Training recurrent models with local denoising objectives at each step—rather than backpropagating through all steps—enables better iterative reasoning while keeping training tractable.
This paper introduces looped flows, a training method for models that solve problems by iteratively refining their internal state over multiple steps. Instead of training all steps together (which is computationally hard), the method trains each step to denoise progressively, creating a flow that can be integrated during inference.