arXiv cs.CL· Cheng Liang, Tengfei Zhang, Ya Zhang, Yanfeng Wang, Zhouhan Lin·· 11 小时前AI 评分34
UpLoop：不增加预训练算力，用 upcycling 训练循环 Transformer
UpLoop: Training Looped Transformers via Upcycling at No Extra Pretraining Compute
AI 导读
UpLoop 提出一种预训练方案，通过 upcycling 训练循环 Transformer，在最终参数量、训练 token 总量和训练 FLOPs 上与标准稠密基线完全持平。
正文 · 原文
Abstract:Looped language models can strengthen reasoning through repeated computation with shared weights, increasing execution depth without adding independent parameters. However, each additional pass increases training floating-point operations (FLOPs), making these gains more expensive to obtain. To address this cost, we propose UpLoop, a pretraining schedule that uses upcycling to train looped Transformers while exactly matching a standard dense baseline in final parameter count, total training tokens, and training FLOPs. Specifically, UpLoop begins with a shallow Transformer, progressively grows its depth by copying learned layers, and uses the computation saved early to support repeated execution of a shared middle block later in training. Experiments across model scales show improvements in language modeling and downstream performance, with the average score across 22 downstream tasks increasing from 22.99 to 27.52 at 1.3B parameters. These gains also extend to mathematical reasoning after supervised fine-tuning.
| Subjects: | Machine Learning (cs.LG); Computation and Language (cs.CL) |
| Cite as: | arXiv:2610.41720 [cs.LG] |
| (or arXiv:2610.41720v1 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2610.41720 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Tengfei Zhang [view email]
[v1]
Fri, 30 Oct 2026 13:29:35 UTC (216 KB)
来源：arXiv cs.CL · arxiv.org