Foundation Models for Tabular Data within Systemic Contexts Need Grounding (2505.19825v1)

Abstract: Current research on tabular foundation models often overlooks the complexities of large-scale, real-world data by treating tables as isolated entities and assuming information completeness, thereby neglecting the vital operational context. To address this, we introduce the concept of Semantically Linked Tables (SLT), recognizing that tables are inherently connected to both declarative and procedural operational knowledge. We propose Foundation Models for Semantically Linked Tables (FMSLT), which integrate these components to ground tabular data within its true operational context. This comprehensive representation unlocks the full potential of machine learning for complex, interconnected tabular data across diverse domains. Realizing FMSLTs requires access to operational knowledge that is often unavailable in public datasets, highlighting the need for close collaboration between domain experts and researchers. Our work exposes the limitations of current tabular foundation models and proposes a new direction centered on FMSLTs, aiming to advance robust, context-aware models for structured data.