Looped Flows Paper Hits 58.8% on ARC-AGI-1 With Recurrent Denoising Reasoning

MiniMax_AI · x · 2026-09-14

A new arXiv paper, "Thinking with Looped Flows," combines flow/diffusion-style denoising with recurrent reasoning. Instead of backpropagating through one or two recurrent updates (BPTT), it trains the recurrence with local denoising objectives — progressively decreasing noise levels and shared noise encourage hidden states that transfer useful computation over time. Inference integrates a probability-flow velocity parameterized by the learned denoiser, so more compute at test time means a finer temporal grid, and different initial noise samples yield multiple valid predictions. Across six reasoning benchmarks, looped flows beat prior looped-model SOTA, reaching 58.8% on ARC-AGI-1 and 12.2% on ARC-AGI-2.

Related event: Looped Flows Paper Combines Diffusion Denoising with Recurrent Reasoning(2 posts)→

Original post →

More from Research

Research channel →