Sokoban on Sokoban (test)
90.6Success RateHiMAC
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| HiMACBackbone=Qwen3-VL-8B2026.03 | 90.6 | 7.32 | |
| HiMACBackbone=Qwen2.5-VL-7B2026.03 | 87.5 | 6.7 | |
| GRPOBackbone=Qwen2.5-VL-7B2026.03 | 83.9 | 5.39 | |
| HiMACBackbone=Qwen2.5-VL-3B2026.03 | 83.8 | 5.31 | |
| GRPOBackbone=Qwen3-VL-8B2026.03 | 82.9 | 5.24 | |
| GiGPOBackbone=Qwen2.5-VL-7B2026.03 | 82.8 | 5.27 | |
| GiGPOBackbone=Qwen2.5-VL-3B2026.03 | 81.1 | 4.86 | |
| GiGPOBackbone=Qwen3-VL-8B2026.03 | 78.1 | 4.67 | |
| GRPOBackbone=Qwen2.5-VL-3B2026.03 | 76.2 | 4.59 | |
| PromptBackbone=Qwen2.5-VL-7B2026.03 | 25 | -0.26 | |
| PromptBackbone=Qwen3-VL-8B2026.03 | 21.9 | -0.95 | |
| PromptBackbone=Qwen2.5-VL-3B2026.03 | 14.1 | -0.88 |