If you are interested in applying for a PhD position in my group, please check the research conducted in my group and our recent publications to ensure there is an alignment between what you intend to do and our research activities. Please email me your CV, transcripts of your undergraduate/MSc studies, and a short PhD proposal. There will be interviews for shortlisted candidates.
Note that due to the large number of email enquiries, it is not possible for me to reply to every email on PhD application. If you don't receive my response in two weeks of sending an email enquiry, please assume that your application is not considered further.
Details about applying for PhD research at King's can be found here.
For funding support for postgraduate research, please refer to the King's Funding & Scholarships. Some of the funding schemes are:
State-of-the-art large language models (LLMs) are typically pre-trained through statistical next-token prediction and often post-trained with reinforcement learning (RL) using reward models. While these methods enable strong performance on reasoning tasks such as mathematics and coding, they face key limitations such as: (1) sensitivity to task frequency and input perturbations; (2) dependence on task-specific annotations and preference data, which limits scalability; (3) unreliable self-generated feedback that fails to capture shared reasoning patterns. These constraints arise because LLMs are optimised to solve tasks directly rather than to learn how to reason. For better generalisation, we argue that models need to instead actively engage in learning-to-reason or meta-reasoning processes, and develop the ability to monitor, regulate, and adapt their reasoning strategies when needed.
This project aims to implement the Bayesian meta-reasoning framework we described in our position paper in ICML 2025. The new framework will enable models to move beyond task-specific optimisation towards adaptive reasoning strategies that generalise across domains. The research will first develop unified benchmarks and metrics to evaluate meta-reasoning across domains, moving beyond accuracy to measures of calibration error, logical consistency, and cross-domain generalisation. It will then design adaptive architectures that dynamically combine reasoning skills, using approaches such as Mixture-of-Experts and Bayesian inverse planning. Self-play will be explored to discover scalable and multifaceted rewards without costly human annotations. Finally, latent-space reasoning methods will be explored to reduce cascading errors and improve efficiency. Techniques such as diffusion-based reasoning and looped transformers will be explored. If time permits, the project may extend to interpretable meta-knowledge consolidation, using mechanistic interpretability to link reasoning skills to model components. This would allow selective fine-tuning and more efficient training,
References:
Yan, H., Zhang, L., Li, J., Shen, Z. and He, Y., 2025. LLMs Need a Bayesian Meta-Reasoning Framework for More Robust and Generalizable Reasoning. In 2025 International Conference on Machine Learning (ICML).
Pregnancy care draws on diverse guidelines, hospital protocols, and patient information, but this wealth of evidence also creates challenges for women seeking clear, reliable answers. Guidance can be fragmented, inconsistent, or even contradictory, leaving patients uncertain and clinicians struggling to provide consistent advice. At the same time, large language models (LLMs) and conversational AI tools are becoming increasingly common in digital health, yet current systems largely provide generic responses and fail to address the complexities of conflicting information.
Existing automated approaches to guideline conflict detection are limited, often relying on model-level “black box” decisions. Such approaches suffer from several issues: they may favour whichever recommendation appears most frequently, or give undue weight to options mentioned at the beginning or end of a document. Beyond accuracy, this kind of opaque decision-making undermines patient trust, as it provides no clear rationale for why one recommendation is preferred over another.
Our project is therefore novel in advancing fundamental methods for: formally representing heterogeneous guideline knowledge; algorithmically detecting inconsistencies; evaluating and weighting conflicting evidence; and designing explainable reconciliation strategies suitable for patient use. By shifting the focus from simple question answering towards conflict-aware, explainable reasoning, this project addresses a key technical and clinical gap.
The overarching aim is to develop a transparent and trustworthy QA framework that can provide reliable answers even when knowledge sources disagree. Specifically, the project will (1) create formal representation models for pregnancy-related guidance, (2) design algorithms to detect and categorise conflicts, (3) develop evidence weighting and reconciliation strategies, and (4) integrate explainable mechanisms into a patient-facing QA system. Pregnancy care is chosen as the initial testbed given its direct impact on maternal and fetal outcomes, but the proposed methods are domain-agnostic and can be applied across healthcare more broadly.
References:
Braun, T., Rothermel, M., Rohrbach, M., and Rohrbach, A. (2025). Defame: Dynamic evidence-based fact-checking with multimodal experts. In: International Conference on Machine Learning (ICML).
Tang, X., Zou, A., Zhang, Z., Li, Z., Zhao, Y., Zhang, X., Cohan, A., and Gerstein, M. (2024). Medagents: Large language models as collaborators for zero-shot medical reasoning. In Findings of the Association for Computational Linguistics (ACL).
Venktesh, V. and Setty, V. (2025). Factir: A real-world zero-shot open-domain retrieval benchmark for fact-checking. In: Proceedings of the ACM Web Conference 2025 (WWW). Resource Track.
Wu, H., Zeng, Q., and Ding, K. (2024). Fact or facsimile? evaluating the factual robustness of modern retrievers. In: Proceedings of the 34th ACM International Conference on Information and Knowledge Management (CIKM).