From Task-Specific Automated Valuation to Real Estate Foundation Models: Conceptual Boundaries, Constitutive Criteria and a Research Agenda

Automated Valuation Models (AVMs) have progressed from hedonic and spatial statistics to machine-learning architectures, yet most remain optimized for one target, market, and observation period. This review asks whether the foundation model paradigm defines a substantively different research object. A structured, concept-centric review combines direct evidence from valuation, AVMs, multimodal property analysis, and professional standards with primary research on tabular, temporal, geospatial, urban, and multimodal foundation models. This framework separates computational foundationality from real-estate domain validity. A Real Estate Foundation Model (REFM) is a domain-scale, pre-trained, and adaptable model that learns reusable representations from broad real estate evidence and demonstrates support for heterogeneous tasks across non-identical settings. Four criteria define the computational baseline: domain-broad pre-training, a shared reusable backbone, adaptability to heterogeneous tasks, and demonstrated non-trivial transfer. A fifth criterion concerns the semantic and institutional foundation of real estate and establishes “real estate domain validity”. The core model’s capability is distinct from its implementation. The study concludes that a core model applied to real estate data is not necessarily a core model of real estate markets: foundation-model status must be demonstrated through the reuse of the representation, transferability, and semantic validity when compared with strong, specialized alternatives.