Cohort 3 (2026 Entry)

Research: Exploring Accuracy and Performance Tradeoffs for Compiler Optimisations of AI Model Deployment

Supervisors: Ajitha Rajan, Adriana Sejfia

My research interests lie in reliable and efficient machine learning systems, particularly the testing and evaluation of the software and hardware infrastructure underlying modern ML workloads. I am interested in how we can improve the reliability of ML systems across compilers, runtimes, and heterogeneous hardware, including GPUs and other accelerators. Previously, I worked on systems verification, compiler optimisation, fuzzing, and robust machine learning. Including developing testing approaches for system simulators, energy-efficient software testing, and working on hardware-aware ML optimisation and quantisation infrastructure. I completed my undergraduate and master's studies in Computer Science at King's College London before joining the University of Edinburgh.

Photo: Chris Scott

Research: Understanding vulnerability exploitability towards improving vulnerability detection

Supervisors: Adriana Sejfia, Ajitha Rajan

Fascinated by what happens when good code goes bad and how we can fix it. My research interests are software engineering, software vulnerabilities and software testing. I build automated tools that improve software engineering workflows.

Research: Compressing LLM Reasoning while Preserving Legibility

Supervisors: Shay Cohen, Ivan Titov

Ed's research will explore how LLMs can help formalise maths problems from natural language, and push at the frontier of machine-assisted reasoning. His current project proposes to explore methods that allow translation between compressed or latent reasoning processes and full-form natural language such that reasoning quality, token budget, and legibility are maintained.

Research: Multi-agent simulation and constrained, multi-objective reinforcement learning technologies for install scheduling optimisation

Supervisors: Fengxiang He

Finbar's PhD project aims to develop advanced reinforcement learning algorithms within a multi-agent simulation framework to model real-world conditions, generate high-fidelity reward structures, and optimise scheduling and resource allocation under uncertainty.

Research: AI Assistants for Video Games

Supervisors: Kartic Subr, Amos Storkey

Guolong's project will investigate the applicability of AI-based code synthesis towards developing abstract representations and interactions in a video game environment. It will discover rule-based causal relationships from game roll-outs towards AI assistants that will guide players interactively through an instance of a game.

Research: Efficient and Reliable Solutions to large-scale graph optimisation problems based on Local Sub-graph Joining

Supervisors: Liang Zhao, Tianqi Hou

Hi, I am a robotics guy who enjoys a good robotics challenge. My previous work used reinforcement learning for energy efficient microrobot navigation in dynamic environments, quadruped navigation over uneven 3D terrain, and joint optimisation of robot morphology and control. I also used RL to control a robotic arm that helps a person carry water steadily. My current research focuses on SLAM and graph optimisation for more accurate and reliable mapping. I am always interested in how robots can learn to make better decisions and move reliably in complex environments. I enjoy exchanging ideas with others and exploring new perspectives across robotics and learning model to see how far we can push what robots can do.

Research: AI-Enabled Verification of Semantic Equivalent Between Binaries

Supervisors: Elizabeth Polgreen, Michael O'Boyle

I first came to the University of Edinburgh for my bachelor's in computer science - where I received a Class Medal upon completion. During my undergraduate studies, I grew interests in Natural Language Processing, Programming Languages and eventually some Systems tasks. During my undergraduate thesis, under Elizabeth Polgreen's supervision, I devised a method to obtain post-hoc formal proofs of equivalence for tensor programs. Subsequently, I joined Skyscanner, completing rotations as a frontend engineer and a product manager for the data platform. I'm now interested how we can use formal methods to gain confidence in the functionality of our code. My initial efforts will target equivalence verification between binary programs. In my personal life, I like to spend time cooking and trying out new hobbies and hoping some of them stick - currently cycling and fermentation.

Research: AI Methods to Identify Trends in Simulation and Characterisation of Wafer Chip Design

Supervisors: Ajitha Rajan, Isobel Nicholson, John McPhillimy

Simulation and characterization roles within ST Micro require rapid assimilation of large amounts of previously unexplored data. This must be quickly and succinctly summarized into meaningful models to drive development. Owing to the large size of these datasets (10s to 10,000s of datapoints), it is increasingly challenging to spot outliers; identify trends; compare with previous results, and summarise the full dataset. John's PhD project will address these challenges in the context of simulation and characterisation of ST Micro devices.

Research: Towards Coding Agents that Reliably Observe, Verify and Optimize Compute Systems

Supervisors: Luo Mai, Wenda Li

I want to build the next generation of "operating systems" for AI, and eventually, have AI build them for us. Before coming to Edinburgh, I completed my master's degree at the University of Science and Technology of China, where my research focused on improving the efficiency of machine learning training and inference by orchestrating and co-optimizing compute, communication, and memory resources. During my PhD, I plan to continue improving the efficiency and scalability of ML workloads through systems research. At the same time, I am interested in exploring how AI agents can autonomously and rigorously design and implement efficient, scalable, and reliable systems with minimal human supervision.

Research: LLM-Based AI Companion for Multi-Agent Collaboration

Supervisors: Amos Storkey, Henry Gouk

I'm a PhD student working on machine learning and multi-agent systems. I'm broadly interested in reliable and interpretable machine learning systems and am currently working on methodological developments in reinforcement learning. Outside of my area of study, I enjoy learning about things related to ecology, earth sciences and languages.

Research:

Supervisors: Henry Gouk

Photo: Chris Scott

Research: Multi-Modal Memory models for vision, language and action

Supervisors: Laura Sevilla-Lara, Alessandro Suglia

By the end of my bachelor's degree in computer science and AI at Loughborough University, I developed a strong conviction that studying the human mind can inspire new and efficient algorithms in AI. After a year as a software developer, I attended UCL's MSc in Cognitive and Decision Sciences, where I engaged with many topics such as embodied cognition, language and memory.

It is these three topics that form the basis of my research in the ML Systems CDT. By developing long and short-term memory mechanisms for Vision-Language-Action models, I aim to enable these systems to continually learn and adapt from physical interactions for long time horizons. Currently, I am creating an efficient method to address the stability-plasticity dilemma in neural networks by identifying and sparsely updating the model parameters.

Research: Deep Learning based Solutions to large-scale graph optimisation problems

Supervisors: Liang Zhao, Tianqi Hou

Siqi's current project focuses on developing efficient algorithms and implementations for large-scale graph optimisation problems by using deep learning methods. Such problems frequently arise in domains like SLAM, machine learning, structure-from-motion, and sensor networks, where scalability and computational efficiency are critical. The research aims to design novel linearisation, factorisation, and incremental updating techniques that leverage structured sparsity to reduce computational complexity while maintaining high accuracy.

Photo: Chris Scott

Research: Cool Cores: thermal management for next-generation computing systems through a cross-stack approach

Supervisors: Boris Grot

I completed my undergraduate degree in Electronic Engineering at the University of Southampton. During my time at Southampton, I became interested in VLSI, digital design, and computer architecture, and worked in chip design for two years following my graduation. My research interests lie broadly in computer architecture and in understanding how novel computer systems should be designed for the challenges of current and future workloads, such as machine learning. I am also interested in how advanced process technologies enable new computer architectures which are more performant and energy efficient. In my free time I enjoy spending time outdoors hiking, running, and cycling.

Photo: Chris Scott

Research:Computer architecture, 3D-stacked and memory-centric system designs

Supervisors: Boris Grot

My research focuses on computer architecture, with a particular emphasis on 3D-stacked and memory-centric system designs for machine learning workloads. Currently, I am exploring memory-stacked GPU architectures for large language model (LLM) inference, investigating how emerging packaging technologies and memory hierarchies can address the performance, bandwidth, and energy efficiency challenges of modern AI systems. I am interested in designing next-generation computing platforms that bridge the gap between rapidly growing model demands and hardware capabilities. Before moving to Edinburgh, I earned my bachelor's degrees in Physics and Electrical & Computer Engineering, as well as a master's degree in Electrical Engineering, from the University of Rochester. I also interned at Intel and Ampere Computing, working on system architecture, performance optimization, and hardware-software co-design.