Agreement with process human labels on CUAVerifierBench Browserbase OM2W
78AccuracyUniversal Verifier (UV)
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Universal Verifier (UV)Verifier=UV, Backbone=GPT-5.22026.04 | 78 | 57 | 0.43 | 29 | 20 | |
| Universal VerifierVerifier=UV, Base Model=GPT-5.22026.04 | 78 | 57 | 0.43 | 29 | 20 | |
| WebVoy.Verifier=WebVoy., Backbone=GPT-5.22026.04 | 75 | 56 | 0.4 | 23 | 26 | |
| WebJudgeVerifier=WebJudge, Backbone=GPT-5.22026.04 | 73 | 49 | 0.32 | 36 | 25 | |
| WebJudgeVerifier=WebJudge, Backbone=o4-mini2026.04 | 68 | 53 | 0.34 | 12 | 38 | |
| WebJudgeVerifier=WebJudge, Base Model=o4-mini2026.04 | 68 | 53 | 0.34 | 12 | 38 | |
| WebVoy.Verifier=WebVoy., Backbone=GPT-4o2026.04 | 55 | 47 | 0.22 | 5 | 56 | |
| WebVoyagerVerifier=WebVoy., Base Model=GPT-4o2026.04 | 55 | 47 | 0.22 | 5 | 56 |