Web Navigation on VisualWebArena
54Published VWA Success RateSGV
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SGVWhere compute is spent (Eq. 3 term)=Cverify ↑ (two-pass self-grounded verifier), Paid=per-task2026.05 | 54 | 2.2 | |
| WALTWhere compute is spent (Eq. 3 term)=Cpre ↑ (offline tool discovery, 100 steps/tool), Paid=pre-eval2026.05 | 52.9 | — | |
| WALTModel=GPT-5, Access=Proprietary2026.06 | 52.9 | — | |
| WebChallengerModel=GLM-4-32B + Qwen3-VL-4B-Instruct, Access=Open-Source, Evaluation Strategy=zero-shot2026.06 | 48.7 | — | |
| Mobile-Agent-v3.5Model=Qwen3-VL-32B, Access=Open-Source, Evaluation Strategy=fine-tuned2026.06 | 46.6 | — | |
| Agent-as-AnnotatorsModel=Qwen3.5-9B, Access=Open-Source, Evaluation Strategy=fine-tuned2026.06 | 33.9 | — | |
| ExACTWhere compute is spent (Eq. 3 term)=Nrollout ↑ (reflective MCTS), Paid=per-task2026.05 | 33.7 | — | |
| GenericAgentModel=GPT-4o, Access=Proprietary2026.06 | 26.7 | — | |
| Tree SearchWhere compute is spent (Eq. 3 term)=Nrollout ↑ (best-first search), Paid=per-task2026.05 | 26.4 | — | |
| WebDreamerModel=Qwen2-VL-7B, Access=Open-Source, Evaluation Strategy=fine-tuned2026.06 | 21.9 | — | |
| GenericAgentModel=Claude 3.5 Sonnet, Access=Proprietary2026.06 | 21 | — | |
| GenericAgentModel=GPT-4o-mini, Access=Proprietary2026.06 | 16.9 | — | |
| Tree SearchModel=Llama-3-70B-Instruct, Access=Open-Source, Evaluation Strategy=zero-shot2026.06 | 16.7 | — |