跳到正文
arXiv cs.AI· Abhash Kumar Jha, Diana Alexandra Onu\c{t}u, Neeratyoy Mallik, Swagatam Haldar, Sam Laing, Niccol\`o Ajroldi, Shiwei Liu, Joaquin Vanschoren, Aaron Klein·· 4 小时前AI 评分26

用贝叶斯优化摊销 Scaling Law 构建成本

Amortizing Scaling Law Construction Costs

AI 导读

研究提出将 Scaling Law 的数据收集建模为贝叶斯优化问题,并在采集过程中逐步扩大计算预算,以匹配实践中按算力排序评估配置的方式,显著提升恢复效率。再用代理模型虚构的评估补全已观测配置,从而在无需训练全部配置的情况下准确拟合 Scaling Law,计算量最多节省 10–100 倍。

正文 · 原文

View PDF HTML (experimental)

Abstract:Scaling laws guide the design choices for training large foundation models, but deriving them involves training an exhaustive grid over hyperparameters, token budgets, and parameter counts, which is computationally expensive. Fitting a scaling law, however, only requires the best-loss frontier across compute scales, discarding most of the trained configurations. We propose a framework for efficient scaling law construction that formulates data collection as a Bayesian optimization problem, and introduce metrics for comparing scaling law fitting methods under constrained compute budgets. We find that progressively expanding the compute budget during acquisition, mirroring the compute-ordered evaluation of configurations in practice, substantially improves recovery efficiency. Augmenting the observed configurations with surrogate-fantasized evaluations then recovers the broader experimental grid, allowing accurate scaling law fitting without training every configuration. Together, these can closely match scaling law fits over a full dense grid at computational savings of up to $10\text{--}100\times$.
Comments: 5 pages, 2 figures, workshop
Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
Cite as: arXiv:2609.05016 [cs.LG]
  (or arXiv:2609.05016v2 [cs.LG] for this version)
  https://doi.org/10.48550/arXiv.2609.05016

arXiv-issued DOI via DataCite

Submission history

From: Abhash Jha [view email]
[v1] Fri, 4 Sep 2026 11:25:41 UTC (2,670 KB)
[v2] Tue, 17 Nov 2026 13:09:58 UTC (1,484 KB)

来源:arXiv cs.AI · arxiv.org