---
id: 20261031-T0-04
title: "数据受限时代怎么扩模型？超参数迁移加集成来救"
title_en: "Hyperparameter Transfer and Ensembling Improve Data-Constrained Scaling"
url: https://ai.daily.yangsir.net/daily/20261031-T0-04
issue_date: 2026-10-31
publish_date: 2026-10-30T04:00:00.000Z
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2610.37842
---

# 数据受限时代怎么扩模型？超参数迁移加集成来救

这篇arXiv论文研究数据受限条件下的大模型扩展。当算力预算增长快于高质量训练数据增长时，模型训练必须同时考虑算力和数据约束。论文指出已有数据受限扩展方案存在两个重要问题，并提出结合超参数迁移与集成的方法来改进。具体改进幅度和实验设置摘要中未给出。对训练资源有限、数据量不足的团队，这类方法可能决定在有限数据下能否继续扩大模型规模。

## English Version

**Hyperparameter Transfer and Ensembling Improve Data-Constrained Scaling**

This arXiv paper studies scaling large language models under data constraints. As compute budgets outpace growth in high-quality training data, training must account for both compute and data limits. The paper identifies two important problems with existing data-constrained scaling recipes and proposes combining hyperparameter transfer with ensembling to improve them. Specific improvement margins and experimental setups are not given in the abstract. For teams with limited compute and insufficient data, such methods may determine whether they can keep scaling models at all.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2610.37842)

**详情页**：https://ai.daily.yangsir.net/daily/20261031-T0-04

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*