Han Xuanyuan

Anthropic Safety Fellow

Education

University of Cambridge, Churchill College

2018 – 2022

MEng in Computer Science — Distinction (equiv. GPA 4.0)

2021 – 2022

Dissertation: Interpretability of Graph Neural Networks; advised by Pietro Liò.
Highlighted modules: graph representation learning, NLP, reinforcement learning, and probabilistic ML.
Churchill College Scholarship (2022); Hack Cambridge award winner (2022).

BA in Computer Science — First Class Honours (equiv. GPA 4.0)

2018 – 2021

Dissertation: Privacy-Preserving Machine Learning using Homomorphic Encryption.
Highlighted modules: deep learning, machine learning, computer vision, and information theory.
Churchill College Scholarship (2021); Hack Cambridge finalist (2021); Distinguished Talk, Churchill CS Talks (2020).

Gower College and Bishop Gore School, Swansea, Wales

2011 – 2018

A-levels: 4A*s / 98.6% overall with a perfect score in Maths — top student in all subjects. GCSEs: 12A*s.
British Maths Olympiad (Distinction, 2016); Senior Team Maths Challenge (1st in Wales, 2017); Netcraft Award (top-10 Computer Science A-level nationally, 2018).

Anthropic, London — AI Safety Fellow

Aug 2026 – Present

Full-time AI safety research in the Anthropic Fellows Program, based at the London Initiative for Safe AI (LISA).

SPAR (Supervised Program for Alignment Research), Remote — Research Fellow

2026 (part-time)

Developed temporal crosscoders for mechanistic interpretability, tracking how sparse, interpretable features emerge and shift across token positions (mentor: Dmitry Manning-Coe).
Work accepted to the ICML 2026 Mechanistic Interpretability Workshop.

Tower Research Capital, London — Quantitative Researcher

Sep 2025 – Present (on sabbatical)

Research predictive models for high-frequency trading, bringing modern machine-learning methods the desk was newly adopting together with the statistical rigour to keep them robust under noisy, non-stationary data.
Train and evaluate models on GPU/HPC clusters in Python, contributing to the shared research tooling and frameworks used across the desk.

DRW, London — Quantitative Researcher

Aug 2023 – Jul 2025

Researched statistical and machine-learning models for high- and mid-frequency trading: prototyping in Python, building and running backtests, and contributing to shared research infrastructure.

Software Developer

Jul 2022 – Aug 2023

Built automated trading systems: exchange connectivity, pricing, execution logic, and latency optimisation.
Provided on-call support for quants and traders, fixing production-critical issues.

Amazon Lab126, Cambridge — SDE Intern (computer vision and camera hardware)

Summer 2021

Proposed and implemented a new face liveness detection algorithm combining CNNs with shape from shading.
Deployed the solution to mobile devices and demonstrated feasibility in both energy and CPU usage.

Informetis Europe, Cambridge — Machine Learning and Data Science Intern

Summer 2020

Developed an explainable deep learning model for smart-meter time series data based on Siamese networks.

Crosscoding Through Time: Sparse Feature Discovery Across Sequence Positions
D. Manning-Coe, H. Xuanyuan, A. Deshpande, A. Shportko, W. Fei. ICML 2026, Mechanistic Interpretability Workshop.
Superposition in Graph Neural Networks
L. Pertl*, H. Xuanyuan*, P. Liò. NeurIPS 2025, UniReps Workshop; Proceedings of Machine Learning Research (PMLR). *Equal contribution.
Global Concept-Based Interpretability for Graph Neural Networks via Neuron Analysis
H. Xuanyuan, P. Barbiero, D. Georgiev, L. Magister, P. Liò. AAAI 2023 (oral presentation).
Shedding Light on Random Dropping and Oversmoothing
H. Xuanyuan, T. Zhao, D. Luo. NeurIPS 2023, New Frontiers in Graph Learning Workshop.
Efficient Privacy-Preserving Inference for Convolutional Neural Networks
H. Xuanyuan, F. Vargas, S. Cummins. ICLR 2022, Pair2Struct Workshop.

Programming	Python, C++, Java, JavaScript, SQL
ML/DL	PyTorch, Hugging Face, NumPy, pandas, scikit-learn, matplotlib
Infrastructure	Slurm, HTCondor, RunPod, CUDA, Linux, Bash, pytest
Research automation	LLM-as-a-judge, agentic autoresearch
Languages	English (native), Mandarin (working)
Interests	Weightlifting, Cooking