List your company
ResearchFormerly WizWand

Video Understanding on LVBench

102.8Overall AccuracySCORE
19.70441.27762.8584.423Nov 17, 2025Dec 24, 2025Jan 30, 2026Mar 9, 2026Apr 15, 2026May 22, 2026Jun 29, 2026
Updated 3mo ago

Evaluation results

MethodLinks
SCORERetention Ratio R=25%, Max Frames=1282026.03102.841.2—
SCORERetention Ratio R=40%, Max Frames=1282026.03102.440.6—
VidCom^2Retention Ratio R=40%, Max Frames=1282026.03102.140.2—
DyCokeRetention Ratio R=40%, Max Frames=1282026.03101.739.8—
VidCom^2Retention Ratio R=25%, Max Frames=1282026.03101.641.3—
FastVIDRetention Ratio R=40%, Max Frames=1282026.03100.339.9—
VanillaRetention Ratio R=100%, Max Frames=1282026.0310038.3—
FastVIDRetention Ratio R=25%, Max Frames=1282026.0399.839.7—
SCORERetention Ratio R=10%, Max Frames=1282026.0399.539.6—
HoliTomRetention Ratio R=40%, Max Frames=1282026.0399.339.1—
HoliTomRetention Ratio R=25%, Max Frames=1282026.0397.138.2—
FastVIDRetention Ratio R=10%, Max Frames=1282026.0395.937.8—
VidCom^2Retention Ratio R=10%, Max Frames=1282026.0395.338.6—
DyCokeRetention Ratio R=25%, Max Frames=1282026.0394.837.3—
HoliTomRetention Ratio R=10%, Max Frames=1282026.0392.336.7—
LongVPO (128f) Stage 2Size=8B, Frames=768, Backbone=InternVL32026.0253.6——
LongVPO (128f) Stage 1Size=8B, Frames=512, Backbone=InternVL32026.0252.4——
AdaQEmbedding Model=LongCLIP, Frames=64, MLLM Backbone=Qwen3-VL-8B2026.0651.4——
LongVPO (256f) Stage 2-iter1Size=8B, Frames=512, Backbone=InternVideo2.52026.0251——
LongVPO (256f) Stage 2-iter2Size=8B, Frames=512, Backbone=InternVideo2.52026.0251——
LongVPO (256f) Stage 1Size=8B, Frames=512, Backbone=InternVideo2.52026.0250.9——
AdaQEmbedding Model=CLIP, Frames=64, MLLM Backbone=Qwen3-VL-8B2026.0650.6——
OneClip-RAGEmbedding Model=CLIP, Frames=64, MLLM Backbone=Qwen3-VL-8B2026.0650.5——
LongVPO (512f) Stage 1Size=8B, Frames=512, Backbone=InternVideo2.52026.0250.4——
Q-FrameEmbedding Model=LongCLIP, Frames=64, MLLM Backbone=Qwen3-VL-8B2026.0650.2——
LongVPO (128f) Stage 2Size=8B, Frames=512, Backbone=InternVL2.52026.0250.1——
LongVPO (256f) Stage 1Size=8B, Frames=512, Backbone=InternVL2.52026.0249.6——
LongVPO (128f) Stage 1Size=8B, Frames=512, Backbone=InternVL2.52026.0249.4——
Top-kEmbedding Model=CLIP, Frames=64, MLLM Backbone=Qwen3-VL-8B2026.0649.2——
AdaQEmbedding Model=LongCLIP, Frames=64, MLLM Backbone=LLaVA-Video-7B2026.0649.1——
AKSEmbedding Model=BLIP, Frames=64, MLLM Backbone=Qwen3-VL-8B2026.0649——
GPT-4oFrame=602026.0348.9——
InternVL3*Size=8B, Frames=5122026.0248.8——
AdaQEmbedding Model=CLIP, Frames=64, MLLM Backbone=LLaVA-Video-7B2026.0648.5——
VideoChat-FlashSize=7B, Frames=5122026.0248.2——
OneClip-RAGEmbedding Model=CLIP, Frames=64, MLLM Backbone=LLaVA-Video-7B2026.0648.2——
AKSEmbedding Model=BLIP, Frames=64, MLLM Backbone=LLaVA-Video-7B2026.0647.6——
InternVideo2.5*Size=8B, Frames=5122026.0247.4——
Q-FrameEmbedding Model=LongCLIP, Frames=64, MLLM Backbone=LLaVA-Video-7B2026.0647.1——
AdaQEmbedding Model=LongCLIP, Frames=64, MLLM Backbone=Qwen2.5-VL-7B2026.0647.1——
AdaQEmbedding Model=CLIP, Frames=64, MLLM Backbone=Qwen2.5-VL-7B2026.0647——
Top-kEmbedding Model=CLIP, Frames=64, MLLM Backbone=LLaVA-Video-7B2026.0646.9——
Top-kEmbedding Model=CLIP, Frames=64, MLLM Backbone=Qwen2.5-VL-7B2026.0646.7——
OneClip-RAGEmbedding Model=CLIP, Frames=64, MLLM Backbone=Qwen2.5-VL-7B2026.0646.5——
Q-FrameEmbedding Model=LongCLIP, Frames=64, MLLM Backbone=Qwen2.5-VL-7B2026.0646.5——
Qwen2.5-Omni-3BModel Scale=3B2026.0646.48——
AKSEmbedding Model=BLIP, Frames=64, MLLM Backbone=Qwen2.5-VL-7B2026.0646.4——
Qwen2.5-VL#S (Model Size)=7B, #F (Number of input frames)=64, Architecture=AR Baseline, reproduced=true2026.0145.3——
InternVL2.5*Size=8B, Frames=5122026.0245.2——
AdaQEmbedding Model=LongCLIP, Frames=32, MLLM Backbone=LLaVA-OneVision-7B2026.0644.8——
VidLaDA#S (Model Size)=8B, #F (Number of input frames)=64, Architecture=DLM Baseline2026.0144.7——
Q-FrameEmbedding Model=LongCLIP, Frames=32, MLLM Backbone=LLaVA-OneVision-7B2026.0644.3——
AdaQEmbedding Model=CLIP, Frames=32, MLLM Backbone=LLaVA-OneVision-7B2026.0644.3——
OneClip-RAGEmbedding Model=CLIP, Frames=32, MLLM Backbone=LLaVA-OneVision-7B2026.0644.1——
Top-kEmbedding Model=CLIP, Frames=32, MLLM Backbone=LLaVA-OneVision-7B2026.0643.8——
Qwen3-VL-8BFrames=64, MLLM Backbone=Qwen3-VL-8B2026.0643.6——
AKSEmbedding Model=BLIP, Frames=32, MLLM Backbone=LLaVA-OneVision-7B2026.0643.5——
EVA-GRPOFrame=26.8*2026.0343.3——
InternVL2.5Size=8B, Frames=642026.0243.2——
LLaVA-Video-7BFrames=64, MLLM Backbone=LLaVA-Video-7B2026.0641.9——
LLaVA-Video#S (Model Size)=7B, #F (Number of input frames)=64, Architecture=AR Baseline, reproduced=true2026.0141.5——
LLaVA-OneVision#S (Model Size)=7B, #F (Number of input frames)=32, Architecture=AR Baseline, reproduced=true2026.0140.7——
ParaVT-8BSetting=tool-augmented setting (<think>→<tool_call>→<answer>)2026.0539.8——
Qwen2.5-VL-7BFrames=64, MLLM Backbone=Qwen2.5-VL-7B2026.0639.3——
LLaVA-OneVision-7BFrames=32, MLLM Backbone=LLaVA-OneVision-7B2026.0638.6——
ReWatch-R1-7BSetting=reasoning-enhanced setting (<think>→<answer>)2026.0538.5——
InternVL2.5#S (Model Size)=7B, #F (Number of input frames)=64, Architecture=AR Baseline2026.0138.4——
AVLM-2BModel Scale=2B2026.0638.22——
Time-R1-7BSetting=reasoning-enhanced setting (<think>→<answer>)2026.0538.2——
Conan-7BSetting=tool-augmented setting (<think>→<tool_call>→<answer>)2026.0538.2——
VideoRFT-7BSetting=reasoning-enhanced setting (<think>→<answer>)2026.0538——
LongVT-RFT-7BSetting=tool-augmented setting (<think>→<tool_call>→<answer>)2026.0537.9——
Video-R1-7BSetting=reasoning-enhanced setting (<think>→<answer>)2026.0536.9——
FrameThinkerFrame=23.92026.0336.6——
LLaDA-V#S (Model Size)=8B, #F (Number of input frames)=32, Architecture=DLM Baseline, reproduced=true2026.0136.4——
VideoLLaMA2.1#S (Model Size)=7B, #F (Number of input frames)=32, Architecture=AR Baseline2026.0136.2——
EVA-KTOFrame=34.5*2026.0336——
Video-R1Frame=322026.0335.3——
GPT-4oSetting=best-setting numbers from official reports2026.0534.7——
VideoChat-R1Frame=322026.0334.3——
Gemini-1.5-ProSize=Undisclosed2026.0233.1——
Gemini-1.5-ProFrame=36002026.0333.1——
Gemini 1.5 ProSetting=best-setting numbers from official reports2026.0533.1——
Qwen3-VL-8BSetting=tool-augmented setting (<think>→<tool_call>→<answer>)2026.0533.1——
Qwen2.5-VL-7BSetting=direct-answer setting (native single-pass prompt)2026.0532.2——
DeepSportFrames=13.22025.1132——
SAGE-7BSetting=tool-augmented setting (<think>→<tool_call>→<answer>)2026.0531.8——
Qwen2.5-VL-7B-InstructFrames=322025.1131.6——
Qwen2.5-VLFrame=322026.0331.6——
GPT4-oSize=Undisclosed2026.0230.8——
VideoAgentFrame=25.52026.0329.3——
EVA-SFTFrame=56.2*2026.0326.5——
Gemma-4-E2B-itModel Scale=4B2026.0624.92——
VideoChat-R1-7BSetting=reasoning-enhanced setting (<think>→<answer>)2026.0523.8——
VideoZoomer-7BSetting=tool-augmented setting (<think>→<tool_call>→<answer>)2026.0522.9——
AV-Reasoner2025.08—33.7—
Chain-of-Time Reasoning (Ours)2026.04—59.9—
CircleRoPEBackbone Model=Qwen3-VL-4B-Instruct, Positional Encoding Variant=CircleRoPE2025.10—35.54—
DELTAVID-Qwen3-VL-8BParameters=8B, Base Model=Qwen3-VL2026.06—46.2—
Echolnk2025.08—37.6—