Semantic Physics: A Field-Theoretic Account of Discourse Mode Transitions in Large Language Models

Semantic Physics introduces a field-theoretic framework for understanding why large language models (LLMs) sometimes respond with short, direct answers (“Dialog” mode) and other times with long, structured reasoning (“Elaboration” mode), even when queries are semantically equivalent. The paper models these discourse modes as attractor basins in a semantic potential landscape Ψ, with information flow governed by a semantic flux equation B=−σ∇ΨB = -\sigma \nabla \PsiB=−σ∇Ψ. Within this framework, epistemic operators (e.g. prove, derive, analyze) act as semantic field gradients that bias models toward Elaboration, while certain formats (e.g. binary questions) bias toward Dialog under resource constraints. The work combines theory with empirical evidence. It proposes quantitative signatures of mode transitions—captured by the Fold-Onset Triplet (FOT: changes in spectral structure, ideational density, and topical coherence)—and validates fold-like dynamics on >33,000 conversational turns. A manual ablation study (N = 75) across GPT-4o, Claude Sonnet 4.5, and Grok-4 Expert provides causal evidence that epistemic operators materially affect mode selection. A deconfounded 2×2 design (operators × question format), analyzed with appropriate paired tests (McNemar), further reveals model-specific “conductivity” parameters σ that quantify how strongly each model responds to semantic gradients vs. format constraints.