Haoyan (Ryan) Luo
PhD, Machine Learning, University of Cambridge
hl678 [AT] cam.ac.uk
About
I am a PhD student in Machine Learning at the University of Cambridge, advised by Prof. Mateja Jamnik. I am affiliated with Darwin College and the Department of Computer Science and Technology. Prior to this, I completed my MRes in Machine Learning at Imperial College London, supervised by Prof. Lucia Specia. I obtained my B.Eng. degree in Computer Science from The Chinese University of Hong Kong, Shenzhen.
I have worked as an AI Associate at J.P. Morgan, designing and implementing solutions for real-world financial applications. I previously interned at Microsoft Research Asia, Shanghai AI Lab and Tencent.
📢 2025 update: open to advise student projects on interp research. For Cambridge students, please see this link.
I am interested in how machine learning models work, how to make them more reliable, and how to use them to solve interesting real-world problems:
- Model mechanisms and interpretability problems: studying and manipulating the representations that shape model behavior [MoDE] [SKOP] [Persistent SAE].
- Applied ML and agentic systems: building scalable methods for automated systems [OpenFE] [TIDE] [PageIndex].
Research
Please find all publications on my Google Scholar.
* denotes equal contributions.
Persistent Sparse Autoencoders: Learning Feature-Specific Timescales in Language Model Representations Paper
Haoyan Luo, Mateo Espinosa Zarlenga, Mateja Jamnik
arXiv 2026
Don't Lose Focus: Activation Steering via Key-Orthogonal Projections Paper Code
Haoyan Luo, Mateo Espinosa Zarlenga, Mateja Jamnik
NeurIPS 2026 (Spotlight)
Tuning Language Models by Mixture-of-Depths Ensemble Paper Poster Code
Haoyan Luo, Lucia Specia
COLM 2026 Workshop on Actionable Interpretability
Industry-Aligned Granular Topic Modeling Paper
Sae Young Moon, Myeongjun Erik Jang, Haoyan Luo, Chunyang Xiao, Antonios Georgiadis, Fran Silavong
EMNLP 2025
From Understanding to Utilization: A Survey on Explainability for Large Language Models Paper
Haoyan Luo, Lucia Specia
arXiv 2024
PageIndex Code Chat Platform
VectifyAI
A document indexing and parsing system for building retrieval-ready representations from complex documents, designed to support downstream LLM and agentic workflows.
🎙️ Singer-adaptive Singing Voice Conversion Paper Code Demo
Haoyan Luo
A singing voice conversion project that adapts a source vocal performance to a target singer while preserving musical content and expression.
🗺️ Steering the Networked Temporal Point Processes via Controlling the Network Graph Paper Code
Haoyan Luo, Shuang Li
A reinforcement learning project for controlling networked temporal point processes by intervening on the underlying graph structure.
🛠️ Deep-Learning-Lookup: Deep Learning Algorithms and Utils in Python Code
Haoyan Luo
A personal Python toolkit collecting deep learning algorithms, reusable utilities, and implementation notes for experimentation.
Recent Posts
From Dice to Agents: Reinforcement Learning in One Set of Symbols
September 30, 2026
Most introductions to reinforcement learning for language models start in the middle, with PPO or GRPO already on the table. This post starts from a dice game instead and adds...
Misc.
I was the president of TEDxCUHKSZ during my undergraduate years (check out our annual events), promoting our slogan: "ideas worth spreading" :p
In my free time, I'm a passionate piano and cello enthusiast. I also enjoy playing football, basketball, and tennis. Go Warriors and KTBFFH💙