Who chooses and who keeps time? Typology-grounded LLM user modelling and explained itineraries for cultural heritage, compared with an icon-based recommender
Abstract Cultural heritage guides face a cold-start problem: they must model a visitor they have never seen. Stereotype-based typologies offer one answer, but icon-based elicitation constrains how visitors can describe themselves, and the resulting recommendations are rarely explained. We present Theoros (ACUX-R 2.0), a mobile guide in which a multimodal large language model (LLM) builds a scrutable, multi-label stereotype model of the visitor from photographs and their own words, and then selects and explains the stops of a walking itinerary from a curated heritage dataset. Unlike LLM trip planners that treat the user as a query, the user model is explicit, editable and shown to the visitor, and every explanation must cite the visitor's profile and the place description; deterministic code computes all times from opening hours, walking distances and the weather. In development tests the LLM chose and explained stops reliably but failed at clock arithmetic, and adding temporal rules to the prompt made it worse (from 7 to 3 of 10 valid plans), while scheduling in code yielded 10 of 10. In an exploratory within-subjects study, 21 participants used Theoros and a faithful re-implementation of the original icon-based recommender on the same task. Both scored excellent on the UEQ-S benchmark; Theoros scored higher in hedonic quality ($r = 0.63$, significant after correction for eight outcomes), while perceived transparency, trust and profile accuracy were similar and near the ceiling. Twelve participants preferred Theoros and five the original. We discuss what this implies for LLM-based stereotype user models.
Authors
- Markos Konstantakis (ORCID: https://orcid.org/0000-0003-0656-693X)
Institutions
- University of the Aegean (GR)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-13
- DOI
- https://doi.org/10.21203/rs.3.rs-11529712/v1
- Primary Topic
- Recommender Systems and Techniques
- Type
- preprint