Embodied Task Completion on ALFWorld (Action Success Rates)
100Pick Success RateRLSD
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| RLSDTraining Protocol=On-Policy Self-Distillation Methods2026.05 | 100 | 87.5 | 92.3 | 58.8 | 80 | 65.2 | 82 | |
| Skill1Training Protocol=Skill-Augmented RL Methods2026.05 | 100 | 98.6 | 97.3 | 99.2 | 96.1 | 96 | 97.5 | |
| SKILL0Training Protocol=Skill-Internalization RL Methods2026.05 | 100 | 84.6 | 91.1 | 84.2 | 88.9 | 55.5 | 85.9 | |
| GRPO+OPSDBackbone=Qwen2.5-3B-Instruct, Validation with skills=false2026.06 | 100 | 82.4 | 85.7 | 75 | 70 | 60 | 81.2 | |
| Skill-GRPOBackbone=Qwen2.5-7B-Instruct, Validation with skills=true2026.06 | 100 | 83.3 | 96.4 | 83.3 | 75 | 78.9 | 88.3 | |
| RLSDBackbone=Qwen2.5-7B-Instruct, Validation with skills=false2026.06 | 100 | 87.5 | 92.3 | 58.8 | 80 | 65.2 | 82 | |
| OPIDBackbone=Qwen2.5-7B-Instruct, Validation with skills=false2026.06 | 100 | 81.8 | 97.1 | 100 | 80.8 | 80 | 90 | |
| SKILLRLTraining Protocol=Skill-Augmented RL Methods2026.05 | 97.9 | 71.4 | 90 | 90 | 95.5 | 87.5 | 89.9 | |
| RetroAgentTraining Protocol=Skill-Augmented RL Methods2026.05 | 97.9 | 90.9 | 99.2 | 92.9 | 85.3 | 91 | 94.9 | |
| GiGPOTraining Protocol=Direct RL Methods2026.05 | 97.7 | 82.7 | 98.8 | 83.7 | 89.3 | 79.2 | 90.8 | |
| D2SkillTraining Protocol=Skill-Augmented RL Methods2026.05 | 97.1 | 100 | 75 | 87.5 | 100 | 78.6 | 90.6 | |
| SDARBackbone=Qwen2.5-3B-Instruct, Validation with skills=false2026.06 | 97.1 | 62.5 | 100 | 61.9 | 75 | 84.2 | 84.4 | |
| SDARTraining Protocol=On-Policy Self-Distillation Methods2026.05 | 94.7 | 75 | 100 | 86.7 | 68.2 | 78.9 | 85.9 | |
| SDARBackbone=Qwen2.5-7B-Instruct, Validation with skills=false2026.06 | 94.7 | 75 | 100 | 86.7 | 68.2 | 78.9 | 85.9 | |
| Skill-GRPOBackbone=Qwen2.5-3B-Instruct, Validation with skills=true2026.06 | 94.3 | 57.1 | 100 | 66.7 | 73.1 | 57.1 | 80.5 | |
| Skill-SDTraining Protocol=On-Policy Self-Distillation Methods2026.05 | 93.9 | 93.8 | 90.9 | 100 | 69.2 | 68.4 | 85.1 | |
| Skill-SDBackbone=Qwen2.5-7B-Instruct, Validation with skills=false2026.06 | 93.9 | 93.8 | 90.9 | 100 | 69.2 | 68.4 | 85.1 | |
| OPIDBackbone=Qwen2.5-3B-Instruct, Validation with skills=false2026.06 | 92.7 | 100 | 88.9 | 70 | 84.2 | 70 | 84.3 | |
| PPOTraining Protocol=Direct RL Methods2026.05 | 92.3 | 64 | 92.5 | 89.5 | 80.3 | 68.8 | 80.4 | |
| GRPO+OPSDTraining Protocol=On-Policy Self-Distillation Methods2026.05 | 91.4 | 61.5 | 100 | 87.5 | 76.5 | 52.2 | 80.4 | |
| GRPO+OPSDBackbone=Qwen2.5-7B-Instruct, Validation with skills=false2026.06 | 91.4 | 61.5 | 100 | 87.5 | 76.5 | 52.2 | 80.4 | |
| GRPOBackbone=Qwen2.5-3B-Instruct, Validation with skills=false2026.06 | 91.2 | 62.5 | 96.2 | 61.9 | 65 | 47.4 | 75 | |
| GRPOBackbone=Qwen2.5-7B-Instruct, Validation with skills=false2026.06 | 91.2 | 87.5 | 96.2 | 81 | 65 | 57.9 | 81.2 | |
| GRPOTraining Protocol=Direct RL Methods2026.05 | 90.8 | 66.1 | 89.3 | 74.7 | 72.5 | 64.7 | 77.6 | |
| Skill-GRPOBackbone=Qwen2.5-3B-Instruct, Validation with skills=false2026.06 | 88.9 | 71.4 | 58.8 | 70.6 | 40.7 | 29.2 | 60.2 | |
| SKILLCTraining Protocol=Skill-Internalization RL Methods2026.05 | 88.5 | 78.6 | 91.2 | 94.7 | 100 | 88.9 | 90.6 | |
| Skill-GRPOBackbone=Qwen2.5-7B-Instruct, Validation with skills=false2026.06 | 88.5 | 66.7 | 65.2 | 61.1 | 57.7 | 73.1 | 69.5 | |
| Skill-SDBackbone=Qwen2.5-3B-Instruct, Validation with skills=false2026.06 | 88.2 | 50 | 96.2 | 52.4 | 65 | 57.9 | 73.4 | |
| RLSDBackbone=Qwen2.5-3B-Instruct, Validation with skills=false2026.06 | 87.9 | 75 | 90.9 | 75 | 73.1 | 68.4 | 79.7 | |
| RLOOTraining Protocol=Direct RL Methods2026.05 | 87.6 | 78.2 | 87.3 | 81.3 | 71.9 | 48.9 | 75.5 | |
| SDARBackbone=Qwen3-1.7B-Instruct, Validation with skills=false2026.06 | 73.5 | 25 | 76.9 | 33.3 | 40 | 36.8 | 53.9 | |
| GRPOBackbone=Qwen3-1.7B-Instruct, Validation with skills=false2026.06 | 71.1 | 41.7 | 36.4 | 40 | 31.8 | 31.6 | 46.1 | |
| OPIDBackbone=Qwen3-1.7B-Instruct, Validation with skills=false2026.06 | 65.9 | 72.7 | 66.7 | 40 | 63.2 | 45 | 58.9 | |
| EvolveRTraining Protocol=Skill-Augmented RL Methods2026.05 | 64.9 | 33.3 | 46.4 | 13.3 | 33.3 | 33.3 | 43.8 | |
| ReflexionTraining Protocol=Training-free Methods2026.05 | 62 | 41.6 | 44.9 | 30.9 | 36.3 | 23.8 | 42.7 | |
| Mem0Training Protocol=Training-free Methods2026.05 | 54 | 55 | 26.9 | 36.4 | 20.8 | 7.7 | 33.6 | |
| Skill-SDBackbone=Qwen3-1.7B-Instruct, Validation with skills=false2026.06 | 52.9 | 37.5 | 69.2 | 42.9 | 60 | 36.8 | 52.3 | |
| Skill-PromptBackbone=Qwen2.5-3B-Instruct, Validation with skills=true2026.06 | 51.7 | 66.7 | 48.4 | 0 | 4.3 | 10 | 28.9 | |
| Skill-PromptBackbone=Qwen2.5-7B-Instruct, Validation with skills=true2026.06 | 51.7 | 50 | 32.3 | 5.3 | 4.3 | 0 | 23.4 | |
| OPSDTraining Protocol=On-Policy Self-Distillation Methods2026.05 | 50 | 60 | 22.7 | 21.4 | 17.6 | 9.5 | 32.8 | |
| OPSDBackbone=Qwen2.5-7B-Instruct, Validation with skills=false2026.06 | 50 | 60 | 22.7 | 21.4 | 17.6 | 9.5 | 32.8 | |
| RLSDBackbone=Qwen3-1.7B-Instruct, Validation with skills=false2026.06 | 50 | 37.5 | 61.5 | 19 | 50 | 21.1 | 42.2 | |
| OPSDBackbone=Qwen2.5-3B-Instruct, Validation with skills=false2026.06 | 48.8 | 41.7 | 16.7 | 0 | 15.8 | 16.7 | 28.1 | |
| ReActTraining Protocol=Training-free Methods2026.05 | 48.5 | 35.4 | 34.3 | 13.2 | 18.2 | 17.6 | 31.2 | |
| VanillaBackbone=Qwen2.5-3B-Instruct, Validation with skills=false2026.06 | 44.4 | 11.1 | 6.2 | 15.4 | 28.6 | 12.5 | 21.9 | |
| GRPO+OPSDBackbone=Qwen3-1.7B-Instruct, Validation with skills=false2026.06 | 38.2 | 50 | 30.8 | 28.6 | 30 | 21.1 | 32 | |
| VanillaBackbone=Qwen2.5-7B-Instruct, Validation with skills=false2026.06 | 36.1 | 22.2 | 3.1 | 0 | 0 | 0 | 12.5 | |
| Zero-ShotTraining Protocol=Training-free Methods2026.05 | 33.4 | 21.6 | 19.3 | 6.9 | 2.8 | 3.2 | 14.8 | |
| Skill-GRPOBackbone=Qwen3-1.7B-Instruct, Validation with skills=true2026.06 | 31.4 | 42.9 | 51.9 | 8.3 | 11.5 | 7.1 | 28.1 | |
| Skill-GRPOBackbone=Qwen3-1.7B-Instruct, Validation with skills=false2026.06 | 27.6 | 54.5 | 22.7 | 27.3 | 0 | 19.2 | 21.1 | |
| OPSDBackbone=Qwen3-1.7B-Instruct, Validation with skills=false2026.06 | 26.3 | 33.3 | 9.1 | 0 | 4.5 | 5.3 | 14.1 | |
| VanillaBackbone=Qwen3-1.7B-Instruct, Validation with skills=false2026.06 | 25 | 22.2 | 3.1 | 0 | 21.4 | 4.2 | 12.5 | |
| ExpeLTraining Protocol=Training-free Methods2026.05 | 21 | 67 | 55 | 52 | 71 | 6 | 46.3 | |
| Skill-PromptBackbone=Qwen3-1.7B-Instruct, Validation with skills=true2026.06 | 10.3 | 50 | 16.1 | 0 | 0 | 5 | 9.4 |