Web agent meta-evaluation on AGENTREWARDBENCH
89.21Overall PrecisionGUIDE
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| GUIDEBackbone=gemini-3.0-flash2026.04 | 89.21 | 100 | 78.9 | 91.7 | 100 | 90.9 | |
| GUIDE (o4-mini)Backbone=o4-mini2026.04 | 86.67 | 100 | 76.6 | 89.7 | 100 | 87 | |
| Rule-based2026.04 | 83.8 | 25 | 85.2 | 79 | 100 | 83.3 | |
| WebJudge (o4-mini)Backbone=o4-mini2026.04 | 82 | 100 | 74.5 | 81.2 | 100 | 90 | |
| GPT-4o (A11y Tree)Backbone=GPT-4o, Representation=Accessibility Tree2026.04 | 69.8 | 77.8 | 63 | 70.2 | 94.6 | 63 | |
| Claude 3.7 SonnetBackbone=Claude 3.7 Sonnet2026.04 | 69.4 | 71.4 | 64.8 | 69.3 | 85.3 | 66.7 | |
| Autonomous Eval2026.04 | 67.6 | 83.3 | 61.2 | 67.6 | 96.4 | 59.3 | |
| NNetNav2026.04 | 52.5 | 20.8 | 54.5 | 54.3 | 77.3 | 43.2 |