Thinking With Looped Flows Hits 58.8% on ARC-AGI-1, Joining the Looped-Flow Reasoning Wave

alec_helbling · x · 2026-09-17

The FRM authors highlight a converging research direction: looped flow architectures pairing parallel generation with iterative refinement. The new paper Thinking with Looped Flows (arXiv 2609.11801) trains recurrence with local denoising objectives — building temporal association across objectives via progressively decreasing, shared noise — so early updates learn to support later ones despite gradients covering only a few steps. Inference integrates a probability-flow velocity field on a finer temporal grid, and different noise seeds yield multiple valid predictions. Across six reasoning benchmarks it beats prior looped-model SOTA, scoring 58.8% on ARC-AGI-1 and 12.2% on ARC-AGI-2.

Original post →

More from Research

Research channel →