Neurosymbolic spatial reasoning: teaching language models to plan over qualitative geographic knowledge

PhD Students Only About the Project About the project Large language models are fluent but unreliable at spatial reasoning. Given a geographic knowledge graph, they recall isolated facts well, yet they struggle with questions that require several spatial relationships to be chained together. They also cannot reliably write the formal spatial queries that would let a database reason on their behalf. Part of the difficulty lies in how geographic knowledge is represented. In most geographic knowledge graphs, spatial relationships are computed in advance from geometry and stored as flat facts. Systems built on them are therefore tested on retrieving stored answers, not on reasoning. This project takes a different route. It does not ask the model to compute geometry or to write database queries. It investigates whether a model can generate abstract, verifiable reasoning programs over a compact vocabulary of qualitative spatial relations (containment, adjacency, proximity and direction). A symbolic engine then executes each program and checks its result, so that every answer can be traced to the relations that support it. The central question is whether a small, closed and compositional target language makes spatial reasoning at inference time tractable where general query generation has failed. The work combines current research on language models with knowledge representation and qualitative spatial reasoning, and its results are relevant wherever people need trustworthy answers to questions about places. Research questions Can a language model translate unseen natural-language questions about places into valid, executable qualitative reasoning programs at inference time? Does a compact, closed relational vocabulary yield more reliable program generation than full geospatial query languages such as GeoSPARQL? How does reliability change with compositional depth, from single-relation questions to nested and multi-hop reasoning across different families of relations? What you will do You will design the target reasoning language and the methods by which a language model produces programs in it. You will build the symbolic component that executes and verifies those programs over geographic knowledge graphs, and you will evaluate the combined system against established geographic question-answering approaches. There is room to shape the direction of the project around your own strengths, whether these lie in machine learning, in knowledge graphs or in formal reasoning. You will join a research group with a long record in qualitative spatial reasoning, place ontologies and geographic knowledge graphs, and you will be able to draw on its existing work as a starting point. Candidate profile We are looking for a candidate with a strong background in computer science, artificial intelligence or a related quantitative discipline. Solid programming skills in Python and familiarity with machine learning are expected. Experience with large language models, knowledge graphs, semantic web technologies or symbolic reasoning is an advantage. An interest in combining neural and symbolic methods is essential. No prior training in geography is required. Funding notes This project is open to self-funded applicants and to applicants who hold, or are applying for, funding from other sources. English language Applicants need IELTS 6.5 overall with at least 6.0 in each subskill, or an equivalent score in another test accepted by the University. A degree taught in English and awarded by a UK institution may exempt an applicant from the test. How to apply Please contact Dr Abdelmoty before applying to discuss the project, your background and possible start dates. Email: abdelmotyai@cardiff.ac.uk Submit a PhD application through the Cardiff University Computer Science and Informatics programme page, choosing Doctor of Philosophy and Full-Time. Name this project title and Dr Abdelmoty as the proposed supervisor, and state your source of funding in the funding section. Keywords Large language models, neurosymbolic AI, spatial reasoning, knowledge graphs, question answering, semantic web, qualitative spatial representation. References Cohn, A. G. and Blackwell, R. E. (2024). Evaluating the Ability of Large Language Models to Reason about Cardinal Directions. 16th International Conference on Spatial Information Theory (COSIT 2024), LIPIcs 315, 28:1-28:9. Ji, Y., Gao, S., Nie, Y., Majić, I. and Janowicz, K. (2025). Foundation Models for Geospatial Reasoning: Assessing the Capabilities of Large Language Models in Understanding Geometries and Topological Spatial Relations. International Journal of Geographical Information Science, 39(9), 1866-1903. doi:10.1080/13658816.2025.2511227 Khalid, I., Nourollah, A. M. and Schockaert, S. (2025). Large Language and Reasoning Models are Shallow Disjunctive Reasoners. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics, 8843-8869. Mai, G., Janowicz, K., Zhu, R., Cai, L. and Lao, N. (2021). Geographic Question Answering: Challenges, Uniqueness, Classification, and Future Directions. AGILE: GIScience Series, 2, 8. Abdelmoty, A. I., Muhajab, H. and Satoti, A. (2024). Spatial Semantics for the Evaluation of Administrative Geospatial Ontologies. ISPRS International Journal of Geo-Information, 13(8), 291. doi:10.3390/ijgi13080291