Han Xuanyuan

Xuanyuan (軒轅) is my last name

I'm a researcher working on AI alignment and interpretability. This summer I'm joining Anthropic as an AI safety fellow.

My goal is to develop the tools and methods we need to deploy AI agents safely in the real world. I work toward this through interpretability, taking an empirical approach to understanding what models represent and compute internally. I believe this will help enable their wider adoption in areas such as autonomous scientific discovery.

Before returning to AI research, I spent four years in high frequency trading, most of it as a quantitative researcher at Tower Research and DRW, building statistical models of the financial markets and the production systems that traded on them.

I obtained my Bachelor's and Master's in Computer Science from the University of Cambridge, where I worked with Pietro Liò on graph neural networks, graduating with First Class Honours and Distinction.

I grew up in the Netherlands and Wales. When I'm not working, I'm usually lifting weights or playing video games.

I'm always happy to talk. The best way to reach me is by email.

Email  /  Google Scholar  /  LinkedIn  /  CV

Han Xuanyuan

News

Jul 2026: Started working full time on AI safety.
May 2026: One paper accepted at the ICML 2026 Mechanistic Interpretability Workshop.
Apr 2026: Accepted into the Anthropic Fellows programme.
Mar 2026: One paper on superposition published in PMLR.
Feb 2026: Accepted into SPAR as a research fellow.

Selected Papers

Crosscoding Through Time
Crosscoding Through Time: Sparse Feature Discovery Across Sequence Positions
D. Manning-Coe, Han Xuanyuan, A. Deshpande, A. Shportko, W. Fei
Mechanistic Interpretability Workshop, ICML, 2026
Paper | Code

Uses crosscoders to discover sparse, interpretable features and track how they emerge and shift across token positions in a sequence.

Superposition in Graph Neural Networks
Superposition in Graph Neural Networks
L. Pertl, Han Xuanyuan, P. Liò
Proceedings of Machine Learning Research (PMLR), 2026
Paper | Code

Extends the superposition hypothesis from language models to graphs, studying how GNNs represent more features than they have dimensions.

Global Concept-Based Interpretability for GNNs
Global Concept-Based Interpretability for Graph Neural Networks via Neuron Analysis
Han Xuanyuan, P. Barbiero, D. Georgiev, L. C. Magister, P. Liò
AAAI Conference on Artificial Intelligence (AAAI), 2023
Paper | Code

Analyses individual neurons to extract human-interpretable, global concepts learned by graph neural networks.

Shedding Light on Random Dropping and Oversmoothing
Shedding Light on Random Dropping and Oversmoothing
Han Xuanyuan, T. Zhao, D. Luo
NeurIPS Workshop: New Frontiers in Graph Learning, 2023
Paper

Examines how random-dropping techniques interact with oversmoothing in deep graph neural networks.

Efficient Privacy-Preserving Inference
Efficient Privacy-Preserving Inference for Convolutional Neural Networks
Han Xuanyuan, F. Vargas, S. Cummins
ICLR Workshop on PAIR²Struct, 2022
Paper

Reduces the cost of running convolutional networks under privacy-preserving (encrypted) inference.