Large Language Models for Control
Adil Rasheed, Oscar Ravik, Omer San
TL;DR
The paper addresses the challenge of designing control systems that adapt to changing conditions without extensive control-engineering. It proposes a framework where large language models operate in a loop with goals, historical data, and predictive tools to generate actuator commands, evaluated on a greenhouse testbed. Key contributions include demonstrating that prompt-only LLMs can provide viable control, showing how SQL and predictor-assisted variants alter behavior and adaptability, and exposing a transparent prompt–evidence–decision chain for auditability. The findings suggest LLMs can serve as a flexible, auditable, human-in-the-loop control paradigm with practical implications for cyber-physical systems and data-centric infrastructures.
Abstract
This paper investigates using large language models (LLMs) to generate control actions directly, without requiring control-engineering expertise or hand-tuned algorithms. We implement several variants: (i) prompt-only, (ii) tool-assisted with access to historical data, and (iii) prediction-assisted using learned or simple models to score candidate actions. We compare them on tracking accuracy and actuation effort, with and without a prompt that requests lower actuator usage. Results show prompt-only LLMs already produce viable control, while tool-augmented versions adapt better to changing objectives but can be more sensitive to constraints, supporting LLM-in-the-loop control for evolving cyber-physical systems today and operator and human inputs.