Lanhao Niu

I’m Lanhao Niu, a Master’s student in Cognitive Science at the University of Edinburgh. My research interests lie in mechanistic interpretability, representation learning, and reasoning in large language models. I am broadly interested in how models internally represent structured knowledge, how those representations support reasoning and computation, and how these mechanisms can be studied in controlled settings.

One strand of my work, with Prof. Sashank Varma, studies the representational geometry of language models: how concepts, numbers, and context-dependent meanings are structured internally, and which aspects of this geometry are shared across model architectures. A second, related strand, with Prof. Siddharth Narayanaswamy, studies learned probabilistic inference, asking whether language models can use symbolic program structure and sampler trajectories to learn state-conditioned MCMC proposals that generalize across probabilistic programs.

Across both strands, I’m interested in the relationship between representation and reasoning: how internal structure supports computation, and what that structure can tell us about how models behave.

I’m currently looking for research assistant and PhD opportunities!