List your company
ResearchFormerly WizWand

Audio-Visual Question Answering on AVQA

94.2AccuracyV-LynX-7B
42.51255.93169.3582.769Mar 7, 2024Jul 20, 2024Dec 3, 2024Apr 18, 2025Aug 31, 2025Jan 14, 2026May 30, 2026
Updated 5mo ago

Evaluation results

MethodLinks
V-LynX-7BEvaluation Protocol=Task-specific, Model Scale=≥ 7B, ∆Params.=195.0M2026.0594.2
Qwen3-OmniReasoning=true, Model Size=30B, Zero-shot=true2026.0493.8
PAVE-7BEvaluation Protocol=Task-specific, Model Scale=≥ 7B, ∆Params.=256.7M2026.0593.8
V-LynX-0.5BEvaluation Protocol=Task-specific, Model Scale=< 7B, ∆Params.=68.7M2026.0593.1
Crab+2026.0392.16
CAT-7BEvaluation Protocol=fine-tuned2024.0392
CAT-7B-FTEvaluation Protocol=Task-specific, Model Scale=≥ 7B2026.0592
AVRTReasoning=true, Model Size=3B, In-domain=true2026.0491.1
LLaVA-OV-7B-FTEvaluation Protocol=Task-specific, Model Scale=≥ 7B, ∆Params.=161.5M2026.0590.8
AVRTReasoning=true, Model Size=7B, In-domain=true2026.0490.4
PAVE-0.5BEvaluation Protocol=Task-specific, Model Scale=< 7B, ∆Params.=127.6M2026.0590.4
PSTP-NetEvaluation Protocol=fine-tuned2024.0390.2
PSTP-Net2026.0390.2
PSTP-NetEvaluation Protocol=Task-specific, Model Scale=< 7B2026.0590.2
SaSR-Net2026.0389.9
EgoAVU (LoRA)Fine-tuning Strategy=LoRA2026.0289.7
EgoAVU (Full)Fine-tuning Strategy=Full2026.0289.5
Qwen2.5-Omni2026.0289.4
HCRN + HAVFEvaluation Protocol=fine-tuned2024.0389
HCRN+HAVF2026.0389
Qwen2.5 OmniReasoning=false, Model Size=3B, Zero-shot=true2026.0488.3
HGA + HAVFEvaluation Protocol=fine-tuned2024.0387.7
Full TokenBackbone=Qwen2.5-Omni 7B, Comp.↑ (%)=02026.0587.6
Full TokenBackbone=Qwen2.5-Omni 3B, Comp.↑ (%)=02026.0587.6
ContextGuardBackbone=Qwen2.5-Omni 3B, Comp.↑ (%)=612026.0587.3
OmniZipBackbone=Qwen2.5-Omni 3B, Comp.↑ (%)=542026.0587.3
FastVBackbone=Qwen2.5-Omni 3B, Comp.↑ (%)=502026.0587.2
ContextGuardBackbone=Qwen2.5-Omni 7B, Comp.↑ (%)=552026.0587.1
RandomBackbone=Qwen2.5-Omni 3B, Comp.↑ (%)=502026.0587.1
FastVBackbone=Qwen2.5-Omni 7B, Comp.↑ (%)=502026.0587
OmniZipBackbone=Qwen2.5-Omni 7B, Comp.↑ (%)=542026.0587
RandomBackbone=Qwen2.5-Omni 7B, Comp.↑ (%)=502026.0586.7
V-RTSReasoning=true, Model Size=7B, Zero-shot=true2026.0486.6
LLaVA-OV-0.5B-FTEvaluation Protocol=Task-specific, Model Scale=< 7B, ∆Params.=35.2M2026.0586.4
LLaVA-OV-7BEvaluation Protocol=Zero-shot2026.0585.6
Qwen2.5 OmniReasoning=false, Model Size=7B, Zero-shot=true2026.0484.9
v-SALMONN-o1Reasoning=true, Model Size=7B, Zero-shot=true2026.0484.8
Full TokenBackbone=Video-SALMONN2+ 3B, Comp.↑ (%)=02026.0581.6
ContextGuardBackbone=Video-SALMONN2+ 3B, Comp.↑ (%)=562026.0581.4
Kimi-VL-ThinkingReasoning=true, Model Size=Teacher, Zero-shot=true2026.0481.4
OmniZipBackbone=Video-SALMONN2+ 3B, Comp.↑ (%)=492026.0581.3
RandomBackbone=Video-SALMONN2+ 3B, Comp.↑ (%)=502026.0581.2
FastVBackbone=Video-SALMONN2+ 3B, Comp.↑ (%)=502026.0581
TSV MergingEvaluation Protocol=Zero-shot, Composition Strategy=TSV Merging2025.0580.9
OptMergeEvaluation Protocol=Zero-shot, Composition Strategy=OptMerge2025.0580.82
DAMCEvaluation Protocol=Zero-shot, Composition Strategy=DAMC2025.0580.78
NaiveMCEvaluation Protocol=Zero-shot, Composition Strategy=NaiveMC2025.0580.26
VideoEvaluation Protocol=Zero-shot, Input Modality=Video2025.0579.2
Crab2025.0378.94
VALOR2025.0378.9
Task ArithmeticEvaluation Protocol=Zero-shot, Composition Strategy=Task Arithmetic2025.0578.62
Iso-CEvaluation Protocol=Zero-shot, Composition Strategy=Iso-C2025.0577.51
LLaVA-OV-0.5BEvaluation Protocol=Zero-shot2026.0577.4
WUDI MergingEvaluation Protocol=Zero-shot, Composition Strategy=WUDI Merging2025.0576.86
Ties MergingEvaluation Protocol=Zero-shot, Composition Strategy=Ties Merging2025.0575.84
VisionEvaluation Protocol=Zero-shot, Input Modality=Vision2025.0575.55
Dynamic FSQ (Pure Semantic)Approx. Rate=5.0Hz, Backbone=LLaMA 3.1 8B, Acoustic reconstruction probe loss (L_Recon)=false2026.0572.46
TAPFDataset=AudioSet, Quantizer=RVQ8, Frame Rate=50, Token Rate=4002026.0472.08
Dynamic FSQ (w/ Recon)Approx. Rate=5.0Hz, Backbone=LLaMA 3.1 8B, Acoustic reconstruction probe loss (L_Recon)=included2026.0571.39
Dynamic FSQApprox. Rate=6.0Hz, Backbone=LLaMA 3.1 8B2026.0570.15
A-V Static FusionDataset=AudioSet, Quantizer=RVQ8, Frame Rate=50, Token Rate=4002026.0469.52
TAPFDataset=AudioSet, Quantizer=FSQ, Frame Rate=50, Token Rate=502026.0469.41
Weight AverageEvaluation Protocol=Zero-shot, Composition Strategy=Weight Average2025.0569.39
OmniZipBackbone=Video-SALMONN2+ 7B, Comp.↑ (%)=492026.0568.8
Full TokenBackbone=Video-SALMONN2+ 7B, Comp.↑ (%)=02026.0567.8
WavTokenizerDataset=Several, Quantizer=VQ, Frame Rate=75, Token Rate=752026.0467.34
WavTokenizerApprox. Rate=40/75Hz, Backbone=LLaMA 3.1 8B2026.0567.32
ContextGuardBackbone=Video-SALMONN2+ 7B, Comp.↑ (%)=552026.0566.9
RandomBackbone=Video-SALMONN2+ 7B, Comp.↑ (%)=502026.0566.8
FastVBackbone=Video-SALMONN2+ 7B, Comp.↑ (%)=502026.0566.1
DAC TokenizerDataset=Several, Quantizer=RVQ9, Frame Rate=31.25, Token Rate=2812026.0465.61
DAC TokenizerApprox. Rate=75-100Hz, Backbone=LLaMA 3.1 8B2026.0565.61
Dynamic FSQApprox. Rate=4.0Hz, Backbone=LLaMA 3.1 8B2026.0565.26
AF3 (think)Reasoning=true, Model Size=Teacher, Zero-shot=true2026.0464.3
VideoLLaMA2 w/ FastAVFLOPs=56, Latency=0.32, Memory=19G2026.0162.3
VideoLLaMA2FLOPs=100, Latency=0.43, Memory=22G2026.0161.4
video-SALMONN2 w/ FastAVFLOPs=58, Latency=0.29, Memory=21G2026.0158.4
Speech TokenizerDataset=LibriSpeech, Quantizer=RVQ8, Frame Rate=50, Token Rate=4002026.0458.39
Speech TokenizerApprox. Rate=50Hz, Backbone=LLaMA 3.1 8B2026.0558.39
A-V Static FusionDataset=AudioSet, Quantizer=FSQ, Frame Rate=50, Token Rate=502026.0458.32
Fixed-Stride FSQApprox. Rate=~5.5Hz, Backbone=LLaMA 3.1 8B2026.0557.77
video-SALMONN2FLOPs=100, Latency=0.44, Memory=28G2026.0157.6
SYLLABLELMApprox. Rate=~4-5Hz, Backbone=LLaMA 3.1 8B2026.0555.26
AudioEvaluation Protocol=Zero-shot, Input Modality=Audio2025.0547.57
X-InstructBLIPzero-shot=true2025.0344.5