List your company

Video Understanding on 5 Video Benchmarks suite (MMVU, MVBench, VMME, VMMES, VMMMU)

74.2MMVU ScoreZPPO
33.84844.32454.865.276Jun 16, 2026

Evaluation results

MethodLinks
ZPPOBackbone=Qwen3.5-9B2026.0674.271.770.272.882.174.2
Qwen3.5-9BBackbone=Qwen3.5-9B2026.0673.871.470.572.480.973.8
GRPOBackbone=Qwen3.5-9B, Prompt Replay Buffer Augmentation=true2026.0673.171.569.371.28273.4
On-DistillBackbone=Qwen3.5-9B, Prompt Replay Buffer Augmentation=true2026.067370.469.671.680.773.1
Off-DistillBackbone=Qwen3.5-9B, Prompt Replay Buffer Augmentation=true2026.0672.569.668.571.580.872.6
ZPPOBackbone=Qwen3.5-4B2026.0670.669.668.570.280.271.8
Qwen3.5-4BBackbone=Qwen3.5-4B2026.0669.270.268.169.980.371.5
GRPOBackbone=Qwen3.5-4B, Prompt Replay Buffer Augmentation=true2026.0669.269.168.268.779.470.9
On-DistillBackbone=Qwen3.5-4B, Prompt Replay Buffer Augmentation=true2026.0668.767.86769.378.970.3
Off-DistillBackbone=Qwen3.5-4B, Prompt Replay Buffer Augmentation=true2026.0667.967.466.468.878.269.7
Qwen3.5 + ZPPOModel Size=2B2026.065962.762.463.968.163.2
Qwen3.5 + GRPOModel Size=2B, Prompt Replay Buffer=true2026.0656.761.761.262.667.361.9
Qwen3.5Model Size=2B2026.0653.760.759.961.267.660.6
Qwen3.5 + On-DistillModel Size=2B, Prompt Replay Buffer=true2026.0652.459.958.659.465.559.2
Qwen3.5 + Off-DistillModel Size=2B, Prompt Replay Buffer=true2026.0651.559.257.95965.458.6
Qwen3.5 + ZPPOModel Size=0.8B2026.0642.556.653.954.956.452.8
Qwen3.5 + GRPOModel Size=0.8B, Prompt Replay Buffer=true2026.0641.353.552.25352.550.5
Qwen3.5Model Size=0.8B2026.063951.650.351.648.848.3
Qwen3.5 + On-DistillModel Size=0.8B, Prompt Replay Buffer=true2026.0636.849.247.849.645.745.8
Qwen3.5 + Off-DistillModel Size=0.8B, Prompt Replay Buffer=true2026.0635.448.946.848.545.445