About

Hi! My name is Linglong Meng, and my Chinese name is “孟令茏”. In Chinese, Linglong conveys intelligence and a quick, agile mind.

I am in the Honors Program with a major in Mathematics at the University of Michigan. I am very fortunate to be advised by Prof. Laura Balzano in the SPADA Lab at the Department of Electrical Engineering and Computer Science. I also closely collaborate with Prof. Gabriel Poesia in Poetry Lab.

Research

Research principle: How do structured parametric models acquire structured representations and computational mechanisms through learning, and how do these mechanisms enable generalization and reasoning?

My research interests grew out of my early work with classical sparse coding methods in machine learning. I study structured parametric models through modern mathematical perspectives, with a primary focus on deep neural networks.

By connecting training dynamics, optimization strategies, and representation structure, I hope to guide the design of future model architectures that are more efficient, capable, and interpretable. My interests span two closely related directions:

  1. Learning dynamics and optimization theory

    How do these models learn?

    I study training dynamics, the implicit bias of optimization algorithms, low-dimensional learning dynamics, and the emergence of features throughout training.

  2. Representations, computations, and reasonings

    What do these models learn and how do models use what they learned?

    I investigate structured representations, sparse and low-rank structure, functional circuits, and mechanistic explanations of learned computation.

News

I will present our work on Reliable Sparse Representation Learning at the 2026 Michigan AI Symposium. Welcome to our poster!

I will present our work on mechanistic interpretability of low-noise level diffusion at DeepMath 2026. Welcome to our poster, and see you in Columbus!