Abstract
I'm a PhD student at Imperial College London, supervised by Dr. Matthew Wicker. My research is about formal verification in machine learning models, currently I am particularly working on using formal verification tools in connection with generalization bounds.
- Formal Verification
- Generalization Bounds
- Continual Learning
- Model Robustness
1Introduction
While the empirical success of machine learning models and AI is very evident, the theoretical understanding for empirical patterns is often lacking. The main motivation behind my work is to mathematically understand and certify different machine learning model behaviours. Before my PhD, I studied Computer Science at The University of Manchester (BSc) and Imperial College London (MSc).
I'm always happy to talk about research and possible collaborations, so feel free to get in touch.
2Recent Developments
- Blog post: Bred to Break: What a 1970 Corn Blight Teaches Us About Distilled LLMs.
- “Certification-Enhanced Generalization Bounds” was accepted at NeurIPS 2026.
- Presented “Provably Safe Model Updates” at SaTML 2026.
- “Provably Safe Model Updates” was accepted at SaTML 2026.
- Started my PhD at Imperial College London.
- Finished my MSc in Computing (AI&ML) at Imperial College London.
3Publications
Full list on Google Scholar ↗3.12026
-
[1]NeurIPS '26
Figure 1 of [1]: In various settings our bounds are always-sound, tighter alternatives to existing analytical bounds. @article{elmecker2026certification, title={Certification-Enhanced Generalization Bounds}, author={Elmecker-Plakolm, Leo and Wicker, Matthew}, journal={arXiv preprint arXiv:2610.06238}, year={2026} } -
[2]SaTML '26
Figure 1 of [2]: Illustration of our method on a simple ‘blobs’ dataset, where the model sequentially learns pairs of classes. The true class label for each data point corresponds to the color of the point and the background region to the output of the model. Top: Standard training forgets previously learned classes. Bottom: Training within our locally invariant domain successfully preserves performance on earlier tasks. @inproceedings{elmecker2026provably, title={Provably safe model updates}, author={Elmecker-Plakolm, Leo and Fasterling, Pierre and Sosnin, Philip and Tsay, Calvin and Wicker, Matthew}, booktitle={2026 IEEE Conference on Secure and Trustworthy Machine Learning (SaTML)}, pages={671--688}, year={2026}, organization={IEEE} } -
[3]Preprint
3.22025
-
[4]COLT '25COLT '25 Certifiably Safe Post-Training.
3.3Blog Posts
-
[B1]Oct '26
Lead figure of [B1]: The more the models are distilled from one teacher, the fewer distinct attacks it takes to break them all.
No publications in this category yet.
* denotes equal contribution
ACurriculum Vitae
Download PDFA.1Education
-
Jan 2026 – Dec 2028 (expected)
PhD in Foundational ML/AI
Imperial College London · Supervisor: Dr. Matthew Wicker
-
Sep 2024 – Sep 2025
MSc Computing (AI & ML)
Imperial College London · Distinction
-
Sep 2019 – Jun 2022
BSc (Hons) Computer Science
The University of Manchester · First Class
A.2Experience
-
Aug 2022 – Sep 2024
Software Engineer
Bloomberg L.P. · London
-
Jun 2021 – Aug 2021
Summer Intern, Technology
Barclays UK · Glasgow
-
Jun 2020 – Sep 2020
Full Stack Software Developer
Imago · Manchester
Correspondence
- leo.elmecker-plakolm24@imperial.ac.uk
- Office
- Room 357, ACEX Building
Imperial College London
South Kensington, London SW7 2AZ