Mengxi Wu 武梦溪

Research

My research focuses on the foundations of efficiency in large language model (LLM) pre-training. I am interested in methods that scale from small proxy models to trillion-parameter foundation models, as well as the mathematical principles underlying training dynamics. I work on hyperparameter optimization, analytically linking regularization, model architecture, and dataset properties. I also study optimization algorithms and develop new optimizers. Previously, my work included transfer learning with theoretical guarantees, adversarial machine learning for 3D point cloud processing, and streaming algorithms for time-varying volume data.