Looped Flows: Denoising-Trained Recurrent Reasoning Tops Loop-Model ARC-AGI
A team led by ayhozade publishes "Thinking with Looped Flows," proposing a new reasoning paradigm: training looped/recurrent architectures with local denoising objectives so that neural networks learn to "think by denoising," gaining stronger reasoning without extra parameters. The method reaches 58.8% on ARC-AGI-1, a new SoTA for looped models.
Confirmed
- Looped models repeatedly update a shared hidden state with the same parameters at inference time, effectively "thinking longer" and spending more test-time compute on harder problems
- Training is hard: full backpropagation across many loop steps is costly and unstable, so gradients are typically truncated after one or two updates, leaving early steps without useful learning signal
- The fix is a local denoising objective that gives every update an effective learning signal
- @burkov notes the method addresses the long-standing instability of multi-step backpropagation in looped models
- Sharers (@erikjbekkers, @LucaAmb, @NandoDF) frame it as a continuation of the flow-based language model line
- ARC-AGI-1 score: 58.8%
Why it matters
- It boosts reasoning without adding parameters, opening a new path for test-time compute scaling
- Denoising objectives give early loop steps direct learning signals, resolving a chronic training instability and potentially advancing recurrent architectures in language models
2026-09-11 ~ 2026-09-13 · 6 related posts
Primary sources
- [source] Looped Flows: Recurrent Denoising Reasoning Hits SoTA on ARC-AGI-1 and 2 — erikjbekkers · 2026-09-11
- Looped Flows: New Training Method Boosts Looped-Model Reasoning on 5 of 6 Benchmarks — omarsar0 · 2026-09-13
- [source] Looped Flows: A Training Method to Stabilize Recurrent Compute in Looped Models — burkov · 2026-09-13