Looped transformers improve performance without extra training cost
UpLoop: Training Looped Transformers via Upcycling at No Extra Pretraining Compute
Machine LearningComputation and Language
Summary
Training language models to think more deeply by repeating certain parts usually means more computing power is needed. The authors propose UpLoop, which starts with a small model and grows it by reusing layers, balancing training so it doesn’t use more computing resources than usual. This approach helps the models get better at understanding language and solving problems, like math, without extra cost in training. Their tests showed that this method improves results across many language tasks and mathematical reasoning after fine-tuning.
What this means in practice
- •For natural language processing engineers: Implement language models that improve accuracy without increasing training costs by using UpLoop’s progressive layer reuse technique.
- •For machine learning platform operators: Reduce compute resource demands during language model training by adopting UpLoop’s efficient pretraining schedule to match baseline requirements.
Authors
Cheng Liang, Tengfei Zhang, Ya Zhang, Yanfeng Wang, Zhouhan Lin
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.