Looped Flows Train Recurrent Reasoning With Local Denoising and Hit 58.8% on ARC-AGI-1
arXiv·medium signal
arXiv 2609.11801 avoids backpropagating through many recurrent updates. It trains a looped model with local denoising objectives at decreasing noise levels, then runs inference as integration of a probability flow. A finer time grid buys more compute at test time, and different noise seeds give multiple valid answers. Across six reasoning benchmarks it beats prior looped models, reaching 58.8% on ARC-AGI-1 and 12.2% on ARC-AGI-2.