Reward Modeling on OSWorld-Verified (Class-Imbalanced, Human Evaluation) 1.0 (test)
88.5PrecisionVAGEN
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| VAGENVerified Agent Backbone=UI-TARS-1.5-7B2026.01 | 88.5 | 90.1 | 89.3 | 93.4 | |
| WebRLVerified Agent Backbone=UI-TARS-1.5-7B2026.01 | 76.6 | 56.2 | 64.8 | 82.2 | |
| ZeroGUIVerified Agent Backbone=UI-TARS-1.5-7B2026.01 | 76.6 | 79.4 | 78 | 86.7 | |
| FullTrajEvalVerified Agent Backbone=UI-TARS-1.5-7B2026.01 | 74.8 | 76.9 | 75.8 | 85.3 | |
| DistRLVerified Agent Backbone=UI-TARS-1.5-7B2026.01 | 73.7 | 53.8 | 62.2 | 81.2 | |
| AndroidGenVerified Agent Backbone=UI-TARS-1.5-7B2026.01 | 73.6 | 60.4 | 66.3 | 82 | |
| DigiRLVerified Agent Backbone=UI-TARS-1.5-7B2026.01 | 68.4 | 52.9 | 59.7 | 79.8 |