Looped transformers reuse computation across steps or depth. The appeal is an adjustable trade-off between parameter count, inference compute and iterative refinement—but the design also changes optimization and evaluation questions.
Loopie and LoopFormer are useful entry points because they make distinct claims about scaling and variable compute. They are research materials, not product documentation, and their reported results should be read with their methods and publication status in view.
A careful explainer separates architecture ideas from deployment recommendations and avoids treating a single benchmark result as a general measure of model quality.