Structured Parsing as Emergent Motor Circuits in LLMs: Empirical Evidence for Dual-Layer Neural Architecture

This paper provides empirical evidence that LLMs process structured formats (JSON/XML) through a dedicated "structural motor circuit" that is functionally independent of semantic processing. Four categories of experiments — three behavioral (DeepSeek-V3) and one internal attention analysis (Llama-3.3-70B, Qwen2.5-72B) — demonstrate: (1) structure-content double dissociation, (2) noise-robust silent recovery with 0% upper-layer awareness, (3) directional switching inertia correlated with closing-tag explicitness, and (4) structural attention heads concentrated in shallow layers (top 20%), separated from deep semantic layers. Cross-format (JSON vs XML) and cross-model (Meta vs Alibaba) consistency confirms this as an emergent property of the Transformer architecture, not a training-specific artifact. The structural circuit functions analogously to procedural memory in biological systems.