“Thinking with Looped Flows” Looped models can spend more compute by recurrently updating hidden states, but truncated backprop makes it hard for early steps to learn computations that stay useful later. This paper trains recurrence through progressively easier denoising objectives with decreasing noise, turning reasoning into a flow from noise to solution while carrying a recurrent state across steps. This able to obtain more stable test-time scaling and 58.8% on ARC-AGI-1 and 12.2% on…