Looped Flows: New Training Method Boosts Looped-Model Reasoning on 5 of 6 Benchmarks

omarsar0 · x · 2026-09-13

A new paper introduces Looped Flows, a looped architecture with a novel training method: local denoising objectives borrowed from flow models, with stepwise-decreasing shared noise samples that tie each hidden-state update to the next. This fixes the usual problem that gradients only reach the last updates, and the authors report wins over prior looped models on five of six reasoning benchmarks—more reasoning without extra parameters.

Related event: Looped Flows: denoising-based recurrent reasoning tops ARC-AGI(3 posts)→

Original post →

More from Research

Research channel →