2026 07 Icml

This week, we are presenting two papers at the International Conference on Machine Learning (ICML) in Seoul, Korea. In Enhancing LLM Training via Spectral Clipping [CODE], we propose SPECTRA, a simple wrapper that consistently improves coordinate-wise optimizers (AdamW, Signum, etc.), leading to better convergence and generalization. In Improved Convergence Analysis of Topology Dependence in Decentralized SGD, we develop a tighter convergence analysis of decentralized SGD, offering a more precise understanding of how network topology affects its convergence rate.