G-Research 2025 PhD prize winners: EPFL

25 September 2025

News
Quantitative research

Every year, G-Research runs a number of different PhD prizes in maths and data science with academic institutions in the UK, Europe and beyond.

Each prize is worth up to £10,000 and is open to final or penultimate year PhD students at specific universities, working across areas including machine learning, quantitative finance and mathematics.

We’re pleased to announce the next PhD prize winners, which ran in conjunction with EPFL – Swiss Federal Technology Institute of Lausanne.

Learn more about our prizes

Alessandro Favero

“As AI models grow in scale and capability, our fundamental understanding of how they work often lags.

“My research aims to bridge this gap. I use tools from statistical physics to develop simple, predictive theoretical models that explain these complex systems.

“Currently, my work focuses on quantifying generalisation and creativity in generative models – especially diffusion.

“In parallel, I work on the science of post-training for large-scale model editing and merging, to make these systems more adaptable and aligned.”

Shaobo Cui

“My research explores how AI systems can reason about causality under uncertainty, a crucial step toward trustworthy and human-aligned AI.

“I developed the BoCUF framework, which models causal reasoning with supporters and defeaters, and introduced CESAR, a token-level metric for quantifying causal strength in uncertain contexts.

“I further proposed DoGE, a geometric embedding method to capture conditional dichotomy between competing causal factors, along with benchmarks and metrics to assess whether large language models maintain causal epistemic consistency.

“These contributions pave the way for reliable, interpretable causal reasoning in next-generation AI systems.”

Youssef Allouah

“My research investigates the foundational principles of trustworthy machine learning by analysing the intrinsic cost of trust – the quantifiable trade-offs between a model’s utility and its guarantees of privacy and robustness.

“My doctoral work began by tackling robustness, where I established fundamental performance limits of collaborative learning under realistic data heterogeneity. This investigation naturally led me to consider privacy, where I uncovered that the costs of these two guarantees are not merely additive but interact to create a fundamental privacy-robustness-utility “trilemma”.

“I then applied these foundational principles of privacy and robustness to the problem of machine unlearning, developing certifiable and resource-efficient methods for data removal.

“Building on this foundation, my future research will focus on extending these trust principles to the increasingly complex large-scale models that define our future, with the goal of making rigorous guarantees practical for real-world AI systems.”