One-step next-observation prediction on Maze (test)
98Token F1PatchWorld-Residual
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| PatchWorld-ResidualBackbone=DeepSeek-V4-Flash, LLM Usage=CodeGen, Infer. LLM?=No2026.05 | 98 | 93 | |
| Word2WorldBackbone=Qwen3.5-4B, LLM Usage=SFT, Infer. LLM?=Yes2026.05 | 97 | 89 | |
| PatchWorld-ResidualBackbone=Qwen3-Coder-480B, LLM Usage=CodeGen, Infer. LLM?=No2026.05 | 97 | 91 | |
| PatchWorld-SimpleBackbone=DeepSeek-V4-Flash, LLM Usage=CodeGen, Infer. LLM?=No2026.05 | 90 | 80 | |
| WorldCoderBackbone=Mimo-v2.5, LLM Usage=CodeGen, Infer. LLM?=No2026.05 | 88 | 78 | |
| PatchWorld-SimpleBackbone=Mimo-v2.5, LLM Usage=CodeGen, Infer. LLM?=No2026.05 | 88 | 75 | |
| PatchWorld-ResidualBackbone=Mimo-v2.5, LLM Usage=CodeGen, Infer. LLM?=No2026.05 | 87 | 82 | |
| PoE-WorldBackbone=Mimo-v2.5, LLM Usage=CodeGen, Infer. LLM?=No2026.05 | 86 | 78 | |
| LLM-DirectBackbone=Qwen3-Coder-480B, LLM Usage=ICL, Infer. LLM?=Yes2026.05 | 83 | 75 | |
| LLM-DirectBackbone=Mimo-v2.5, LLM Usage=ICL, Infer. LLM?=Yes2026.05 | 83 | 75 | |
| LLM-DirectBackbone=DeepSeek-V4-Flash, LLM Usage=ICL, Infer. LLM?=Yes2026.05 | 83 | 73 | |
| WorldCoderBackbone=Qwen3-Coder-480B, LLM Usage=CodeGen, Infer. LLM?=No2026.05 | 83 | 76 | |
| PoE-WorldBackbone=Qwen3-Coder-480B, LLM Usage=CodeGen, Infer. LLM?=No2026.05 | 83 | 77 | |
| PatchWorld-SimpleBackbone=Qwen3-Coder-480B, LLM Usage=CodeGen, Infer. LLM?=No2026.05 | 80 | 69 | |
| PoE-WorldBackbone=DeepSeek-V4-Flash, LLM Usage=CodeGen, Infer. LLM?=No2026.05 | 76 | 61 | |
| WorldCoderBackbone=DeepSeek-V4-Flash, LLM Usage=CodeGen, Infer. LLM?=No2026.05 | 73 | 61 |