Agent Task on ScienceWorld (Avg Reward, Success Rate)
67.5Success RateE3-TIR
Evaluation results
| Method | Links | ||
|---|---|---|---|
| E3-TIRBackbone=Qwen2.5-3B-Instruct, Expert Data Trajectories=2K2026.04 | 67.5 | — | |
| W2SG with MCTS2025.07 | 66.8 | 58.2 | |
| SFT-then-RLBackbone=Qwen2.5-3B-Instruct, Expert Data Trajectories=2K2026.04 | 65 | — | |
| Ceiling Model2025.07 | 63.5 | 56.9 | |
| Zero-RLBackbone=Qwen2.5-3B-Instruct, Expert Data Trajectories=2K2026.04 | 63.5 | — | |
| W2SG with Tree DPO2025.07 | 61.1 | 55.4 | |
| SFT Strong ModelBase Model=Llama-2-13b+ETO2025.07 | 61.1 | 54.9 | |
| SFT Strong Model + Best of N2025.07 | 60.7 | 55.3 | |
| SFT Strong ModelBase Model=Llama-2-13b+SFT2025.07 | 59.2 | 53.6 | |
| SFT Weak ModelBase Model=Llama-2-7b+SFT2025.07 | 55.5 | 41.2 | |
| Only SFTBackbone=Qwen2.5-3B-Instruct, Expert Data Trajectories=2K2026.04 | 43 | — |