Strategic Evaluation of Planning Strategies for LLM Agents in Cyber-Physical Systems
This research evaluates how different planning strategies influence outcomes in cyber-physical systems, moving beyond simple success metrics. It introduces a physics-grounded method to test three properties: whether architecture changes results, if execution fidelity needs more than mode agreement, and if scenario diversity helps select the best plan. The findings show that planning strategy heterogeneity and execution fidelity are crucial for reliable autonomous control.