Embodied Reasoning on ALFWorld Unseen
99.3Pick&Place Success RateRESKILL
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| RESKILLModel Scale=Qwen3-8B, Strategy=w/ RL2026.06 | 99.3 | 91.2 | 100 | 95.3 | 92.7 | |
| INSPOModel Scale=Qwen3-8B, Strategy=w/ RL2026.06 | 88.1 | 89.2 | 81.5 | 87.8 | 85.5 | |
| RESKILLModel Scale=Qwen3-4B-Instruct, Strategy=w/ RL2026.06 | 87.4 | 89.6 | 96.3 | 89.6 | 89.8 | |
| GRPOModel Scale=Qwen3-4B-Instruct, Strategy=w/ RL2026.06 | 86.2 | 72 | 68.5 | 75.9 | 75.3 | |
| INSPOModel Scale=Qwen3-4B-Instruct, Strategy=w/ RL2026.06 | 82.9 | 87 | 83.3 | 79.6 | 78 | |
| SkillRLModel Scale=Qwen3-8B, Strategy=w/ RL2026.06 | 82.9 | 78.3 | 94.4 | 82.6 | 85.8 | |
| GRPOModel Scale=Qwen3-8B, Strategy=w/ RL2026.06 | 82.1 | 81.8 | 79.6 | 81.6 | 81.1 | |
| MemRL (+RL)Model Scale=Qwen3-8B, Strategy=w/ RL2026.06 | 82.1 | 83.3 | 75.9 | 81.6 | 82.3 | |
| MemRL (+RL)Model Scale=Qwen3-4B-Instruct, Strategy=w/ RL2026.06 | 78.9 | 76.6 | 94.4 | 79.1 | 80.2 | |
| Best†Model Scale=Qwen3-8B, Strategy=w/ Skill2026.06 | 76.4 | 80.4 | 72.2 | 74.6 | 70.7 | |
| Sonnet 4.5Model Scale=Sonnet 4.5, Strategy=Proprietary2026.06 | 73.2 | 72 | 100 | 76.1 | 79.2 | |
| EvolveRModel Scale=Qwen3-8B, Strategy=w/ RL2026.06 | 70.7 | 89.2 | 83.3 | 82.8 | 83.1 | |
| EvolveRModel Scale=Qwen3-4B-Instruct, Strategy=w/ RL2026.06 | 69.9 | 77.9 | 94.4 | 77.9 | 77.3 | |
| SkillRLModel Scale=Qwen3-4B-Instruct, Strategy=w/ RL2026.06 | 69.9 | 86.9 | 88.9 | 82.1 | 83.9 | |
| ReActModel Scale=Qwen3-8B, Strategy=Baseline2026.06 | 69.2 | 65 | 66.7 | 64.9 | 62 | |
| Best‡Model Scale=Qwen3-8B, Strategy=w/ Evolve2026.06 | 57.7 | 76.4 | 57.4 | 65.7 | 65.9 | |
| Best†Model Scale=Qwen3-4B-Instruct, Strategy=w/ Skill2026.06 | 50.4 | 68 | 74.1 | 63.4 | 57.5 | |
| Best‡Model Scale=Qwen3-4B-Instruct, Strategy=w/ Evolve2026.06 | 45.5 | 58.2 | 55.6 | 52.7 | 53.6 | |
| ReActModel Scale=Qwen3-4B-Instruct, Strategy=Baseline2026.06 | 30.8 | 28.9 | 66.7 | 32.8 | 35 |