GradES: Significantly Faster Training in Transformers with Gradient-Based Early Stopping
Published in arXiv preprint, 2025
A novel gradient-based early stopping method that operates at the component level within transformer architectures, achieving 1.57x to 7.22x training speedup while improving model generalization.
Recommended citation: Qifu Wen, Xi Zeng, Zihan Zhou, Shuaijun Liu, Mehdi Hosseinzadeh, Ningxin Su, Reza Rawassizadeh. "GradES: Significantly Faster Training in Transformers with Gradient-Based Early Stopping." arXiv preprint arXiv:2509.01842, 2025.
Download Paper