arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2601.01332cs.CLcs.LG
收藏

高效计算训练:基于测试时间计算意识的提前停止

FLOP-Efficient Training: Early Stopping Based on Test-Time Compute Awareness

Ascend Team, Huawei Technologies, Toronto, Canada(华为技术有限公司,多伦多,加拿大);University of Waterloo, Waterloo, Canada(滑铁卢大学,滑铁卢,加拿大)

Hossam Amer, Maryam Dialameh, Hossein Rajabzadeh, Walid Ahmed, Weiwei Zhang, Yang Liu

AI总结:

本文提出了一种基于测试时间计算意识的高效训练方法,通过提前停止算法减少训练计算量,同时保持或提升模型准确性。

摘要:

增加训练计算量(以FLOPs衡量)长期以来已被证明可以提高大语言模型的准确性,但训练仍然耗费资源。先前的研究表明,增加测试时间计算(TTC)——例如通过迭代采样——可以使较小的模型在总体成本更低的情况下与 much 更大的模型竞争或超越。我们引入了TTC意识训练,其中中间检查点和相应的TTC配置可以一起达到或超过完全训练模型的准确性,同时要求显著更少的训练FLOPs。基于这一见解,我们提出了一种提前停止算法,该算法联合选择检查点和TTC配置以在不牺牲准确性的情况下最小化训练计算。为了使这一方法实用,我们开发了一种高效的TTC评估方法,以避免穷举搜索,并且我们正式化了一个平衡点界限,以确定增加推理计算是否补偿了减少的训练计算。实验表明,训练FLOPs可以减少高达92%,同时保持并有时显著提高准确性。这些结果强调了在模型开发中平衡训练和推理计算的新视角,使更快的部署周期和更频繁的模型刷新成为可能。代码将公开发布。

英文摘要:

Scaling training compute, measured in FLOPs, has long been shown to improve the accuracy of large language models, yet training remains resource-intensive. Prior work shows that increasing test-time compute (TTC)-for example through iterative sampling-can allow smaller models to rival or surpass much larger ones at lower overall cost. We introduce TTC-aware training, where an intermediate checkpoint and a corresponding TTC configuration can together match or exceed the accuracy of a fully trained model while requiring substantially fewer training FLOPs. Building on this insight, we propose an early stopping algorithm that jointly selects a checkpoint and TTC configuration to minimize training compute without sacrificing accuracy. To make this practical, we develop an efficient TTC evaluation method that avoids exhaustive search, and we formalize a break-even bound that identifies when increased inference compute compensates for reduced training compute. Experiments demonstrate up to 92\% reductions in training FLOPs while maintaining and sometimes remarkably improving accuracy. These results highlight a new perspective for balancing training and inference compute in model development, enabling faster deployment cycles and more frequent model refreshes. Codes will be publicly released.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑