Desktop Task Success on WindowsAgentArena
44.2Task Success RateRL + corrected reward estimator (rˆt)
Evaluation Results
| Method | Links | |
|---|---|---|
| RL + corrected reward estimator (rˆt)Model Type=per-OS models2026.06 | 44.2 | |
| RL + corrected reward estimator (rˆt)Model Type=unified model2026.06 | 40.3 | |
| RL + raw evaluator reward (r˜t)Model Type=unified model2026.06 | 38.3 | |
| RL + raw evaluator reward (r˜t)Model Type=per-OS models2026.06 | 35.7 | |
| UI-TARSTraining Setup=Zero-shot baseline2026.06 | 33.1 |