Video Reasoning on SEED-Bench L3 OOD R1
49.3AccuracyAPPO
Evaluation Results
| Method | Links | |
|---|---|---|
| APPOBackbone=Qwen2.5-VL-7B, Resolution=224x224, Sampling Rate=1 fps, Max Frames=602026.02 | 49.3 | |
| DAPOBackbone=Qwen2.5-VL-7B, Resolution=224x224, Sampling Rate=1 fps, Max Frames=602026.02 | 48.7 | |
| GRPOBackbone=Qwen2.5-VL-7B, Resolution=224x224, Sampling Rate=1 fps, Max Frames=602026.02 | 48.2 | |
| SFTBackbone=Qwen2.5-VL-7B, Resolution=224x224, Sampling Rate=1 fps, Max Frames=602026.02 | 43.9 | |
| APPOBackbone=Qwen2.5-VL-3B, Resolution=224x224, Sampling Rate=1 fps, Max Frames=602026.02 | 35 | |
| APPOSize=7B, Training Data=34K, Resolution=224 x 224, Sampling Rate (fps)=1fps, Maximum Frames=30, Zero-shot evaluation=true2026.02 | 34.7 | |
| DAPOBackbone=Qwen2.5-VL-3B, Resolution=224x224, Sampling Rate=1 fps, Max Frames=602026.02 | 31.8 | |
| GRPOBackbone=Qwen2.5-VL-3B, Resolution=224x224, Sampling Rate=1 fps, Max Frames=602026.02 | 31 | |
| Base ModelBackbone=Qwen2.5-VL-7B, Resolution=224x224, Sampling Rate=1 fps, Max Frames=602026.02 | 29.6 | |
| Video-R1Size=7B, Training Data=260K, Resolution=224 x 224, Sampling Rate (fps)=1fps, Maximum Frames=30, Zero-shot evaluation=true2026.02 | 28.8 | |
| SFTBackbone=Qwen2.5-VL-3B, Resolution=224x224, Sampling Rate=1 fps, Max Frames=602026.02 | 28.6 | |
| VideoChat-R1Size=7B, Training Data=18K, Resolution=224 x 224, Sampling Rate (fps)=1fps, Maximum Frames=30, Zero-shot evaluation=true2026.02 | 27.8 | |
| VideoRFTSize=7B, Training Data=310K, Resolution=224 x 224, Sampling Rate (fps)=1fps, Maximum Frames=30, Zero-shot evaluation=true2026.02 | 27.7 | |
| Base ModelBackbone=Qwen2.5-VL-3B, Resolution=224x224, Sampling Rate=1 fps, Max Frames=602026.02 | 27 | |
| GRPO-CARESize=7B, Training Data=260K, Resolution=224 x 224, Sampling Rate (fps)=1fps, Maximum Frames=30, Zero-shot evaluation=true2026.02 | 26 | |
| TW-GRPOSize=7B, Training Data=1K, Resolution=224 x 224, Sampling Rate (fps)=1fps, Maximum Frames=30, Zero-shot evaluation=true2026.02 | 25.6 |