PhD studentship in Efficient LLM Inference

One fully funded, full-time PhD position to work with Edoardo Ponti in the Institute for Language, Cognition and Computation at the School of Informatics, University of Edinburgh and Euan Wielewski, Head of Applied AI, NatWest Group.

This project aims to increase the efficiency of Large Language Models (LLMs) and multimodal foundation models, focusing on adaptive architectures, and broaden their “horizon” to longer contexts and reasoning chains.

The use of Large Language Models has become pervasive in AI applications. However, their deployment remains slow, costly, energy-demanding, and high-carbon. This project seeks to design new model architectures capable of accelerating LLM deployment by achieving conditional computation (where the amount of compute dedicated to a task is proportional to its complexity and achieves a better efficiency-performance Pareto frontier compared to dense architectures). The project will focus on adaptive tokenization, including multi-token prediction, and dynamic pooling of representations. The goal is to increase LLM throughput while enhancing their capabilities to model long-context sequences and generate long chains of thought for reasoning.

Candidate’s profile 

• A good Bachelors degree (2.1 or above or international equivalent) and/or Masters degree in a relevant subject (physics, mathematics, engineering, computer science, or related subject)
• Proficiency in English (both oral and written)
• Strong foundation in natural language processing and deep learning, with a particular emphasis on large language models, transformer architectures, and their training methodologies
• Proficiency in Python and PyTorch, with experience in developing and training deep learning models, and familiarity with distributed computing environments
• Hardware knowledge of current accelerators (e.g. GPUs) is highly desirable

Studentship and eligibility

The funded studentship starting in the academic year 2024/25 covers:

• Full time PhD tuition fees for a student with a Home fee status (£4,786 per annum)
• A tax-free stipend of GBP £19,237 per year for 3.5 years
• Additional programme costs of £1000 per year

Application Information

Applicants should apply via the University’s admissions portal (EUCLID) and apply for the following programme: https://postgraduate.degrees.ed.ac.uk/?r=site/view&edition=2024&id=491 with a start date of 01-05-25.

Applicants should state “Efficient LLM Inference” and the research supervisor (Edoardo Ponti) in their application and Research Proposal document.

Complete applications submitted by 30 April 2025 will receive full consideration; after that date applications will be considered until the position is filled. The anticipated start date is 01-05-25.

Applicants must submit:

• All degree transcripts and certificates (and certified translations if applicable)
• Evidence of English Language capability (where applicable)
• A short research proposal (max 2 pages)
• A full CV and cover letter describing your background, suitability for the PhD, and research interests (max 2 pages)
• Two references (note that it the applicant’s responsibility to ensure reference letters are received before the deadline)

Only complete applications (i.e., those that are not missing the above documentation) will progress forward to Academic Selectors for further consideration.

To be onboarded to NatWest Group’s cloud computing environment, the successful candidate will need to undergo background checks.

Environment

The School of Informatics is one of the largest in Europe and currently the top Informatics institute in the UK for research power, with 40% of its research outputs considered world-leading, and almost 50% considered top grade for societal impact.

NatWest Group is one of the UK’s largest banking and financial services institutions, serving over 19 million customers across multiple brands.

Type / Role:

Subject Area(s):

Location(s):