Faculty of Engineering and Physical Sciences EPSRC Project Proposals 2025/26 - Robust Large Language Models for Intelligent Software Development

Eligibility: UK Only

Funding: EPSRC Doctoral Landscape CASE Competition Award in collaboration with TurinTech AI, providing full academic fees, together with a tax-free maintenance grant at the standard UKRI rate of £20,780 per year and an additional top-up of £4,000 per year for 3.5 years.

Lead Supervisor’s full name & email address

Professor Zheng Wang: z.wang5@leeds.ac.uk

Co-supervisor’s full name & email address

Dr Chunwei Xia: c.xia@leeds.ac.uk

Project summary

Large Language Models (LLMs) are revolutionising software development by automating code generation and optimisation. However, applying LLMs to software development faces one glaring problem: correctness. Asking LLMs to generate the correct code remains a matter of luck. This project aims to make LLMs reliable for software engineering, enabling them to produce accurate and correct code.

This project will develop techniques to help software engineers complete previously costly and challenging tasks in real-life settings. If successful, this project will lead to fundamental breakthroughs in ML-based code reasoning.

Large language models (LLMs) hold immense potential in supporting software engineering tasks like code translation and optimisation, many of which currently require extensive human involvement and are expensive. Automating these tasks can thus offer substantial cost savings. However, applying LLMs to code generation faces one glaring problem: correctness. Asking LLMs to produce correct code remains a matter of luck - they are often wrong than right in many code-related tasks.

Our vision is to make LLMs practical and reliable for code generation. To this end, we will develop new learning algorithms and machine learning (ML) model architectures to extract information from structured data, such as program data and dependence graphs. This will enable ML to take advantage of the structured syntax and semantics of programming languages to reason about data flows and dependencies essential for code generation. We will find ways to scale LLMs and formal methods so that they can handle large and complex programs in real-life settings.

If successful, this project will lead to fundamental breakthroughs in ML-based code reasoning. Working with our industry partners (TurinTech AI), we will demonstrate how our techniques can assist in code generation and optimisation tasks in real-world industry settings, helping software engineers complete previously costly and challenging software engineering tasks.

Please state your entry requirements plus any necessary or desired background

A first class or an upper second class British Bachelors Honours degree (or equivalent) in an appropriate discipline.

Advert information