Andrew Liu

I am a second-year Masters student in the NLP Group at the University of Toronto, supervised by Professor Gerald Penn.

I recieved my Bachelors in Computer Science from the University of Waterloo in 2024. During my undergrad, I had the good fortune to work with Professor Kimon Fountoulakis and Professor Jimmy Lin.

Email  /  CV  /  Google Scholar  /  LinkedIn

My research in the past has focused on both interpretability in language models as well as multilingual capabilities of information retrieval systems. Currently my research focuses on topological and geometric methods for understanding language models. I additionally do a bit of work in NLP for healthcare.

Current interpretability techniques identify patterns that are interpretable both syntactically and semantically; however, these patterns do not constitute knowledge.

Despite the large amounts of incidental multilingual texts in training data, they play a very minor role in cross-lingual transfer capabilities.

Influence functions have emerged as a feasible way to assign blame to datapoints. Our findings show influence scores may encode semantics in the same strange way that Sentence-BERT vectors do, and possibly exhibit a higher degree of robustness to semantics-preserving variance.

A controlled benchmark containing a curated corpus to fairly evaluate deep-research agents. It enables clear, reproducible comparisons of reasoning and retrieval, revealing large performance differences across systems.

Teaching Experience

(Fall 2024) TA for CSC110: Foundations of Computer Science @ University of Toronto