Web Agent Navigation on WebArena
65.6Success RateClaude Sonnet 4.6
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Claude Sonnet 4.6Model Type=Proprietary2026.05 | 65.6 | — | |
| CUA-GYM-A17BTraining=RLVR on CUA-GYM, Base Model=Qwen3.5-397B-A17B2026.05 | 56 | — | |
| Qwen3.5-397B-A17BModel Type=Base Model2026.05 | 54 | — | |
| DRIVE2026.04 | 45.9 | — | |
| CUA-GYM-A3BTraining=RLVR on CUA-GYM, Base Model=Qwen3.5-35B-A3B2026.05 | 44.5 | — | |
| Go-Browse2026.04 | 41.2 | — | |
| Qwen3.5-35B-A3BModel Type=Base Model2026.05 | 40.8 | — | |
| AgentOccam2026.04 | 37.2 | — | |
| SteP2026.04 | 37 | — | |
| AdaRubric-DABackbone=Qwen2.5-7B, Fine-tuning Protocol=DPO2026.03 | 27.8 | 15.5 | |
| SkillWeaver2026.04 | 25.1 | — | |
| AdaRubric-GMBackbone=Qwen2.5-7B, Fine-tuning Protocol=DPO2026.03 | 25.1 | 12.8 | |
| AdaRubric-WMBackbone=Qwen2.5-7B, Fine-tuning Protocol=DPO2026.03 | 24.3 | 12 | |
| WebArena2026.04 | 21.4 | — | |
| PrometheusBackbone=Qwen2.5-7B, Fine-tuning Protocol=DPO2026.03 | 21 | 8.7 | |
| G-EvalBackbone=Qwen2.5-7B, Fine-tuning Protocol=DPO2026.03 | 20.1 | 7.8 | |
| Random pairsBackbone=Qwen2.5-7B, Fine-tuning Protocol=DPO2026.03 | 17.4 | 5.1 | |
| SFT-Success onlyBackbone=Qwen2.5-7B, Fine-tuning Protocol=SFT2026.03 | 16.7 | 4.4 | |
| BaseBackbone=Qwen2.5-7B, Fine-tuning Protocol=Zero-shot2026.03 | 12.3 | — |