Posted: 9th July 2025

Mihaela Stoian, a final-year DPhil student in Computer Science, has been awarded the Oxford PhD Runner-up Prize by G-Research. 

The annual prize recognises outstanding DPhil students at the University of Oxford in their final or penultimate year, across fields including (but not limited to) Machine Learning, Computer Science, Quantitative Finance, Mathematics, Statistics, Physics and Engineering. 

Mihaela’s research, supervised by Professor Thomas Lukasiewicz, focuses on neuro-symbolic methods – an area of artificial intelligence that combines the strengths of neural networks with symbolic reasoning. Her work aims to improve the reliability of AI systems by making sure they respect background knowledge and real-world rules, both while learning and when making predictions. 

She is particularly interested in deep generative models for synthesising realistic tabular data. These models are often used to generate outputs that look like the real data – such as medical records – but without compromising privacy. However, standard models can sometimes produce unrealistic or contradictory outputs, such as a record where a patient’s minimum haemoglobin level is higher than their maximum. Standard models discard these outputs and generate new samples, which can result in potentially indefinite inference times – wasting time and resources. 

To address this, Mihaela developed a new framework that allows deep generative models to learn from and follow complex constraints – including logical rules and inequalities – from the start. This makes the synthetic data far more accurate and useful, especially in sensitive areas like healthcare. Her method is the first to guarantee that every generated sample fully meets the required constraints, with little or no extra time needed to produce results. Crucially, it works with any deep generative model architecture. 

Her research also tackles the problem of scaling neuro-symbolic methods to more complex settings, such as autonomous driving. In these situations, AI models must avoid producing contradictory or dangerous outputs. Mihaela has developed a memory-efficient technique that allows logic-based rules to be used in such settings, improving the model’s performance even when there is limited annotated data. 

Her wider research vision is to bring neuro-symbolic AI closer to real-world deployment, helping to create AI systems that are more trustworthy, safe and robust. 

Further details of the 2025 prize winners are available via G-Research: 
https://www.gresearch.com/news/g-research-2025-phd-prize-winners-university-of-oxford/