Looped Flows: A Training Method to Stabilize Recurrent Compute in Looped Models
burkov · x · 2026-09-13
Looped models reuse the same parameters to iteratively update a shared hidden state, gaining extra inference-time compute without growing model size. But training them is hard: full backprop through many recurrent steps is costly and unstable, so gradients are usually truncated after one or a few updates, leaving early steps unsupervised and recurrences diverging or settling into wrong solutions.
The proposed looped flows approach trains a stateful denoiser on a sequence of local denoising objectives at progressively lower noise levels, with shared noise across steps to encourage temporal consistency of the recurrent state; inference then integrates the resulting probability flow.
Related event: Looped Flows: Denoising-Trained Recurrent Reasoning Tops Loop-Model ARC-AGI(6 posts)→
More from Research
- NVIDIA Research releases Foundation4D: unified 4D perception model turns camera video into 3D scenes, no lidar needed — Scobleizer · 2026-10-28
- Moving Atoms Engine turns real-world spaces and cities into high-resolution 4D models for sim — Scobleizer · 2026-10-28
- Rachel Thomas points to classic paper on articulating the value of not building tech — math_rachel · 2026-10-28
- Meta's Omnilingual MT covers 1,600 languages, 1B–8B models rival 70B LLMs — kalyan_kpl · 2026-10-28
- Supercharging LLM classification with logprobs: soft classifiers you can threshold-tune — JnBrymn · 2026-10-28
- cryoET Data Portal releases three datasets benchmarking crossed laser phase plate — ZhongingAlong · 2026-10-28