Loopie shows that an AI model can reuse the same Transformer layers twice and still beat a larger standard model trained for the same amount of time. The saved memory is turned into more model capacity.

Normal Transformers get bigger by stacking more and more different layers.

A looped Transformer tries another approach： reuse the same layer more than once， so the model can process information again without storing a completely new layer for every step.

This idea already existed， but it usually lost to simply building a larger normal Transformer once training cost was counted.

Loopie proposes the practical fix. It stores fewer layers， runs each layer twice， and uses the saved GPU memory to train a wider model more efficiently.

Under the paper's matched training-time experiment， Loopie eventually beat the larger conventional Transformer baseline.

AI models may not need endless new parameters to become better. Carefully reusing the parameters they already have could become another way to scale them.

• arxiv. org/abs/2607.16051

Title： "Loop the Loopies！"