What is a looped flow model?

A reasoning model that reuses one denoising network across many recurrent steps, where each step denoises a progressively cleaner solution candidate. Training uses local denoising objectives instead of backpropagation through time, and inference integrates the learned denoiser like a probability flow, so more compute steps means better answers.

How much better is Looped Flows than TRM on ARC-AGI?

With the same 5–7M parameter architecture, it improves pass@2 accuracy from 44.6% to 58.8% on ARC-AGI-1 and from 7.8% to 12.2% on ARC-AGI-2, and resolves 90.9% of TRM's Sudoku-Extreme failures.

Can it produce multiple different valid solutions?

Yes. Because inference samples from a probability flow, different initial noise yields different valid answers — 91.4% coverage of distinct valid solutions on 8×8 N-Queens from 20 inferences, which deterministic looped models cannot do.

Looped Flows: 44.6% → 58.8% on ARC-AGI-1

The looped-reasoning line — HRM, TRM, and their descendants — has been the quiet overperformer on ARC-AGI: tiny 5–27M parameter models beating far larger transformers on structured puzzles. Their weakness was always the training recipe. Full backpropagation through dozens of recurrent steps is expensive and unstable, so everyone truncates gradients to one or two local updates. The early recurrence steps never learn to set up the later ones. A new paper, Thinking with Looped Flows (arXiv:2609.11801, Sept 10), fixes exactly that — and the jump is the largest the line has seen since TRM itself.

The problem: early loop steps fly blind

Looped models reuse one network over and over, updating a hidden state until a solution crystallizes. But because gradients only flow through the last step or two, an update at iteration 3 is never trained to be useful at iteration 20. The paper quantifies the fallout on Sudoku-Extreme: TRM fails on 12.6% of instances, and 88.3% of those failures are non-convergence — the recurrence simply never settles. The architecture is asked to do iterative refinement, but the training signal doesn't teach iteration.

The method: steal from diffusion, not backprop

The fix is elegant: train the recurrence with local denoising objectives borrowed from flow matching. Each recurrent step denoises an interpolant at a progressively lower noise level, with noise shared across steps. No gradient flows between steps — yet the noise schedule chains the steps together, because what step t produces is literally the input distribution for step t+1. The early updates get supervision without backprop-through-time. At inference, the denoiser is integrated like a probability flow (ODE or SDE), so you can buy more compute with a finer temporal grid, and — a genuine bonus — different initial noise samples yield different valid solutions, which looped models built on argmax decode could never do.

The results: state of the art at 7M parameters

Same TRM architecture, same budget, new training objective:

mindmap
  root((Looped Flows
vs TRM, same 5–7M arch))
    ARC-AGI-1 pass@2
      44.6% → 58.8%
    ARC-AGI-2 pass@2
      7.8% → 12.2%
    Sudoku-Extreme
      87.4% → 97.9%
    Maze-Hard
      85.3% → 86.7%
    Failure recovery
      90.9% of TRM failures resolved

ARC-AGI-1 goes from 44.6% to 58.8% ± 1.8 (59.5% with 5-sample ensembling); ARC-AGI-2 from 7.8% to 12.2% ± 1.9. That beats GRAM (52.0/11.1 at 10M params) and dwarfs HRM (40.3/5.0 at 27M). On the multi-solution benchmarks (N-Queens, Graph Coloring) it wins every metric, with coverage of distinct valid solutions reaching 91.4% on 8×8 N-Queens. The ablation table is the most informative part: strip the interpolant and ARC-AGI-1 drops to 51.5%; strip both time conditioning and the interpolant and it collapses to 43.6% — below TRM's baseline. The flow machinery isn't decoration; it's the load-bearing piece.

Limitations the abstract doesn't mention

First, ARC-AGI-2 at 12.2% is still nowhere near frontier LLMs with test-time scaffolding — this line wins on parameter efficiency, not absolute capability. Second, Maze-Hard is only "competitive" with FPRM, not better, and the paper doesn't break down why. Third, the SDE vs ODE ablation shows the stochastic sampler matters mostly for diversity (NQ10 coverage: 54.7% ODE vs 61.5% SDE), meaning the multi-solution story is partly an artifact of sampling temperature, not recurrence quality. Fourth, everything is evaluated on TRM's own preprocessing and small curated datasets (Sudoku, Maze, ARC, N-Queens, Graph Coloring) — no evidence yet this transfers to natural-language reasoning. And the "resolves 90.9% of TRM failures" claim is on the same benchmark family the model was tuned on.

Why builders should care

Two things are portable. One: local objectives can train deep recurrences. If you're building any iterate-and-refine system (diffusion LMs, latent reasoners, iterative solvers), you don't need BPTT — you need a curriculum that makes each step's output the next step's input distribution. Two: stochastic recurrent state gives you solution diversity for free. Sampling 5 flows beat every 20–128-sample ensemble baseline on Sudoku at a fraction of the inference cost. The ARC-AGI efficiency frontier just moved, and it moved at 7M parameters.