Haorui Yu

My research asks what an evaluation of a multimodal AI system can claim when the subject is art. Vision-language models now describe, interpret and critique paintings, and they are scored on how well they do it. I work with classical Chinese painting and heritage collections, where the gap between a model's reading and a scholarly one is wide enough to examine closely. My focus is on the scores themselves: what a benchmark result shows about a model, what it shows only about the test, and where human judgement has to stay in the loop.

My first-authored papers have appeared in Findings of EMNLP 2025, WiNLP 2025, the ACM Journal on Computing and Cultural Heritage (2026) and TimelyAI 2026, with another accepted at CAC 2026. I also maintain VULCA, an open-source Python SDK, command-line tool and MCP server for evaluating images against cultural criteria. I am testing these methods with collections: comparing AI-generated descriptions of museum objects with the museum's own catalogue records, and trialling a quality-checking tool on glass-plate and film scans from the University of Dundee's archives. I am the doctoral researcher representative for the SGSAH Modern Languages Catalyst and a reviewer for ACM Computing Surveys.

Names of Supervisors:

• Dr Oisin Keohane (Primary Supervisor, Philosophy)
• Dr Alison Pease (Computing)
• Professor Natasha Lushetich