A little spiel

I studied Natural Sciences at the University of Cambridge from 2023 to 2026, obtaining a double First, where I placed second in my cohort and received the Holgate-Pollard and J.M. Thoday prizes. During this time, I did research in mathematical and computational genetics at the University and the Cancer Research UK Cambridge Institute, and neuroscience at Caltech.

Science competitions ignited my desire to work hard and do interesting research. Meeting like-minded competitors at camps and competitions was a fantastic experience, and I wholeheartedly recommend trying them out - especially across many disciplines!

Publications

Leung, C.H.C., Houldcroft, C. J. & Scally, A. Statistical inference of virus evolution using ancestral recombination graphs. 2026.10.02.756209 Preprint at https://doi.org/10.64898/2026.10.02.756209 (2026).

Graepel, G. & Leung, C.H.C. Something is rotten in the state of knowledge. (2026). PDF

Research

Spontaneous neuronal activation · 2026–present

Lois group, Caltech

Computational modelling of spontaneous neuronal activity and its role in the restoration of learned behaviours following perturbation. Drawing on attractor network theory and Hebbian plasticity. Also worked on preprocessing of fMRI data (ongoing project). Part of the Cambridge–Caltech exchange, funded by Caltech and St Catharine's College, Cambridge.

Viral ancestral recombination graphs · 2025–present

Department of Genetics, University of Cambridge

Scalable inference of ancestral recombination graphs for viral DNA, using Markov Chain Monte Carlo and perturbation theory. Awarded the J.M. Thoday Prize for the best undergraduate research project.

code

Deep learning for genome organisation · 2025–present

Hannon Group, Cancer Research UK Cambridge Institute

Deep learning for predicting three-dimensional chromatin organisation from sequencing data, in the laboratory of Greg Hannon. Work focused on learning sequence determinants of topologically associating domains and compartment structure. Publication in preparation.

code

Lock-in risk benchmark for LLM systems · 2025

Supervised Program for Alignment Research

Paper under review.

Structural biology of transcription · 2024

Department Cramer, Max Planck Institute for Multidisciplinary Sciences, Göttingen

Studied +1 nucleosome promoter-proximal pausing in eukaryotic transcription. Independently designed biochemical assays, and developed skills in cryo-electron microscopy. Funded by the Max Planck Institute for Multidisciplinary Sciences.

Hidden Markov Models for fast protein kinetics · 2023

Haran Group, Weizmann Institute of Science

Inference of conformational state sequences from single-molecule FRET data using hidden Markov models. Work addressed the challenge of fast dynamics on timescales shorter than photon-arrival intervals. Funded by the Weizmann UK Foundation.

report

Projects

New York Times Pips solver

Belief propagation, message passing, and factor graphs applied to solving New York Times Pips. This is obviously not as efficient as a simple constraint/backtracking solver, but it is a statistical analogue of how we solve puzzles collaboratively.

Try it here!

Action chunking with transformers

Teaching a network to control a robotic arm with action chunking, to perform the task of grasping a champagne flute, flipping it upside down, and manoeuvring it to a new position.

code

World models with JAX

Implementation of LeJEPA architecture, trained using JAX on GPUs.

code

todo

Built in Rust: control your todo list. Used to learn Rust.

Try here!

Cross-lingual interpretability

Oxford AI Safety Initiative, 2026

Using LoRA finetuning to understand generalisation across languages in open-source LLMs.

writeup and code

Multi-agent AlphaZero

Implementation of the multi-agent AlphaZero architecture, trained using JAX on GPUs.

code