Prompt-Induced Path Reconfiguration in Large Language Models

Can a prompt change not only what an LLM says, but which response pathway becomes easiest to execute? This study examines prompt-induced behavioral reconfiguration across four widely available LLM families. Using a new functional measure of informational impedance, it shows that small, contextually relevant instructions can sharply reduce the cost of activating new response rules, often matching the performance of much longer prompts. The results reveal a central safety problem: when task execution and safety constraints compete inside the same context-sensitive inference process, a prompt may open a lower-impedance path toward an otherwise restricted response. The paper therefore argues for a nomological control layer that separates what a model can generate from what a system is authorized to disclose or execute. Relevant to researchers and practitioners working on LLM safety, prompt injection, jailbreak resistance, guardrails, alignment, adaptive prompting, and AI control architectures.