Agent Interaction on Aggregate FTWP, ScienceWorld, WebShop
89.74Format Faithfulness RateGemma2-9B + MISE
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Gemma2-9B + MISEBackbone=Gemma2-9B-Instruct, Technique=MISE2026.04 | 89.74 | 28.07 | 23.25 | |
| Gemma2-9B + PRMBackbone=Gemma2-9B-Instruct, Technique=PRM2026.04 | 88.9 | 26.63 | 21.45 | |
| Gemma2-9B + PPOBackbone=Gemma2-9B-Instruct, Technique=PPO2026.04 | 86.42 | 25.87 | 20.67 | |
| GPT-4oBackbone=GPT-4o2026.04 | 82.13 | 40.4 | 38.29 | |
| LLaMA3-8B + PRMBackbone=LLaMA3-8B-Instruct, Technique=PRM2026.04 | 80.7 | 32.75 | 27.36 | |
| LLaMA3-8B + PPOBackbone=LLaMA3-8B-Instruct, Technique=PPO2026.04 | 78.9 | 28.2 | 22.33 | |
| LLaMA3-8B + MISEBackbone=LLaMA3-8B-Instruct, Technique=MISE2026.04 | 77.57 | 36.48 | 30.47 | |
| LLaMA3-8B + RFTBackbone=LLaMA3-8B-Instruct, Technique=RFT2026.04 | 77.47 | 31.29 | 26.1 | |
| Qwen2-7B + MISEBackbone=Qwen2-7B-Instruct, Technique=MISE2026.04 | 73.14 | 32.12 | 26.22 | |
| GPT-4o-miniBackbone=GPT-4o-mini2026.04 | 72.54 | 32.55 | 29.64 | |
| Qwen2-7B + PPOBackbone=Qwen2-7B-Instruct, Technique=PPO2026.04 | 71.34 | 27.86 | 22.15 | |
| Qwen2-7B + PRMBackbone=Qwen2-7B-Instruct, Technique=PRM2026.04 | 70.92 | 28.28 | 22.95 | |
| LLaMA3-8B + ReActBackbone=LLaMA3-8B-Instruct, Technique=ReAct2026.04 | 68.68 | 28.3 | 23.63 | |
| LLaMA3-8BBackbone=LLaMA3-8B-Instruct2026.04 | 67.83 | 28.92 | 23.06 | |
| LLaMA3-8B + online DPOBackbone=LLaMA3-8B-Instruct, Technique=online DPO2026.04 | 67.47 | 29.54 | 24.4 | |
| Gemma2-9BBackbone=Gemma2-9B-Instruct2026.04 | 61.23 | 9.93 | 9.74 | |
| Qwen2-7BBackbone=Qwen2-7B-Instruct2026.04 | 44.85 | 19.6 | 15.72 |