Agreement with outcome human labels on CUAVerifierBench Browserbase OM2W (n=106)
88AccuracyUniversal Verifier (UV)
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Universal Verifier (UV)Verifier=UV, Backbone=GPT-5.22026.04 | 88 | 65 | 0.58 | 31 | 8 | |
| Universal VerifierVerifier=UV, Base Model=GPT-5.22026.04 | 88 | 65 | 0.58 | 31 | 8 | |
| WebVoy.Verifier=WebVoy., Backbone=GPT-5.22026.04 | 74 | 50 | 0.36 | 18 | 28 | |
| WebJudgeVerifier=WebJudge, Backbone=GPT-5.22026.04 | 74 | 46 | 0.31 | 29 | 26 | |
| WebJudgeVerifier=WebJudge, Backbone=o4-mini2026.04 | 64 | 44 | 0.26 | 12 | 40 | |
| WebJudgeVerifier=WebJudge, Base Model=o4-mini2026.04 | 64 | 44 | 0.26 | 12 | 40 | |
| WebVoy.Verifier=WebVoy., Backbone=GPT-4o2026.04 | 48 | 35 | 0.13 | 12 | 60 | |
| WebVoyagerVerifier=WebVoy., Base Model=GPT-4o2026.04 | 48 | 35 | 0.13 | 12 | 60 |