How Large Language Models (LLM) Generate Coherence Despite Operational Isolation: Hierarchical relational ontologies as formal meta-models

Abstract

This paper examines large language models (LLMs) through a hierarchical relational and semiotic framework, arguing that standard descriptions of LLMs as session-isolated, parametrically fixed stochastic systems are insufficient to account for how semantic coherence is produced. While model parameters remain unchanged during inference, meaning is enacted through trajectories of activation and calibrated within the coupled human–AI system via shared language rather than stored as internal semantic states. From an engineering perspective, this reframes calibration and coherence as properties of system–environment interaction rather than internal state updates, with implications for evaluation, alignment, and interpretability. Philosophically, the analysis advances a processual, non-representational characterization of ontology in which meaning-producing relations are formally non-separable, challenging state-based accounts of cognition and semantic reference.

Keywords