The paper introduces looped flows, a training method for recurrent models that perform iterative reasoning at inference. Standard training backpropagates through only a few updates, which makes it hard for early updates to support later ones. Looped flows instead train each update with a denoising objective, using progressively decreasing noise levels and shared noise-target pairs. This incentivizes recurrent states to carry useful information across steps, even with truncated gradients. After training, inference integrates a probability flow parameterized by the learned denoiser, enabling more computation via a finer temporal grid and multiple valid predictions from different initial noise samples. The method achieves 58.8% on ARC-AGI-1 and 12.2% on ARC-AGI-2, and outperforms prior looped models across six reasoning benchmarks.
Looped flows enable deeper reasoning by chaining denoising steps
A new training scheme uses local denoising objectives with time-dependent noise to make recurrent networks learn transferable computation across many updates.
Independent
Ayhan Suleymanzade · Chanhyuk Lee · Floor Eijkelboom · Nicholas M. Boffi · İsmail İlkan Ceylan · Jinwoo Kim
Research Digest··2 min read
In one line
Looped flows train recurrent models with local denoising objectives, allowing more inference steps and improving reasoning benchmark performance.
Read it as
Suleymanzade et al. propose looped flows, a method that trains recurrent neural networks for many-step inference using local denoising losses.
The authors introduce looped flows, which combine a recurrent state with a denoising objective.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.