IAAILooped Flows
Thinking with Looped Flows
Trains the recurrence of a looped model with local denoising objectives at progressively decreasing noise levels with shared noise, so states learn to carry useful computation forward even though gradients cover only a few updates. Inference integrates a probability flow, and finer time grids spend more compute.
Inside the method
Noise → [ Denoiser + recurrent state ] × steps (decreasing noise) → solution
Simplified conceptual schematic. Consult the paper for the complete architecture.
- Recurrence family
- Core recurrence
- Depth control
- More compute through a finer temporal grid
- KV / state strategy
- Not applicable (non-autoregressive solver)
Reading note
Outperforms prior looped models overall on six reasoning benchmarks, with 58.8% on ARC-AGI-1 and 12.2% on ARC-AGI-2; different initial noise gives multiple valid solutions.
Sources checked 2026-10-28. This catalog does not imply independent reproduction.
Related works
Cite this work
@misc{loopedflows2026,
title = {Thinking with Looped Flows},
author = {Ayhan Suleymanzade and Chanhyuk Lee and Floor Eijkelboom and Nicholas M. Boffi and İsmail İlkan Ceylan and Jinwoo Kim},
year = {2026},
eprint = {2609.11801},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.11801}
}