Yanzhe Chen
I am a member of the FLINT research group at Shanghai Jiao Tong University. My research interests include optimization and large language model training.
Research
Muon Sublates the Edge of Stability in LLM Pretraining
Yanzhe Chen, Qifang Zhao, Xiaoxiao Xu, and Fanghui Liu
We study how conditional loss balance and consecutive-update alignment separate under no-momentum Muon, from controlled networks to language-model pretraining.
Blog
How Muon Sublates the Edge of Stability
A visual guide to how Muon retains a loss-side edge while uncoupling it from temporal reversal, through update geometry, two thresholds, and training evidence.