Mechanistic interpretability
This line of research uses probing, representation analysis, and activation interventions to study ethical-framework representations in open-weight language models. The competing-target study introduces a test of whether steering selectively controls its intended concept.
Geometric analysis and readout interventions examine how intermediate representations align with a model’s output map, and when interpretability tools can recover predictive information.
LLMs for Social Science
Survey experiments reveal useful LLM signal in aggregate attitudes alongside weak prediction of individual responses. This motivates a shift from silicon samples to silicon aggregates, incorporating LLM forecasts and similarities between survey questions into a Fay–Herriot estimator calibrated with human responses.
Computational cognition
Bayesian models of lighthouse illusions were developed and tested in field and virtual-reality experiments.
A Bayesian ideal-observer model was adapted for active foveated gaze prediction, modelling peripheral vision and selecting fixations to maximise information gain.
A real-time emotion-detection video-conferencing interface was developed and evaluated in a facial-occlusion experiment with physiological sensing.
Computational musicology
Quantitative measures of tonal coherence on the Tonnetz distinguish tonal focus from tonal connection across classical and popular repertoires.
An interpretable acoustic feature space supports unsupervised analysis of EDM categories.
Information retrieval
The undergraduate thesis examines whether gains in MAP and NDCG@10 are reflected in rank-biased precision, a metric based on a model of users’ browsing persistence. Using RankLib, it compares learning curves for AdaRank, ListNet, LambdaMART, and Coordinate Ascent across three LETOR benchmarks.
Teaching assistant at the University of Melbourne, 2021–2023.
Machine Learning, Elements of Data Processing, and Design of Algorithms.
I enjoy composing, arranging, and orchestrating music. Arranging music gives me another way to work with structure: through timbre, instrumentation, and the balance between voices.
