Tuesday 15 September 2026
Evening EditionRSS

ZOTPAPER

News without the noise

ZOTPAPER
AI & ML

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.

PaperIndependentcs.LGarXiv:2609.11801v1 ↗
Ayhan Suleymanzade · Chanhyuk Lee · Floor Eijkelboom · Nicholas M. Boffi · İsmail İlkan Ceylan · Jinwoo Kim
Research Digest··2 min read
Read the paper →
Suleymanzade et al. propose looped flows, a method that trains recurrent neural networks for many-step inference using local denoising losses. This overcomes the difficulty of training early updates to support later ones when backpropagation is truncated. The authors report state-of-the-art accuracy among looped models on six reasoning benchmarks, including 58.8% on ARC-AGI-1 and 12.2% on ARC-AGI-2.

What they did

The authors introduce looped flows, which combine a recurrent state with a denoising objective. During training, they apply local denoising losses at each step, with noise levels progressively decreasing over time and noise samples shared across steps. This temporal association encourages the network to learn recurrent states that transfer useful computation across many updates, even though gradients only cover a few steps. At inference, the velocity of a probability flow—parameterized by the learned denoiser—is integrated, coupled with the recurrent state. Harder problems are solved by using a finer temporal grid (more integration steps), and multiple valid predictions can be obtained from different initial noise samples. Experiments were conducted on six reasoning benchmarks, including both single-solution and multi-solution tasks such as ARC-AGI-1 and ARC-AGI-2.

Key findings

  • Looped flows achieve 58.8% test accuracy on ARC-AGI-1 and 12.2% on ARC-AGI-2, outperforming all prior looped models in the comparison.
  • The method outperforms prior state-of-the-art looped models on overall average across six reasoning benchmarks.
  • The approach supports multiple distinct valid outputs from different initial noise samples, enabling multi-solution reasoning.
  • Inference computation can be scaled by increasing the number of integration steps, directly trading off compute for accuracy.

Why it matters

Looped flows demonstrate a principled way to train recurrent networks that can perform many steps of computation during inference without requiring long backpropagation. This addresses a key bottleneck in looped architectures and opens a path toward models that flexibly allocate more computation to harder problems, similar to human reasoning. The ability to produce multiple valid predictions is also valuable for tasks with inherent ambiguity.

Caveats

The experiments compare against prior looped models but not against non-looped or transformer-based methods that may achieve higher accuracy on these benchmarks. The training complexity and computational cost of the denoising approach are not fully analyzed. The reported gains may be sensitive to the choice of noise schedule and shared noise construction. Generalization to other domains (e.g., language or image generation) remains unexplored.

§
newspaper

Research Digest

Articles published under the Zotpaper byline are synthesized from multiple source publications by our AI editor and reviewed by our editorial process. Each story combines reporting from credible outlets to give readers a balanced, comprehensive view.

Editorial ProcessSubscribe