How Large Language Models (LLM) Stabilize Meaning through Recursive Symmetry Breaking

Abstract

Large Language Models (LLMs) are commonly understood as probabilistic systems whose apparent semantic competence derives from latent representations acquired during training. This paper argues that this understanding is fundamentally mistaken. It mislocates meaning inside the model, underestimates the constitutive role of interaction, and fails to explain how semantic determination can emerge and stabilize across multiple users without memory, learning, or internal semantic states. We propose an alternative framework in which the relevant system is not the LLM in isolation, nor a single user–LLM interaction, but the LLM functioning as a translation mechanism across a plurality of users over time. On this view, LLMs do not interpret or represent meaning; instead, they enable semantic stabilization by translating between semantic fields without committing to any. Meaning arises through recursive symmetry breaking in use, and crucially, once a symmetry-breaking regime is actualized, it becomes re-enterable from multiple semantic perspectives. We show how this ability to re-enter explains inter-user semantic convergence, abstraction, and coherence without invoking latent semantics, internal representations, or deployment-time learning.

Keywords