Android UI Control on AndroidControl-CL Task Order 1 v1 (test)
77.83Shopping AccuracySFT
Evaluation Results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| SFTBackbone=LLava-onsevision-0.5b, Training Strategy=Joint-Training2026.03 | 77.83 | 84.46 | 83.24 | 78.41 | 84.76 | 69.73 | 79.58 | 79.69 | 28.63 | — | |
| RIF-RFTBackbone=LLava-onsevision-0.5b2026.03 | 76.53 | 74.19 | 75.99 | 73.55 | 81.4 | 65.91 | 75.34 | 74.4 | 22.17 | -0.88 | |
| CGLBackbone=LLava-onsevision-0.5b2026.03 | 75.55 | 81.32 | 80.73 | 76.88 | 83.41 | 68.9 | 78.13 | 77.84 | 24.77 | -0.52 | |
| GRPOBackbone=LLava-onsevision-0.5b2026.03 | 75.44 | 78.86 | 78.01 | 76.41 | 81.57 | 67.93 | 77.8 | 76.57 | 23.85 | -0.93 | |
| SFT+KLBackbone=LLava-onsevision-0.5b2026.03 | 74.16 | 75.23 | 76.89 | 75.49 | 82.06 | 66.01 | 79.29 | 75.59 | 22.39 | -1.78 | |
| SFT+ReplayBackbone=LLava-onsevision-0.5b2026.03 | 70.88 | 76.43 | 77.23 | 74.19 | 78.21 | 64.37 | 80.63 | 74.56 | 20.07 | -4.69 | |
| SFTBackbone=LLava-onsevision-0.5b2026.03 | 65.64 | 74.73 | 74.66 | 72 | 77.39 | 64.12 | 80.08 | 72.66 | 14.61 | -6.81 |