GUI Agent Task Completion on WeChat Mini-App Synthetic (train val)
71.81VLM Eval Success Ratecode-native-reward (Ours)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| code-native-reward (Ours)Training Environment=Synthetic Environment, Hours per step=0.1013, Cost per step=$02026.02 | 71.81 | 48.99 | |
| VLM-reward (Ours)Training Environment=Synthetic Environment, Hours per step=0.1036, Cost per step=$0.52026.02 | 69.8 | 47.65 | |
| Base ModelTraining Environment=N/A2026.02 | 63.76 | 38.93 | |
| VLM-rewardTraining Environment=Real-World Environment, Hours per step=0.2272, Cost per step=$2402026.02 | 63.76 | 44.3 |