Agreement with process human labels on CUAVerifierBench (Internal Dataset (n=140))
81AccuracyUniversal Verifier (UV)
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Universal Verifier (UV)Verifier=UV, Backbone=GPT-5.22026.04 | 81 | 86 | 0.59 | 24 | 4 | |
| Universal VerifierVerifier=UV, Base Model=GPT-5.22026.04 | 81 | 86 | 0.59 | 24 | 4 | |
| WebJudgeVerifier=WebJudge, Backbone=o4-mini2026.04 | 66 | 70 | 0.32 | 40 | 25 | |
| WebJudgeVerifier=WebJudge, Base Model=o4-mini2026.04 | 66 | 70 | 0.32 | 40 | 25 | |
| WebVoy.Verifier=WebVoy., Backbone=GPT-5.22026.04 | 64 | 65 | 0.34 | 49 | 10 | |
| WebVoy.Verifier=WebVoy., Backbone=GPT-4o2026.04 | 62 | 70 | 0.17 | 31 | 52 | |
| WebVoyagerVerifier=WebVoy., Base Model=GPT-4o2026.04 | 62 | 70 | 0.17 | 31 | 52 | |
| WebJudgeVerifier=WebJudge, Backbone=GPT-5.22026.04 | 61 | 57 | 0.3 | 59 | 4 |