Continual Android Device Control on AndroidControl-CL Task Order 1
81.73Shopping AccuracySFT-Joint-Training
Evaluation Results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| SFT-Joint-TrainingBackbone=QwenVL2.5-3b, Task Order=12026.03 | 81.73 | 86.74 | 85.75 | 82.24 | 88.17 | 77 | 82.76 | 83.48 | 41.66 | — | |
| OursBackbone=QwenVL2.5-3b, Task Order=12026.03 | 80.32 | 85.46 | 81.99 | 82.03 | 88.17 | 75.81 | 82.53 | 82.33 | 38.03 | -0.02 | |
| GRPOBackbone=QwenVL2.5-3b, Task Order=12026.03 | 79.63 | 84.74 | 83.43 | 81.34 | 85.74 | 74.08 | 81.73 | 81.53 | 36.78 | -0.62 | |
| RIF-RFTBackbone=QwenVL2.5-3b, Task Order=12026.03 | 79.53 | 83.35 | 80.77 | 80.25 | 85.94 | 73 | 80.25 | 80.44 | 32.91 | -0.58 | |
| SFT+KLBackbone=QwenVL2.5-3b, Task Order=12026.03 | 79.3 | 83.85 | 82.32 | 79.84 | 85.83 | 74.08 | 80.71 | 80.84 | 34.69 | -1.01 | |
| SFT+ReplayBackbone=QwenVL2.5-3b, Task Order=12026.03 | 74.61 | 83.35 | 81.66 | 79.02 | 85.61 | 72.26 | 82.08 | 79.8 | 30.19 | -3.11 | |
| SFTBackbone=QwenVL2.5-3b, Task Order=12026.03 | 70.54 | 79.52 | 78.9 | 76.41 | 83.24 | 69.24 | 80.42 | 76.9 | 23.53 | -5.73 | |
| Zero-ShotBackbone=QwenVL2.5-3b, Task Order=12026.03 | 48.02 | 57.71 | 49.5 | 52.81 | 55.87 | 42.33 | 48.86 | 50.73 | 1.86 | — |