--- headshot-css: true page-title: Matthew Farrugia-Roberts title: "Welcome! ![](matt-2022.jpg){.headshot style=\"float: right\"}" title-left: true --- to the personal website of: **Matthew Farrugia-Roberts**\ Doctoral student\ Department of Computer Science & Magdalen College\ [University of Oxford](https://www.cs.ox.ac.uk/) 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 ``{=html} or give [anonymous feedback](https://whispermeter.com/feedback/BWHC8S879S). This page includes my [bio](#about-me), [publications](#publications), [teaching](#teaching), [software](#publications), and [affiliations](#affiliations). ::: block Recent announcements (see [news page](news) for more): - I published a workshop paper on how representations develop over long-run training time: ["Structure and scale in simplicial sequence modelling"](https://openreview.net/forum?id=aFlGLdTzuq). - Together with my students Afiq Abdillah Effiezal Aswadi, [Olly Britton](https://ollybritton.com/), and Ross Baker, I published a paper on memorisation versus generalisation under non-stationary training data: ["Temporal task diversity: Inductive biases under non-stationarity in synthetic sequence modelling"](https://arxiv.org/abs/2605.18281). - I rebooted [Hi, JAX! An introduction to vanilla JAX for deep learning research](https://github.com/matomatical/hijax). There are currently 11 lectures (\~13hrs) covering the fundamentals of deep learning with JAX, with a few more lectures on the way soon. Consider [subscribing](https://www.youtube.com/@MatthewFarrugiaRoberts)! Recent writing (see [writing page](far-be-it-from-me) for more): - I wrote a [technical note](epiplexity) on the theoretical concept of epiplexity from the paper [Finzi et al. (2026)](https://arxiv.org/abs/2601.03220), "From Entropy to Epiplexity: Rethinking Information for Computationally Bounded Intelligence." - I wrote an essay on [why and how I try to read all of the titles posted to arXiv in CS/ML categories each day](subscribing-to-arxiv). - I read *Replacing Guilt* by Nate Soares. I took notes and wrote a [detailed summary](replacing-guilt), in addition to my regular reflection on this year's [reading log](readings-2026#replacing-guilt-by-nate-soares). - People sometimes ask me why I love JAX so much. So, I [wrote it down!](jax) ::: ## 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](https://www.cs.ox.ac.uk/people/alessandro.abate/). I collaborate on applying singular learning theory to understand deep reinforcement learning with [Timaeus](https://timaeus.co/). I also help run the [Oxford AI Safety Initiative](https://oaisi.org/). 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](https://www.karimabdelsadek.com/), [Hannah Erlebach](https://hannaherlebach.com/), [Usman Anwar](https://uzman-anwar.github.io/), and [Michael Dennis](https://www.michaeldennis.ai/) at [Krueger AI Safety Lab](https://kasl.ai) / [CBL](https://cbl.eng.cam.ac.uk/) at the University of Cambridge. I also worked on establishing AI safety applications of singular learning theory with [Timaeus](https://timaeus.co/). Before that, I completed a [master's thesis](mthesis) on lossless compression of neural networks supervised by [Daniel Murfet](http://therisingsea.org/), and interned at [CHAI](https://humancompatible.ai) researching the foundations of reward learning with [Adam Gleave](https://www.gleave.me/) and [Joar Skalse](https://www.alignmentforum.org/users/logical_lunatic). I also helped run a virtual AI safety [reading group](https://metauni.org/ai-safety/) at [metauni](https://metauni.org/). 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](https://openreview.net/forum?id=okCqx1RdIw) ([poster](abdelsadek2025-poster.pdf), 3.2MB) to appear at **RLC 2025**. [Preprint](https://arxiv.org/abs/2507.03068) on arXiv. Tweet threads on [results](https://x.com/Karim_abdelll/status/1942627952094847121) and [motivation](https://x.com/MatthewFdashR/status/1942682523735318621). - Joar Skalse(=), **MFR(=),** Alessandro Abate, Stuart Russell, and Adam Gleave, 2023, "Invariance in policy optimisation and partial identifiability in reward learning". [Conference paper](https://icml.cc/virtual/2023/poster/24879) ([poster](skalse2023invariance-poster.pdf)) presented at **ICML 2023.** [Preprint](https://arxiv.org/abs/2203.07475) on arXiv. Science of deep learning, singular learning theory, developmental interpretability: - **MFR,** "Structure and scale in simplicial sequence modelling." [Workshop paper](https://openreview.net/forum?id=aFlGLdTzuq) to appear at HiLD: 4th Workshop on High-dimensional Learning Dynamics, a **workshop at ICML 2026.** [Preprint](https://arxiv.org/abs/2606.01302) on arXiv. - Afiq Abdillah Effiezal Aswadi(=), Oliver Britton(=), Ross Baker(=), **MFR,** "Temporal task diversity: Inductive biases under non-stationarity in synthetic sequence modelling." [Workshop paper](https://tais2026.cc/proceedings/farrugiaroberts-temporal-diversity) ([poster](EffiezalAswadi+2026poster.pdf)) presented at **TAIS 2026.** To appear at Foundations of Deep Generative Models: Understanding Memorization, Generalization, and Reasoning, a **workshop at ICML 2026.** [Preprint](https://arxiv.org/abs/2605.18281) on arXiv. [Code](https://github.com/matomatical/temporal-task-diversity) 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](https://arxiv.org/abs/2601.07524) 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](https://arxiv.org/abs/2502.05475) 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](https://arxiv.org/abs/2501.17745) 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](https://openreview.net/forum?id=45qJyBG8Oj), **TMLR.** [Preprint](https://arxiv.org/abs/2402.02364) on arXiv. ``{=html} - George Wang(=), **MFR(=),** Jesse Hoogland, Liam Carroll, Susan Wei, and Daniel Murfet, 2024, "Loss landscape geometry reveals stagewise development of transformers." [Workshop paper](https://openreview.net/forum?id=2JabyZjM5H) ([poster](wang2024geometry-poster.pdf), 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](https://arxiv.org/abs/2306.02834) on arXiv. ``{=html} - **MFR,** 2024, "Losslessly compressible neural network parameters". [Workshop paper](https://openreview.net/forum?id=VhhsbII0Lk) 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](https://neurips.cc/virtual/2023/poster/71011) ([poster](farrugia2023paths-poster.pdf), 3.9MB) presented at **NeurIPS 2023**. [Preprint](https://arxiv.org/abs/2305.05089) on arXiv. - **MFR,** 2022, *Structural Degeneracy in Neural Networks,* **Master's thesis,** School of Computing and Information Systems, the University of Melbourne. [Available online](mthesis). Computer science education: - **MFR,** Bryn Jeffries, and Harald Søndergaard, 2022, "Teaching simple constructive proofs with Haskell programs". Conference paper: [extended abstract](https://wiki.tfpie.science.ru.nl/TFPIE2022) presented at **TFPIE 2022**, [full paper](http://doi.org/10.4204/EPTCS.363.4) published in **EPTCS**. - **MFR,** Bryn Jeffries, and Harald Søndergaard, 2022, "Programming to learn: Logic and computation from a programming perspective". [Conference paper](https://doi.org/10.1145/3502718.3524814) presented at **ACM ITiCSE 2022**. See also my [Google Scholar profile](https://scholar.google.com/citations?hl=en&user=hIN4ac4AAAAJ&view_op=list_works&sortby=pubdate). ## Teaching Here are some select teaching projects. See my [teaching page](teaching) for a full list. Iliad intensive: - Author, teaching assistant (April, June 2026), [Singular Learning Theory](https://docs.google.com/document/d/1g4Hs0K6uVxn44-CCQzclmhU4Z6EuAz8bpraSsc77l5s/edit?tab=t.8q8aoxfgo9bw) ([lecture notes](https://drive.google.com/file/d/1QK8oiqr9zITD7qpBdWqE01bJLf2MYrJt/view)). Independent: - Teacher (2024--2026), "Hi, JAX!" An introduction to vanilla JAX for deep learning research ([2024 pilot](hijax-2024), [2026 online edition](https://github.com/matomatical/hijax)). University of Oxford: - Head TA (2025), [AI Safety and Alignment](https://robots.ox.ac.uk/~fazl/aisaa/) ([alignment fundamentals colab tutorial](https://colab.research.google.com/drive/1rAGN4Ngn2Rk1sJekfmnei1gTgwlEFqYM?usp=sharing)). - TA (2025), Foundations of Self-Programming Agents. The University of Melbourne: - Guest lecturer (2024), COMP90087 The Ethics of Artificial Intelligence ([recording](future-ethics-lecture)). - 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](https://github.com/matomatical/AI-games)). - Head TA (2017--2021), COMP30026 Models of Computation ([paper 1](https://doi.org/10.1145/3502718.3524814), [paper 2](http://doi.org/10.4204/EPTCS.363.4)). ## Software I'm technically half software engineer, half computer scientist by training. Here are some select open-source software projects. Research and operations: - [jaxgmg](https://github.com/matomatical/jaxgmg), a WIP library for JAX-accelerated Goal MisGeneralisation research. - [matthewplotlib](https://github.com/matomatical/matthewplotlib), a Python terminal plotting library that aspires to *not be painful.* - [wunderbar](https://github.com/matomatical/wunderbar), a robust Python parser for W&B's `.wandb` binary structured log format. - [strux](https://github.com/matomatical/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](https://all-you-need.far.in.net/). - [memograph](https://github.com/matomatical/memograph), a command-line flashcard app implementing [ebisu](https://github.com/fasiha/ebisu)'s Bayesian spaced repetition algorithm. - [Several implementations](https://github.com/matomatical/snake) of a Snake game in Rust and nim. See my [GitHub profile](https://github.com/matomatical/) for a full list. ## Affiliations Current: - Doctoral student at the [Department of Computer Science](https://www.cs.ox.ac.uk/people/matthew.farrugia-roberts/), University of Oxford. - Executive committee member, [Oxford AI Safety Initiative](https://oaisi.org/committee/). - Member of [Magdalen College](https://en.wikipedia.org/wiki/Magdalen_College,_Oxford), University of Oxford. - AI researcher, [AI Existential Safety Community](https://futureoflife.org/about-us/our-people/ai-existential-safety-community/), Future of Life Institute. - Research affiliate at [Timaeus](https://timaeus.co). Past: - Research associate at [Timaeus](https://timaeus.co). - Research assistant (AI alignment & reward hacking) at [Krueger AI Safety Lab](https://kasl.ai) and the [Computational and Biological Learning Lab](https://www.cbl-cambridge.org/), University of Cambridge. - Research assistant (human--agent interaction) at the [School of Computing and Information Systems](https://cis.unimelb.edu.au/), the University of Melbourne. - Research intern at the [Centre for Human-compatible AI](https://humancompatible.ai/), University of California, Berkeley. - Casual tutor and lecturer at the [School of Computing and Information Systems](https://cis.unimelb.edu.au/), 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.