Embodied Task Completion on ALFWorld (Mixed)
98.6In-Domain Completion RateDPEPO
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| DPEPOBackbone=Qwen2.5-7B-Instruct2026.04 | 98.6 | 97.8 | 98.2 | |
| MCTS + SFTBackbone=Qwen2.5-7B-Instruct2026.04 | 91.4 | 92.5 | 91.95 | |
| MCTS + DPOBackbone=Qwen2.5-7B-Instruct2026.04 | 91.4 | 93.3 | 92.35 | |
| AgentGYM (SFT + DPO)Backbone=Qwen2.5-7B-Instruct2026.04 | 76.6 | 63.3 | 69.95 | |
| ETO (SFT + DPO)Backbone=Qwen2.5-7B-Instruct2026.04 | 70.3 | 74.2 | 72.25 | |
| MCTSBackbone=Qwen2.5-7B-Instruct2026.04 | 56.4 | 55.2 | 55.8 | |
| SFTBackbone=Qwen2.5-7B-Instruct2026.04 | 43 | 38.7 | 40.85 |