Thinking with Looped Flows
The paper proposes “looped flows,” a training and inference method for recurrent looped models that aims to make extra test-time computation useful.
The approach trains recurrent updates with local denoising objectives, using decreasing noise levels and shared noise to carry computation across steps. At inference, it integrates a probability-flow velocity from the learned denoiser while maintaining recurrent state. The authors report gains across six reasoning benchmarks, including 58.8% test accuracy on ARC-AGI-1 and 12.2% on ARC-AGI-2. ArXiv · AI/CL/LG's note
The approach trains recurrent updates with local denoising objectives, using decreasing noise levels and shared noise to carry computation across steps. At inference, it integrates a probability-flow velocity from the learned denoiser while maintaining recurrent state. The authors report gains across six reasoning benchmarks, including 58.8% test accuracy on ARC-AGI-1 and 12.2% on ARC-AGI-2. ArXiv · AI/CL/LG's note
score 5