LLMs Show Promise as Interpretable Controllers for Dynamic Systems.

Large Language Models (LLMs) are increasingly used for decision-making, but their potential as controllers for physical systems has been largely unexplored. This research investigates LLMs' ability to act as interpretable controllers within a dynamic thermal environment, assessing their capacity to follow setpoints, interpret natural language commands, reason about actuator effects, and integrate prior model-based knowledge. Five LLMs of varying scales were tested under different scenarios, including those with usage penalties and access to a physics-based prediction tool. The findings indicate that control performance correlates with model complexity: smaller and mid-scale models often misinterpreted actuator dynamics, while high-complexity models such as Qwen-3 14B and GPT-4o demonstrated accurate temperature tracking, stable actuator usage, and provided coherent explanations consistent with physical principles. Crucially, incorporating a physics-based model significantly enhanced control smoothness and energy efficiency by enabling anticipatory decision-making. A detailed analysis of reasoning patterns showed a clear progression from causal misinterpretation in less capable models to cohesive, temporally aware reasoning in more advanced ones. These results suggest that sufficiently capable LLMs, when grounded in domain knowledge, can serve as interpretable controllers, opening doors for hybrid control strategies combining model-based and language-driven approaches.