Interactive Decision-making on AgentGym WebShop, ALFWorld, and BabyAI (test)
78.8WebShop Success RateSTRIDE
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| STRIDEType=Token-level RL, Base Model=Qwen2.5-VL-7B2026.06 | 78.8 | 89.6 | 84.5 | 84.3 | |
| HET-DAPOType=Token-level RL, Base Model=Qwen2.5-VL-7B2026.06 | 77.1 | 88.4 | 83 | 82.8 | |
| AgentSTaRType=Offline RL, Base Model=Qwen2.5-VL-7B2026.06 | 76.5 | 88.6 | 82.7 | 82.6 | |
| DPOType=Offline RL, Base Model=Qwen2.5-VL-7B2026.06 | 75.3 | 86.5 | 78.3 | 80 | |
| AgentTraj-L-SFTType=SFT, Base Model=Qwen2.5-VL-7B2026.06 | 73.5 | 83.2 | 74.2 | 77 | |
| PPOType=Online RL, Base Model=Qwen2.5-VL-7B2026.06 | 68.5 | 83.5 | 69.8 | 73.9 | |
| RWRType=Offline RL, Base Model=Qwen2.5-VL-7B2026.06 | 68.2 | 76.5 | 82.1 | 75.6 | |
| LLM-PlannerType=Prompting, Base Model=Qwen2.5-VL-7B2026.06 | 18.9 | 68.5 | 82.5 | 56.6 |