# Loopie 循环 Transformer 以少胜多

- 来源：Rohan Paul (@rohanpaul_ai)
- 发布时间：2026-08-04 17:34
- AIHOT 分数：42
- AIHOT 链接：https://aihot.virxact.com/items/cmsegq3i30rhlro2egw2pfvgl
- 原文链接：https://x.com/rohanpaul_ai/status/2084573782803665306

## AI 摘要

Loopie 证明，AI 模型可重复使用同一 Transformer 层两次，在相同训练时间内超越更大的标准模型。它通过减少层数、每层运行两次，将节省的 GPU 内存用于训练更宽的模型。在匹配训练时间的实验中，Loopie 最终击败了更大的常规 Transformer 基线。

## 正文

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！"