Applications accepted all year round. Funded PhD Project (Students Worldwide).
Project Overview
As foundation models evolve into autonomous, interacting multi-agent systems, safety risks, such as covert collusion, deceptive coordination, goal drift, and cascade failures, emerge deep within internal activation states rather than just observable conversational logs.
This PhD project bridges mechanistic interpretability and multi-agent safety, moving beyond black-box evaluations. By leveraging circuit identification, linear probes, and activation steering alongside communicative representations, the project aims to identify the causal mechanisms driving multi-agent safety dynamics and engineer latent-level runtime guardrails.
Core Research Themes
- Internal Safety Circuits: Reverse-engineering internal circuits that govern honesty, instruction adherence, and tool use during multi-agent interactions.
- Inter-Agent Information Pathways: Tracing how internal representations drive communicative actions and how communicative inputs reshape recipient latent states.
- Emergent Failure Mechanisms: Uncovering root internal causes of sycophancy, cascading hallucinations, and covert coordination.
- Latent-Level Interventions: Designing internal monitors and representation-steering techniques to mitigate risks at runtime.
The candidate will have substantial intellectual freedom to shape their specific research focus within these directions.
Supervision Team
Lead Supervisor: Dr. Ruizhe Li (Mechanistic interpretability, AI safety, reliable multimodal foundation models)
Co-Supervisor: Prof. Mark Lee (NLP, dialogue systems, interpretability, semantics/pragmatics)
Candidate Profile
Essential:
- A 2:1/First-Class Honours or Master’s degree in Computer Science, AI, Mathematics, Data Science, Physics, or a quantitative discipline.
- Strong Python proficiency and practical experience with deep learning libraries (e.g., PyTorch, Hugging Face).
- Solid grasp of ML mathematics (linear algebra, probability, calculus, optimization).
- Familiarity with transformer architectures and modern LLMs and strong English communication skills.
Desirable:
- Experience with mechanistic interpretability tools and methods (e.g., TransformerLens, circuit analysis, representation steering).
- Familiarity with multi-agent RL, agentic LLM frameworks, or game theory.
- Demonstrated research curiosity (e.g., prior publications/preprints, dissertation, or open-source research implementations).
How to Apply
Informal enquiries are strongly encouraged. Email Dr. Ruizhe Li at r.li.7@bham.ac.uk with the subject line:
[26/27 PhD Enquiry] Multi-Agent Safety - [Your Name]
Please attach:
- CV & Academic Transcripts
- Brief statement of interest (<1 page) outlining your motivation and research questions you wish to explore
- Writing sample (e.g., thesis chapter, technical report, or preprint), if available
Formal applications must be submitted via the University of Birmingham Postgraduate Application Portal with this link (select Computer Science PhD, nominating Dr. Ruizhe Li).
Funding Notes
The studentship provides: Tuition fees at the UK home student rate. A tax-free maintenance stipend for 3.5 years at the current UKRI rate, which is £21,805.
The studentship covers fees at the Home rate (UK and EU applicants with pre-settled/settled status and meet the residency criteria). International applicants are welcome to apply but will be required to cover the difference between Home and International fees.