Web Navigation and Automation on WorkArena Held-out Tasks (test)
70Success RateClaude-3.5-Sonnet
Evaluation Results
| Method | Links | |
|---|---|---|
| Claude-3.5-SonnetBackbone Model=Claude-3.5-Sonnet2025.07 | 70 | |
| o1-MiniBackbone Model=o1-Mini2025.07 | 68.6 | |
| Llama-3.1-405B-InstructBackbone Model=Llama-3.1-405B2025.07 | 58.6 | |
| GPT-4oBackbone Model=GPT-4o2025.07 | 55.7 | |
| Llama-3.3-70B Instruct (Teacher)Backbone Model=Llama-3.3-70B, Model Role=Teacher2025.07 | 44 | |
| Llama-3.1-70B-InstructBackbone Model=Llama-3.1-70B2025.07 | 32.9 | |
| Llama-3.1-8B SFT+RL (Ours)Backbone Model=Llama-3.1-8B, Training Protocol=SFT+RL2025.07 | 28.8 | |
| GPT-4o-MiniBackbone Model=GPT-4o-Mini2025.07 | 28.6 | |
| Qwen-2.5-72B Instruct (Teacher)Backbone Model=Qwen-2.5-72B, Model Role=Teacher2025.07 | 27 | |
| Llama-3.1-8B SFT (Ours)Backbone Model=Llama-3.1-8B, Training Protocol=SFT2025.07 | 26.4 | |
| Qwen-2.5-7B SFT+RL (Ours)Backbone Model=Qwen-2.5-7B, Training Protocol=SFT+RL2025.07 | 25 | |
| Qwen-2.5-7B SFT (Ours)Backbone Model=Qwen-2.5-7B, Training Protocol=SFT2025.07 | 21 | |
| Llama-3.1-8B RL (Ours)Backbone Model=Llama-3.1-8B, Training Protocol=RL2025.07 | 11.5 | |
| Qwen-2.5-7B Instruct (Student)Backbone Model=Qwen-2.5-7B, Model Role=Student2025.07 | 10 | |
| Qwen-2.5-7B RL (Ours)Backbone Model=Qwen-2.5-7B, Training Protocol=RL2025.07 | 7 | |
| Llama-3.1-8B Instruct (Student)Backbone Model=Llama-3.1-8B, Model Role=Student2025.07 | 4.2 |