Melody (Zixuan) Li

Hi there!

I’m Melody Li, a research assistant at Mila – Quebec AI Institute. Broadly, I’m interested in developing neural language systems and deeply understanding their inner workings. My research integrates ideas from mechanistic interpretability, cognitive science, neuroscience, and learning dynamics to uncover how large language models learn and represent information.

I’m currently working with Prof. Taylor Webb on how reasoning circuits emerge during training. Previously, I worked with Prof. Blake Richards and Prof. Guillaume Lajoie, where we traced how training shapes internal representations. Using spectral methods from neuroscience, we discovered that models progress through three geometric phases during training, with the shift to compression marking when models transition from memorization to generalization.

I recently graduated from McGill University, majoring in Computer Science with double minors in Statistics and Interdisciplinary Life Sciences. Outside of research, I enjoy reading, travelling, playing the piano, and building puzzles and LEGO sets. Feel free to reach out!