Maze Navigation on Maze Hard
97.66AccuracyGPT-5
Evaluation results
| Method | Links | |
|---|---|---|
| GPT-5category=Proprietary Models2025.11 | 97.66 | |
| OpenAI o3category=Proprietary Models2025.11 | 93.36 | |
| PTRM# Params=7M, K (stochastic rollouts)=100, D (depth)=16, sigma (exploration parameter)=1.0, architecture=TRM-Att2026.05 | 86.73 | |
| TRM# Params=7M, architecture=TRM-Att2026.05 | 85.3 | |
| Standard TRM, our reproduction# Params=7M, architecture=TRM-Att2026.05 | 83.8 | |
| OpenAI o4-minicategory=Proprietary Models2025.11 | 78.52 | |
| HRM# Params=27M2026.05 | 74.5 | |
| Claude 4.5 Sonnetcategory=Proprietary Models2025.11 | 68.36 | |
| Gemini 2.5 Procategory=Proprietary Models2025.11 | 63.28 | |
| WMAct2025.11 | 50.59 | |
| PPO - Interactivemode=interactive2025.11 | 36.52 | |
| Qwen3-14Bcategory=Opensource Models2025.11 | 28.52 | |
| PPO - EntirePlanmode=single-turn output2025.11 | 26.51 | |
| Qwen3-8Bcategory=Opensource Models2025.11 | 17.76 | |
| GPT-4ocategory=Proprietary Models2025.11 | 1.56 | |
| Qwen2.5-32B-Instructcategory=Opensource Models2025.11 | 0.39 | |
| Qwen3-8B-Ownbackbone=Qwen3-8B2025.11 | 0.2 | |
| Qwen2.5-7B-Instructcategory=Opensource Models2025.11 | 0 |