PhD Studentship: Reliable and Generalisable Knowledge Editing in Large Language Models

UAG; Norwich; Vor Ort; GBP 20.000 - 24.000; Vollzeit

Zusammenfassung

University of East Anglia invites applications for a PhD project on reliable, localised knowledge updating in large language models. You will explore editing methods like ROME and MEMIT, compare with RAG and retrieval-augmented memory, and test on factual benchmarks.

The project is funded for three years, with a standard UK stipend of £21,805 annually, available to UK applicants, and begins on 1 October 2027. Full-time study in Norwich is expected.

Qualifikationen

• Minimum entry requirement is a 2:1 in Computer Science.

Ausbildung

2:1 in Computer Science

Jobbeschreibung

About the Project

Primary Supervisor - Dr Farhana Ferdousi Liza

Emerging techniques in model editing, also discussed in the literature as knowledge editing and memory editing, offer promising approaches for correcting erroneous or harmful information, incorporating specialised knowledge, and improving the usability, security, and accountability of large language models (LLMs). The overall aim of this PhD project is to develop and evaluate methods for reliable, localised, and generalisable knowledge updating in large language models. Using accessible open-weight language models suited to available computational resources, including NVIDIA RTX 6000 Ada 48GB GPUs and UEA high-performance computing (HPC) facilities, you will adapt established editing methods, such as ROME and MEMIT, and compare them with retrieval-augmented generation (RAG), external memory mechanisms, and conventional fine-tuning. Experiments will investigate how models can incorporate new information or knowledge, correct inaccurate information, and remove outdated or unwanted information while minimising unintended changes to existing capabilities. Evaluation will use established factual model-editing benchmarks, such as CounterFact and zsRE, supplemented where appropriate by a domain-specific dataset. Performance will be assessed in terms of editing efficacy (whether the target knowledge is successfully updated), generalisation (whether updates remain effective across paraphrases and related queries), locality (whether unrelated knowledge and capabilities are preserved), and retention (whether updates persist following subsequent edits or conflicting information). Further evaluations will examine factual consistency, general reasoning, and robustness to adversarial prompts. Where feasible, the project will also investigate knowledge provenance and update traceability to support accountability. Comparative experiments will establish the strengths, limitations, and trade-offs of different knowledge-updating approaches. You will also have the opportunity to deploy selected models in embodied AI systems or social robots, such as Furhat, and conduct human-AI interaction experiments to investigate how knowledge updates affect responses, consistency, and behaviour during real-time interactions. The project will contribute practical methods and empirical evidence for developing reliable, trustworthy, and accountable LLMs for knowledge-intensive applications.

Entry Requirements

The standard minimum entry requirement is 2:1 in Computer Science.

Mode of Study

Full-time

Start Date

1 October 2027

Additional Funding Information

This PhD project is in a competition for a Faculty of Science funded studentship. Funding is available to UK applicants and comprises ‘home’ tuition fees and an annual stipend for 3 years. The current UEA / institutional PhD stipend rate is £21,805 and will be subject to any agreed inflationary uplift for 2027/28.