—53.864
AFP+GraphBackbone Model=LLaVA-Video-7B, Initial Keyframe Setting=Top 8, Avg. Frames=2.12025.0844.6—53.162
GPT-4o + VSLSBackbone=GPT-4o, Strategy=VSLS, Frame=322026.0444.5—49.5168.24
GPT-4o (Uniform)Backbone=GPT-4o, Strategy=Uniform, Frame=322026.0443.79—46.3660
GPT-4o + T*Backbone=GPT-4o, Strategy=T*, Frame=322026.0443.62—48.7963.53
AFP+GraphBackbone Model=Qwen2.5-VL-7B, Initial Keyframe Setting=Top 32, Avg. Frames=4.12025.0843.5—46.964
AKS*Backbone Model=LLaVA-Video-7B, Initial Keyframe Setting=Top 32, Avg. Frames=32.02025.0843.5—46.550
AFP+GraphBackbone Model=Qwen2.5-VL-7B, Initial Keyframe Setting=Top 8, Avg. Frames=2.12025.0842.9—46.564
GPT-4o + T*Backbone=GPT-4o, Strategy=T*, Frame=82026.0442.73—47.8264.71
Uniform SamplingBackbone Model=LLaVA-Video-7B, Initial Keyframe Setting=Top 32, Avg. Frames=32.02025.0841.7—46.240
AKS*Backbone Model=LLaVA-Video-7B, Initial Keyframe Setting=Top 8, Avg. Frames=8.02025.0841.4—45.852
GPT-4o (Uniform)Backbone=GPT-4o, Strategy=Uniform, Frame=82026.0440.43—43.6957.65
Uniform SamplingBackbone Model=Qwen2.5-VL-7B, Initial Keyframe Setting=Top 8, Avg. Frames=8.02025.0840.3—38.442.8
Uniform SamplingBackbone Model=LLaVA-Video-7B, Initial Keyframe Setting=Top 8, Avg. Frames=8.02025.0840.2—46.950
AKS*Backbone Model=Qwen2.5-VL-7B, Initial Keyframe Setting=Top 32, Avg. Frames=32.02025.0835.7—42.760
AKS*Backbone Model=Qwen2.5-VL-7B, Initial Keyframe Setting=Top 8, Avg. Frames=8.02025.0835.4—41.556
Uniform SamplingBackbone Model=Qwen2.5-VL-7B, Initial Keyframe Setting=Top 32, Avg. Frames=32.02025.0832.7—36.550
FastVSize=7B, Tokens (%)=25.02026.03—56.8——
FastVSize=7B, Tokens (%)=15.02026.03—51.5——
FastVIDSize=7B, Tokens (%)=25.02026.03—56.3——
FastVIDSize=7B, Tokens (%)=15.02026.03—56.2——
FrameFusionSize=7B, Tokens (%)=25.02026.03—54.8——
FrameFusionSize=7B, Tokens (%)=15.02026.03—53——
LLaVA-OneVision 16 framesLLM Size=7B, Frames=162025.03—55.7——
LLaVA-OneVision 16 frames + BOLTLLM Size=7B, Frames=162025.03—57——
LLaVA-OneVision 32 framesLLM Size=7B, Frames=322025.03—56.4——
LLaVA-OneVision 32 frames + BOLTLLM Size=7B, Frames=322025.03—59.6——
LLaVA-OneVision 8 framesLLM Size=7B, Frames=82025.03—54.2——
LLaVA-OneVision 8 frames + BOLTLLM Size=7B, Frames=82025.03—55.6——
LLaVA-OneVision-7BSize=7B, Tokens (%)=1002026.03—56.6——
PruneVidSize=7B, Tokens (%)=25.02026.03—55.1——
PruneVidSize=7B, Tokens (%)=15.02026.03—55.6——
SemVIDSize=7B, Tokens (%)=25.02026.03—56.4——
SemVIDSize=7B, Tokens (%)=15.02026.03—56.4——
ShareGPT4VideoLLM Size=8B2025.03—41.8——
VanillaSize=7B, Tokens (%)=1002026.03—56.6——
Video-LLaVALLM Size=7B2025.03—39.1——
VideoChat2LLM Size=7B2025.03—39.3——
VisionZipSize=7B, Tokens (%)=25.02026.03—56——
VisionZipSize=7B, Tokens (%)=15.02026.03—54.2——