Truncate or Not? A Scaling Study of Looped Transformers with Truncated Backpropagation
Truncate or Not? A Scaling Study of Looped Transformers with Truncated Backpropagation
Looped Transformers train better than truncated backpropagation with the same computational cost through full backtracking, extended regularity shows that when the model size is adjustable, two loops are the best. Looped Transformers increase computational depth by reusing shared layers without increasing the number of trainable parameters, usually using backtracking training (BPTT), but are often truncated to the last few loops to save training computational cost, which may lead to gradient bias and weaken the loop advantage.
Published on Oct 20
Authors:
,
,
,
,
,
,
Abstract
Looped transformers trained with full backpropagation through time outperform truncated alternatives under matched compute, with scaling laws identifying two loops as optimal when model size is adjustable.
Looped transformers reuse a shared block of layers to increase computational depth without increasing the number of trainable parameters. They are trained with backpropagation through time (BPTT). To save training compute, BPTT is often truncated to the last few loops, but truncation yields biased gradients that may weaken the benefit of looping. In this work, we study the value of truncation in looped transformer training in a systematic way: we jointly vary model size, training tokens, and forward and backward loop counts across nearly 600 training runs and fit a unified scaling law that captures the effects of looping and truncation through effective parameters. Despite the training-compute savings of truncation, our fitted law favors full BPTT when forward and backward loop counts are jointly optimized. In matched-compute runs, full BPTT through fewer loops consistently outperforms truncated settings of the same compute. The law further shows that at fixed model size, the compute-optimal loop count and training-token budget both grow with compute. When model size is also free, the law identifies two fully trained loops as the best looped setting. Experimentally, this setting achieves lower validation loss than the compute-optimal non-looped model with about 27% fewer non-embedding parameters and only 8% more training compute.
View arXiv page View PDF Project page GitHub Add to collection
Get this paper in your agent:
hf papers read 2610.26030
Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash
Models citing this paper 0
No model linking this paper
Cite arxiv.org/abs/2610.26030 in a model README.md to link it from this page.
Datasets citing this paper 0
No dataset linking this paper
Cite arxiv.org/abs/2610.26030 in a dataset README.md to link it from this page.
Spaces citing this paper 0
No Space linking this paper
Cite arxiv.org/abs/2610.26030 in a Space README.md to link it from this page.
Collections including this paper 0
No Collection including this paper
Add this paper to a collection to link it from this page.
Source: Hugging Face Daily Papers · huggingface.co