Web Navigation on WorkArena L1
79.1Success RateGenericAgent
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| GenericAgentModel=GPT-5, Access=Proprietary2026.06 | 79.1 | — | |
| WebChallengerModel=GLM-4-32B + Qwen2.5-VL-7B, Access=Open-Source, Evaluation Strategy=zero-shot2026.06 | 70.9 | — | |
| GenericAgentModel=Claude 3.5 Sonnet, Access=Proprietary2026.06 | 56.4 | — | |
| Agent-as-AnnotatorsModel=Qwen3.5-9B, Access=Open-Source, Evaluation Strategy=fine-tuned2026.06 | 51.5 | — | |
| GenericAgentModel=GPT-oss-120b, Access=Open-Source, Evaluation Strategy=zero-shot2026.06 | 50.9 | — | |
| GenericAgentModel=GPT-4o, Access=Proprietary2026.06 | 45.5 | — | |
| GenericAgentModel=GPT-4o-mini, Access=Proprietary2026.06 | 27 | — | |
| Qwen2.5-7B-Instruct + WEASELTraining Steps=10K steps, Training Method=WEASEL2026.05 | 7.6 | 9.7 | |
| Qwen2.5-7B-Instruct + Pruning + SamplingTraining Steps=10K steps, Training Method=Pruning + Sampling2026.05 | 7 | 9.7 | |
| Qwen2.5-7B-Instruct + FullTraining Steps=52K steps, Training Method=Full2026.05 | 5.2 | 1 | |
| Qwen2.5-7B-InstructZero-shot transfer=true2026.05 | 4.8 | — | |
| Qwen2.5-7B-Instruct + PruningTraining Steps=52K steps, Training Method=Pruning2026.05 | 1.8 | 1.7 |