AI research
Looped Flows Trains Iterative Reasoning Through Local Denoising, Reaching 58.8% Accuracy on ARC-AGI-1
The new method arranges each recurrent-network update into a continuous flow from noise toward an answer, allowing the model to learn state across steps without fully backpropagating through every iteration. At inference time, more integration steps can be used for additional compute, while different initial noise samples can produce multiple candidate solutions.
Recurrent reasoning models can repeatedly apply the same set of parameters, trading fewer weights for more inference-time computation. During training, however, backpropagation is typically limited to only one or a few update steps. When gradients are truncated between steps, early states may not learn to preserve the information that later steps actually need; simply running the loop longer can also lead to stagnation or accumulating errors.
Looped Flows, submitted on September 10, reframes this credit assignment problem as local denoising. The training data first forms a temporal trajectory from the same answer and an initial noise sample. Each update step sees progressively less noise while sharing the problem conditioning and recurrent state. Every step has its own denoising target, and gradients can be stopped between steps. Because adjacent tasks follow the same trajectory and the later one is slightly easier, the state is still incentivized to carry computational results that will be useful in the next step. Its underlying denoiser is built on a Tiny Recursive Model–style architecture rather than a conventional large autoregressive language model.
At inference time, instead of repeatedly processing a fixed clean input, the method numerically integrates the velocity field defined by the stateful denoiser, gradually transporting noise toward a discrete answer. A finer time grid corresponds to more test-time compute, while changing the initial noise can yield different solutions—an especially useful property for problems with multiple valid answers, such as N-Queens and graph coloring. The paper compares Looped Flows with existing looped models across six reasoning benchmarks, reporting test accuracy of 58.8% on ARC-AGI-1 and 12.2% on ARC-AGI-2.
This is not yet a drop-in reasoning replacement for general-purpose LLMs. The results are concentrated on algorithmic and abstract-reasoning tasks with structured verification, while the latency, stability, and calibration costs of increasing the number of integration steps must be evaluated separately. Independent reproduction results are also currently lacking. The next key steps are to release the complete code and training configuration, and to determine whether local denoising can scale to natural language, long-context settings, and tasks without a single verifiable answer.