ZoneTwelve/HRM-Sudoku

HRMACT Sudoku Solver

This repository hosts checkpoints for the Hierarchical Reasoning Model with Adaptive Computation Time (HRMACT), trained on Sudoku puzzles.

• Training steps: 3,000
• Batch size: 512
• Total training time: ~18 hours
• Checkpoint format: .safetensors

📌 Note: The original HRM paper recommends training for 10,000+ steps for best performance. These checkpoints are intended as a lightweight, educational reference.

---

🔗 Code & Usage

For the full implementation, training pipeline, and evaluation tools, see the main repo: 👉 ZoneTwelve/HRM-Sudoku

To run evaluation on a checkpoint:

python evaluate.py <checkpoint>

---

🔍 Evaluation

Evaluation is performed with evaluate.py.

Each checkpoint is evaluated 100 times per difficulty to estimate average performance.

Example format for results:

Model	very-easy	easy	medium	hard	extreme
checkpoint-1	0.00%	0.00%	0.00%	0.00%	0.00%
checkpoint-250	0.00%	0.00%	0.00%	0.00%	0.00%
checkpoint-500	0.00%	0.00%	0.00%	0.00%	0.00%
checkpoint-750	0.00%	0.00%	0.00%	0.00%	0.00%
checkpoint-1000	2.00%	0.00%	0.00%	0.00%	0.00%
checkpoint-1250	15.00%	4.00%	0.00%	0.00%	0.00%
checkpoint-1500	42.00%	18.00%	1.00%	0.00%	0.00%
checkpoint-1750	61.00%	40.00%	1.00%	1.00%	0.00%
checkpoint-2000	64.00%	28.00%	1.00%	1.00%	0.00%
checkpoint-2250	84.00%	67.00%	5.00%	2.00%	0.00%
checkpoint-2500	80.00%	66.00%	22.00%	8.00%	0.00%
checkpoint-2750	91.00%	81.00%	42.00%	13.00%	3.00%
checkpoint-3000	98.00%	95.00%	38.00%	13.00%	1.00%

---

📖 Citation

If you use this model in academic work, please cite:

@misc{wang2025hierarchicalreasoningmodel,
      title={Hierarchical Reasoning Model},
      author={Guan Wang and Jin Li and Yuhao Sun and Xing Chen and Changling Liu and Yue Wu and Meng Lu and Sen Song and Yasin Abbasi Yadkori},
      year={2025},
      eprint={2506.21734},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2506.21734},
}

---

📜 License

Apache License 2.0