I am a researcher at Huawei Noah’s Ark Lab, working in the Reinforcement Learning group on world models, mechanistic interpretability, and continual learning under the guidance of Prof. Jun Wang and Dr. Zafeirios Fountas. My work is broadly about making learning systems more efficient, more adaptive, and better able to reason over long horizons.
I studied Computational Statistics and Machine Learning at UCL. Before that, I completed a BSc in Theoretical Physics, during which I became more interested in computing. After graduating, I spent three years as a software engineer, first in laser technology R&D and then on event-driven trading systems for prediction markets.
Broadly, I want to build data-efficient models that can reason, plan, and adapt beyond their training distribution. My main interests are efficient ML systems, continual learning, mechanistic interpretability, and safe long-horizon AI.
Outside work, I enjoy coding, Formula 1, board games, and art. I am always open to discussing new projects, research ideas, or collaborations; contact me below if you would like to connect.
TLDR: We build an explicit neural global workspace inside an LLM and train it to do mathematical reasoning. We further show that this architecture is scalable through pretraining and that the model increases the synergy of information between its components.
TLDR: We build two Transformer architectures that learn where and when to compute using Bayesian surprise. We route tokens through dynamic layers, which either execute or skip Transformer blocks (Attention + MLP) over time. We show the resulting reductions in self-attention computation and KV-cache requirements, as well as the accuracy trade-offs of conditional computation.
TLDR: We build ASAL++, a method that uses multimodal foundation models to guide open-ended-like search through simulations. ASAL++ continuously proposes new targets stochastically based on a simulation's visual history and explores possible evolutionary strategies.
Contact
Send me a short note, or use one of the email addresses below.
Email addresses: