Summer research in School of Computer Science

The School of Computer Science is breaking new ground in the theory and practice of computational systems and their applications, it is a progressive, inclusive department, providing specialist teaching and conducting world-leading research in fundamental and applied computer science.

Summer research projects available for summer 2026

Project 1

Evaluating the shift in LLM thoughts in multi-agent systems.

Research lead - Dr Abhirup Ghosh

Research objective

This project aims to evaluate how the internal representations of Large Language Models (LLMs) evolve as multiple agents interact to achieve a task through collaboration. By examining how token embeddings and their relationships shift during multi-agent interaction, we will investigate when and why a collaboration gap arises in multi-agent systems. We will evaluate scenarios in which agents, configured using system prompts, differ in their initial beliefs and must converge to produce a correct final answer. Using tasks where agents possess correct but incomplete information, we will test the hypothesis that the token embeddings associated with key symbols will evolve to reflect changes in the inferred relationships among them.

Research prerequisites - Basic understanding of machine learning and large language models, hands-on coding experience in Python.

You will have the opportunity to gain 

• Hands-on understanding of multi-agent LLMs.
• An understanding of embedding spaces for LLMs.

Project 2

What is true and what is not in GraphRAG?

Research lead - Dr Anelia Kurteva

Research objective

In this project, we aim to investigate Large Language Models (LLMs) sensitivity to hallucinations and incorrect information by injecting them with knowledge graphs of both correct and incorrect facts via GraphRAG. By injecting correct and incorrect knowledge graphs via GraphRAG, the research evaluates how models handle contradictory information and whether they can effectively identify and filter out factual inconsistencies. The project examines model behaviour in the presence of contradictory inputs, specifically assessing an LLM's capacity for conflict resolution and error detection within its retrieval‑augmented generation pipeline.

Research prerequisites - Basic understanding of technologies to carry out this project - LLMs, ontologies, knowledge graphs.

You will have the opportunity to

• Study how LLMs handle contradictory knowledge graphs using GraphRAG.
• Learn techniques for detecting hallucinations, conflicts, and errors in retrieval‑augmented generation.

Project 3

Regulatory document summarisation with dynamic chunking and adaptive length control using LLMs.

Research lead - Dr Mubashir Ali

Research objective

This project explores how large language models (LLMs) can be used to summarise long and complex regulatory and legal documents in a reliable and controllable way. It investigates multiple summarisation strategies, i.e. whole‑document, truncation‑based, and hierarchical summarisation and compares their effectiveness on long regulatory texts. The project also examines fixed‑size, semantic, and dynamic chunking strategies and how these interact with LLM‑based summarisation. A key focus is adaptive length control to ensure summaries meet required length and detail constraints while preserving legal meaning and structure, leading to a comparative analysis and prototype pipeline for accurate, scalable, and controllable summarisation of regulatory documents.

You will have the opportunity to

• Explore LLM summarisation methods for long regulatory and legal documents.
• Learn adaptive length control and chunking techniques for accurate, controllable summaries.

Project 4

Active sensing for remote visit.

Research lead - Prof Eyal Ofek

Research objective

This project investigates how a mixed reality display enables users to view their environment as if they are in a remote site. We will look at building a world model on the fly, of the remote site, using both live observations by multiple sensors in the remote site to build a unified display for the user. Special attention will be done on filling in missing information to generate a complete plausible representation of the remote site (including extrapolation, using past models, and rules) as well as planning future capture based on the user's behaviour to minimise missing information.

You will have the opportunity to 

• Build real‑time mixed‑reality world models from live multi‑sensor data.
• Learn methods to fill missing information and plan future capture based on user behaviour.

The University of Birmingham is at the forefront of developing technology for technical AI Safety. This summer, we offer a unique opportunity for taught students to study alongside leading researchers in AI Safety as part of a collaborative initiative with the University of Manchester, funded by the Advanced Research and Invention Agency (ARIA).

This project offers a generous scholarship worth up to £5000, enabling participants to fully engage in an intensive research-based learning experience in AI Safety. Successful scholarship awardees will pay no programme fees and receive free accommodation at the University of Birmingham halls of residence.

Learn more about the scholarship and how to apply.

Project 1

Automated Manufacturing Design

Research team

Mirco Giacobbe, Leonardo Stella, Adam Szekely, Gokhan Tut

Research objective

Predictive modelling of biopharmaceutical stability remains one of the most significant bottlenecks in modern drug development. While traditional machine learning methods offer speed and flexibility, they often lack the physical grounding required to satisfy rigorous regulatory standards for shelf-life estimation.

In this project, you will explore the development of a framework for assessing biologic degradation by bridging the gap between high-dimensional structural data and fundamental chemical kinetics. The goal is to understand whether a physics-informed modelling framework can provide the level of safety, reliability, and interpretability required to support critical decision-making within the strict regulatory environment of biopharmaceutical manufacturing.

Project 2

Hardware-level Verification

Research team

Mirco Giacobbe, Edwin Hamel-de le Court, Edoardo Manino

Research objective

Hardware-level considerations are often overlooked when evaluating the safety of AI models. Yet details such as quantisation, sampling, and software implementation can substantially alter model behaviour during deployment. AI models, including neural networks, are not immune to these effects.

In this project, you will study research at the intersection of the digital and physical worlds while investigating several key questions: Do AI deployment pipelines truly preserve the intended behaviour of models? Are AI-based digital controllers robust and stable when interacting with a physical environment? Can we formally and automatically verify that AI decision-making systems remain safe when implemented on real hardware?

Project 3

Privacy-preserving Verification

Research team

Pascal Berrang, Mirco Giacobbe, Xiao Yang

Research objective

Formal verification of systems has focused for decades on the challenge of scalability. However, ensuring AI Safety at a global scale introduces an additional challenge–confidentiality. Existing verification technologies typically require access to a formal and detailed description of the system, which can create tension between providers’ desire to protect their intellectual property and regulators’ need to ensure safety.

This project investigates how one can convince an external verifier that a system is safe without revealing the system itself. Participants will explore novel questions at the intersection of formal verification, cryptography, and AI, studying new approaches to safety assurance in settings where confidentiality is essential.

Project 4

Formal Certification for AI Safety

Research team

Mrudula Balachander, Mirco Giacobbe, Grigory Neustroev, Diptarko Roy

Research objective

Formally verifying an AI system involves determining whether it satisfies a precise specification of its intended behaviour. Standard testing techniques are easy to implement but inherently non-exhaustive and cannot provide formal guarantees of safety, whether absolute or probabilistic.

In this project, participants will learn advanced techniques at the intersection of logic and AI while working with a team investigating formal specification languages for AI Safety and algorithms for verifying them. A central goal is to understand what constitutes a formal certificate of safety and how learning systems might eventually be enabled to generate proofs of their own safety.