Describing an object involves more than naming its features. We may also need to explain where those features are, how they relate to one another, and what changes when we consider a different perspective. For AI-assisted interpretation, these relationships raise an important question: can a convincing explanation also be spatially correct?
Testing relationships, not just vocabulary
A REEVALUATE-supported study investigates this question through human anatomy. The researchers evaluated eight large language models using a synthetic, graph-based dataset derived from the Foundational Model of Anatomy.
The questions tested both direct spatial relationships and answers requiring several connected reasoning steps.
Strong performance, but persistent weaknesses
Results varied considerably between models and relationship types. Although leading models performed very strongly on many relations, all showed weaknesses with more abstract concepts such as medial and lateral.
More reasoning steps did not consistently mean worse performance. The structure of the relationships also mattered. The researchers additionally identified cases where instruction-tuned models accepted incorrect premises instead of challenging them.
Together, these findings suggest that a model’s ability to produce fluent answers should be assessed separately from its ability to reason consistently about relationships.
Why this matters for cultural heritage
For cultural heritage, spatial descriptions can influence how people understand an object, an architectural feature or a reconstructed environment.
Imagine a digital guide describing a sculpture. Saying that one detail is “to the left” of another may depend on whether the description follows the viewer’s perspective or the figure’s own orientation. An explanation can sound natural while leaving that distinction unclear.
This illustrates a possible relevance to REEVALUATE: AI-supported interpretation should be checked against the underlying collection information and the intended viewpoint. Teams could test descriptions with known answers, make reference points explicit, and retain expert review where spatial mistakes could mislead visitors.
These are implications for cultural heritage practice, rather than museum applications tested in the study. The research itself concerns anatomical relationships.
Read the publication
Evaluating Large Language Models’ Ability to Reason About Anatomical Spatial Relations
Maria Tsourma; Robert E. Blackwell; Anthony G. Cohn; Anastasios Drosou; Manolis Koubarakis.
Spatial Cognition XIV, Spatial Cognition 2026, Springer, pp. 108–130. First published online on 1 September 2026; proceedings citation year: 2027.
REEVALUATE is funded by the European Union under Grant Agreement No. 101132389. Views and opinions expressed are those of the authors and do not necessarily reflect those of the European Union or the European Research Executive Agency. Neither the European Union nor the granting authority can be held responsible for them.
