to the personal website of:

Matthew Farrugia-Roberts
Doctoral student
Department of Computer Science & Magdalen College
University of Oxford

My research aim is to understand the foundations of intelligence, learning, and computation, and to use this understanding to anticipate risks to humanity from future advanced intelligent systems.

Contact: Email matthew@far.in.net, schedule a 1-on-1, or give anonymous feedback.

This page includes my bio, publications, teaching, software, and affiliations.

§About me

I’m currently a PhD student at the University of Oxford, studying the science of deep learning and agent foundations under the supervision of Professor Alessandro Abate. I collaborate on applying singular learning theory to understand deep reinforcement learning with Timaeus. I also help run the Oxford AI Safety Initiative.

My broad research goal is to understand and anticipate catastrophic and existential risks from future advanced AI systems. I previously studied goal misgeneralisation with Karim Abdel Sadek, Hannah Erlebach, Usman Anwar, and Michael Dennis at Krueger AI Safety Lab / CBL at the University of Cambridge. I also worked on establishing applications of singular learning theory to AI safety with Timaeus. Before that, I completed a master’s thesis on lossless compression of neural networks supervised by Daniel Murfet, and interned at CHAI researching the foundations of reward learning with Adam Gleave and Joar Skalse. I also helped run a virtual AI safety reading group at metauni.

My background is in statistical machine learning, theoretical computer science, and software engineering. I studied computer science and machine learning at the University of Melbourne and ETH Zürich. I also worked for several years as a tutor and lecturer at the University of Melbourne, teaching mostly theoretical computer science, algorithms and data structures, and classical AI.

§Publications

Reward ambiguity and generalization in reinforcement learning:

• Karim Abdel Sadek(=), MFR(=), Usman Anwar, Hannah Erlebach, Christian Schroeder de Witt, David Krueger, and Michael Dennis, 2025, “Mitigating goal misgeneralization via minimax regret”. Conference paper (poster, 3.2MB) to appear at RLC 2025. Preprint on arXiv. Tweet threads on results and motivation.

• Joar Skalse(=), MFR(=), Alessandro Abate, Stuart Russell, and Adam Gleave, 2023, “Invariance in policy optimisation and partial identifiability in reward learning”. Conference paper (poster) presented at ICML 2023. Preprint on arXiv.

Science of deep learning, singular learning theory, developmental interpretability:

• MFR, “Structure and scale in simplicial sequence modelling.” Workshop paper under review. Preprint in preparation.

• Afiq Abdillah Effiezal Aswadi(=), Oliver Britton(=), Ross Baker(=), MFR, “Temporal task diversity: Inductive biases under non-stationarity in synthetic sequence modelling.” Workshop paper (poster) presented at TAIS 2026. Preprint on arXiv. Code on GitHub.

• Chris Elliott(=), Einar Urdshals(=), David Quarel(=), MFR, and Daniel Murfet, 2026, “Stagewise reinforcement learning and the geometry of the regret landscape.” Journal paper under review. Preprint on arXiv.

• Simon Pepin Lehalleur(=), Jesse Hoogland(=), MFR(=), Susan Wei, Alexander Gietelink Oldenziel, George Wang, Liam Carroll, and Daniel Murfet, 2025, “You are what you eat: AI alignment requires understanding how data shapes structure and generalisation.” Position paper under review. Preprint on arXiv.

• Liam Carroll, Jesse Hoogland, MFR, and Daniel Murfet, 2025, “Dynamics of transient structure in in-context linear regression transformers.” Conference paper under review. Preprint on arXiv.

• Jesse Hoogland(=), George Wang(=), MFR, Liam Carroll, Susan Wei, and Daniel Murfet, 2025, “Loss landscape degeneracy and stagewise development in transformers.” Journal paper, TMLR. Preprint on arXiv.

• George Wang(=), MFR(=), Jesse Hoogland, Liam Carroll, Susan Wei, and Daniel Murfet, 2024, “Loss landscape geometry reveals stagewise development of transformers.” Workshop paper (poster, 7.3MB) presented at HiLD: 2nd Workshop on High-dimensional Learning Dynamics, a workshop at ICML 2024. Best papers of HiLD award.

Neural network geometry:

• MFR, 2024, “Proximity to losslessly compressible parameters”. Conference paper under review. Preprint on arXiv.

• MFR, 2024, “Losslessly compressible neural network parameters”. Workshop paper presented at Machine Learning and Compression Workshop, a workshop at NeurIPS 2024.

• MFR, 2023, “Functional equivalence and path connectivity of reducible hyperbolic tangent networks”. Conference paper (poster, 3.9MB) presented at NeurIPS 2023. Preprint on arXiv.

• MFR, 2022, Structural Degeneracy in Neural Networks, Master’s thesis, School of Computing and Information Systems, the University of Melbourne. Available online.

Computer science education:

• MFR, Bryn Jeffries, and Harald Søndergaard, 2022, “Teaching simple constructive proofs with Haskell programs”. Conference paper: extended abstract presented at TFPIE 2022, full paper published in EPTCS.

• MFR, Bryn Jeffries, and Harald Søndergaard, 2022, “Programming to learn: Logic and computation from a programming perspective”. Conference paper presented at ACM ITiCSE 2022.

See also my Google Scholar profile.

§Teaching

Here are some select teaching projects. See my teaching page for a full list.

University of Oxford:

• Head TA (2025), AI Safety and Alignment (alignment fundamentals colab tutorial).
• TA (2025), Foundations of Self-Programming Agents.

Independent:

• Teacher (2024–2025), “Hi, JAX!” (webpage 2024, online).

The University of Melbourne:

• Guest lecturer (2024), COMP90087 The Ethics of Artificial Intelligence (recording).
• Lecturer (2019), COMP90059 Introduction to Python Programming.
• Head TA (2018) and coordinator (2019), COMP20007 Design of Algorithms.
• Head TA (2017–2021), COMP30024 Artificial Intelligence (projects repository).
• Head TA (2017–2021), COMP30026 Models of Computation (paper 1, paper 2).

§Software

I’m technically half software engineer, half computer scientist by training. Here are some select open-source software projects.

Research and operations:

• jaxgmg, a WIP library for JAX-accelerated Goal MisGeneralisation research.
• matthewplotlib, a Python terminal plotting library that aspires to not be painful.
• wunderbar, a robust Python parser for W&B’s .wandb binary structured log format.
• strux, a minimal JAX/dataclass wrapper.

Other fun stuff:

• How many days has it been since someone posted a paper to arXiv with “all you need” in the title? Find out here.
• memograph, a command-line flashcard app implementing ebisu’s Bayesian spaced repetition algorithm.
• Several implementations of a Snake game in Rust and nim.

See my GitHub profile for a full list.

§Affiliations

Current:

• Doctoral student at the Department of Computer Science, University of Oxford.
• Executive committee member, Oxford AI Safety Initiative.
• Member of Magdalen College, University of Oxford.
• AI researcher, AI Existential Safety Community, Future of Life Institute.
• Research affiliate at Timaeus.

Past:

• Research associate at Timaeus.
• Research assistant (AI alignment & reward hacking) at Krueger AI Safety Lab and the Computational and Biological Learning Lab, University of Cambridge.
• Research assistant (human–agent interaction) at the School of Computing and Information Systems, the University of Melbourne.
• Research intern at the Centre for Human-compatible AI, University of California, Berkeley.
• Casual tutor and lecturer at the School of Computing and Information Systems, the University of Melbourne.

Any views expressed on this website are not intended to represent the views of any of my current or previous affiliated institutions.