Beyond Hooking Onto the World: Referential Profiles and the Numerical Structure of LLM Grounding

This paper explores the grounding problem in large language models (LLMs), proposing a shift from classical symbol grounding to a more nuanced vector grounding approach. It argues that reference in LLMs is context-sensitive and discourse-level, shaped by usage patterns rather than static links to objects, and emphasizes the importance of numerical realization in understanding how LLMs parameterize linguistic traces. The findings suggest that while LLMs do not achieve human-like reference, they may exhibit derivative, profile-based forms of reference that are mathematically structured through mechanisms like weights and attention states, which is crucial for practitioners aiming to enhance LLM interpretability and grounding in practical applications.