Computer Science · University of Warwick

Lucas Tompson

I have just finished my second year studying computer science at the University of Warwick. My interests involve the maths behind machine learning, in order to help create better models and understand why they work.

Lucas Tompson
Education BSc Computer Science · University of Warwick · 2024–2027
Year 175.8% · First↗

The foundations: programming, maths, computer architecture and security.

8 modules · view year
Year 2First predicted↗

The theory-heavy year. Algorithms, AI, databases and software engineering.

9 modules · view year
Year 3Final year↗

Where I specialise — machine learning modules, plus the thesis.

7 modules · view year
Toolkit

Mostly Python and PyTorch these days — that's what the research runs on. Comfortable enough in Java, C, SQL and JavaScript to be useful, and I reach for Git, Linux and Jupyter without thinking about it. The rest I tend to pick up when a project needs it.

Python PyTorch Java C SQL JavaScript Flask scikit-learn Git Linux Jupyter
Experience

Previous Employment and Education

  1. June - September 2026

    Software Engineer · Barclays Investment Bank

    Canary Wharf, London, UK

    Working within the Fixed Income team.

  2. September 2023 – September 2024

    Network Installations Assistant · IT Workhouse

    Kingsbridge, UK

    Designed and fitted bespoke network systems for clients end to end — scoping requirements, configuring hardware, and providing ongoing support.

  3. Various between 2019 - 2025

    Kitchen Supervisor · Café Alf Resco

    Dartmouth, UK

    Several years at a busy Dartmouth café, working up to the supervisor team — running the kitchen under pressure, owning the rota and staff leave, and helping the accounts team with payroll and daily sales.

  4. 2016 – 2023

    Churston Ferrers Grammar School

    Churston, UK
    Mathematics A* Further Maths A* Computer Science A Physics A

    Finished with an EPQ graded A* — a self-directed research dissertation, “To what extent has the internet affected political discourse?” — my first taste of framing a question and writing it up properly.

Featured · ongoing research

Post-training of Large Language Diffusion Models

My dissertation: continual learning in diffusion LLMs. Most models read one token at a time, left to right, but masked diffusion reveals them in a random order, training on an ensemble of orderings at once.

I'm looking into how that affects catastrophic forgetting. Supervised by Dr Paris Giampouras.

Read the write-up →
Autoregressive Masked diffusion
forgetting 𝓕ₐ → new-task adaptation budget ε →
Fig. 1 — conjectured stability–plasticity frontier. Lower is better.
Writing

Personal Projects and Research

Some write-ups about projects I've done recently. You can find most of them on GitHub.

Grokking detection with variational inference

Watching a transformer for the moment generalisation suddenly happens, I estimate the loss landscape's local learning coefficient to flag the phase-transition early, across many seeds.

PyTorch · devinterpRead →

A genetic algorithm for the travelling salesman

Building my own evolutionary solver for an NP-hard problem — tournament selection, crossover and mutation to dodge local minima

Python · MatplotlibRead →

Shipping a full-stack artwork platform

Accounts, hashed auth and a content-moderation pipeline over a database. Synchronous and asynchronous interaction so the gallery felt live.

Flask · SQLAlchemyRead →