---
license: mit
library_name: pytorch
tags:
  - looped-transformer
  - recursive-reasoning
  - sudoku
datasets:
  - sapientinc/sudoku-extreme
---

# InfiLoop: Sudoku-Extreme

The 6.83M-parameter InfiLoop model from [Scaling to Tens of Thousands of Test-Time Iterations with Loop-Native Attention Residuals](https://arxiv.org/abs/2610.11570), trained on Sudoku-Extreme. This is the EMA checkpoint at step 71,610 used for the paper's Sudoku results.

Code: [github.com/pixeli99/InfiLoop](https://github.com/pixeli99/InfiLoop/tree/main/sudoku)

## Usage

```bash
git clone https://github.com/pixeli99/InfiLoop.git && cd InfiLoop/sudoku
pip install -r requirements.txt
python data.py --output data/sudoku
hf download pengxiang/InfiLoop-Sudoku model.pt --local-dir checkpoints

# evaluation across depths (paper: 97.9% exact accuracy)
python verify.py --data data/sudoku --checkpoint checkpoints/model.pt --output verified.json

# fixed-depth evaluation, 1,872 Transformer layers
python run.py evaluate --data data/sudoku --checkpoint checkpoints/model.pt \
  --h-cycles 6 --l-cycles 12 --segments 12
```

## Citation

```bibtex
@article{li2026infiloop,
  title   = {Scaling to Tens of Thousands of Test-Time Iterations with Loop-Native Attention Residuals},
  author  = {Li, Pengxiang and Muhtar, Dilxat and He, Di and Su, Guinan and Yin, Lu and Liu, Shiwei},
  journal = {arXiv preprint arXiv:2610.11570},
  year    = {2026}
}
```