Decoded
Feed

Thinking with Looped Flows

arXiv:2609.11801LGAI10 sep 20261 min readHN ×11PDF

Train looping reasoning models as step-by-step denoisers so more loops at test time help.

58.8%

ARC-AGI-1 test accuracy

Same 7M architecture as TRM, much higher score

How it works

  1. Loops train poorly

    Looped models backprop through only one or a few updates, so early steps never learn to help later ones.

  2. Denoise, step by step

    Each loop step denoises a noisy solution, with noise shrinking over steps and the same noise sample shared.

  3. Better, scalable reasoning

    Beats prior looped models on five of six benchmarks; finer time grids add compute, different noise gives varied answers.

updatenext steppredictionflow stateat t=1Problem+ noiseStatefuldenoiserRecurrentstateODE/SDE stepSolutionupdatenext steppredictionflow stateat t=1Problem+ noiseStatefuldenoiserRecurrentstateODE/SDE stepSolution
From the paper