MisakiTaro0414/looped-flows

Thinking with Looped Flows

Ayhan Suleymanzade^1,*   Chanhyuk Lee²   Floor Eijkelboom³   Nicholas M. Boffi⁴   İsmail İlkan Ceylan^5,6,7,†   Jinwoo Kim^2,†

¹EPFL   ²KAIST   ³University of Amsterdam   ⁴Carnegie Mellon University   ⁵TU Wien   ⁶AITHYRA   ⁷University of Oxford   ^*Work done while a visiting researcher at AITHYRA   ^†Equal advising

Paper · Code · Project page

Overview

Looped flows train a looped model as a stateful denoiser with local denoising objectives at decreasing noise levels and a shared noise sample, so the recurrent state learns to carry computation across steps. Inference integrates the learned probability flow (ODE, or SDE with $\gamma>0$) coupled with the recurrent state: more steps give more computation, and different initial noise gives different valid solutions.

Available Checkpoints

File	Task	Size	Accuracy
looped-flows-sudoku.pt	Sudoku-Extreme	21 MB	98.2
looped-flows-maze.pt	Maze-Hard 30x30	28 MB	86.9
looped-flows-arc1.pt	ARC-AGI-1	1.8 GB	61.4
looped-flows-arc2.pt	ARC-AGI-2	2.5 GB	12.5

Usage

git clone https://github.com/MisakiTaro0414/looped-flows.git
cd looped-flows                      # set up the environment and build the dataset as in the README

# Download one checkpoint (or scripts/download_checkpoints.sh for all four)
hf download MisakiTaro0414/looped-flows looped-flows-sudoku.pt --local-dir checkpoints

# Evaluate with the paper's inference setting for the task
CKPT=checkpoints/looped-flows-sudoku.pt TASK=sudoku bash scripts/eval.sbatch

Training Details

• 2-layer, 512-wide stateful denoiser (5–7M parameters) built on the TRM backbone
• Looped flow training with $k=16$ steps, shared noise, ACT halting, Adam-atan2, batch size 768, EMA 0.999
• Per-task recipes (noise scale, time sampler, pseudotargets, weight decay) in scripts/task_recipe.sh of the code repository
• Trained on single NVIDIA GPUs (H100, L40S, A40, A10, RTX 3090; 24–94 GB)

Citation

@article{suleymanzade2026thinking,
  title={Thinking with looped flows},
  author={Suleymanzade, Ayhan and Lee, Chanhyuk and Eijkelboom, Floor and Boffi, Nicholas M and Ceylan, Ismail Ilkan and Kim, Jinwoo},
  journal={arXiv preprint arXiv:2609.11801},
  year={2026}
}

License

MIT