Interactive Task Completion on ALFWorld cleaned 100-task v3 (mix of val-seen and val-unseen)
97Success RateReflexion
Evaluation results
| Method | Links | |
|---|---|---|
| ReflexionSetting / note=ReAct + reflection; multiple trials2026.05 | 97 | |
| ReflActSetting / note=goal-state reflection agent2026.05 | 93.3 | |
| GiGPOSetting / note=Qwen2.5-7B RL post-training2026.05 | 90.8 | |
| Const.-context skillsSetting / note=Qwen3-8B SFT+RL skill modules2026.05 | 89.6 | |
| GRPOSetting / note=Qwen2.5-7B RL post-training2026.05 | 77.6 | |
| SKILLREVISESetting / note=Qwen3-8B; 25-task absorption set2026.05 | 71 | |
| SimpleMem + GRPOSetting / note=memory + RL baseline2026.05 | 62.5 | |
| Mem0 + GRPOSetting / note=memory + RL baseline2026.05 | 54.7 | |
| ExpeLSetting / note=experience-to-insight baseline2026.05 | 46.3 | |
| EvolveRSetting / note=self-evolving memory baseline2026.05 | 43.8 | |
| MemPSetting / note=procedural memory baseline2026.05 | 41.4 | |
| Mem0Setting / note=external memory baseline2026.05 | 33.6 |