Thinking with Looped Flows

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

Loops train poorly
Looped models backprop through only one or a few updates, so early steps never learn to help later ones.
Denoise, step by step
Each loop step denoises a noisy solution, with noise shrinking over steps and the same noise sample shared.
Better, scalable reasoning
Beats prior looped models on five of six benchmarks; finer time grids add compute, different noise gives varied answers.

Which comes out on top on ARC-AGI-1 pass@2?

try it

Drag the number of inference steps up and watch accuracy change.

Inference steps8 steps

A finer time grid spends more compute and lifts accuracy, then gains level off.Illustrative model, not the paper's data

6 terms

Looped model Network reusing the same weights repeatedly to update a hidden state.'

Like restoring a photo through progressively clearer versions, each pass keeping notes for the next, instead of guessing the whole picture at once.

Where it breaks

Here the notes are learned hidden features, trained only with local losses.

From the paper