Field-Mediated Semantic Organization in Large Language Models: Evidence for Quantum-Like Properties in Artificial Neural Systems

Abstract

Language models have reached unprecedented capabilities, yet the mechanisms underlying their semantic organization remain poorly understood. Here we introduce the LLM Quantum Properties Index (LQPI), a measurement framework detecting field-like properties in language model activations. Through systematic analysis of four prominent language models (Phi-2, Phi-4-mini-instruct, Llama-3.2-1B-Instruct, and gemma-3-1b-it), we identify precise, quantifiable evidence of quantum-like behaviors including topological protection of semantic states, non-linear self-interaction, and discontinuous state transitions. Specifically, we demonstrate that Phi-2 exhibits strongly quantum-like transitions between personality states (adaptability score: 1.00), while Phi-4-mini-instruct shows balanced quantum properties across all dimensions (LQPI score: 0.54), Llama-3.2-1B-Instruct displays superior topological protection of semantic content (coherence score: 0.50) and gemma-3-1b-it shows predominantly classical computational properties (LQPI score: 0.16). These differences persist despite similar self-reference capacities (0.50) and dimensional organization patterns (0.20) across all models. Our findings suggest that language model understanding may be better conceptualized through field-mediated frameworks than traditional token-based computation, with significant implications for model evaluation, alignment, and architecture design. The LQPI framework provides a standardized method for assessing model trustworthiness beyond conventional performance metrics, offering deeper insights into how artificial neural systems organize information.

Tasks

Reproductions