Looped Flows trains a 7-million-parameter reasoner through denoising

A 7-million-parameter recurrent model called Looped Flows reports 58.8% accuracy on ARC-AGI-1 and 12.2% on ARC-AGI-2. Those scores improve on the same-size Tiny Recursive Model by 14.2 and 4.4 percentage points, respectively. The paper, Thinking with Looped Flows, adapts diffusion-style training to recurrent reasoning so the model can improve with additional inference steps and sample multiple valid solutions.

Why long loops strain training

ARC-AGI evaluates abstract reasoning through grid-transformation puzzles. Each task provides a few input-output examples, then asks the system to infer the underlying rule and apply it to a new grid. The benchmark favors models that can discover compact procedures from limited evidence.

Small recurrent architectures provide the relevant baseline. The 27-million-parameter Hierarchical Reasoning Model reported 40.3% on ARC-AGI-1 by repeatedly refining a latent state. The TRM repository later reduced the architecture to a shared 7-million-parameter network and reported 44.6% on ARC-AGI-1.

Training becomes difficult when a model performs many recurrent updates at inference but receives gradients through only one or two updates during training. Full backpropagation through time requires storing the entire update sequence, increasing memory use and optimization difficulty as the loop grows. Truncating that sequence saves resources, but early updates receive little information about how they affect the eventual answer.

A noise schedule aligns the updates

Looped Flows assigns each recurrent update a local denoising objective. Early steps operate on heavily corrupted solution states, while later steps receive progressively cleaner states. Every step uses the same underlying noise sample, creating a coherent trajectory from noise to answer and encouraging the hidden state to preserve useful information across updates.