LLM-Based Closed-Loop Control for Industrial Processes | IECON 2026

Vidyashree Rayar presents an LLM-based closed-loop control framework for multivariable industrial processes. Traditional control approaches often rely on fixed control logic and manual tuning, which can limit their ability to adapt to changing process conditions. This work explores whether fine-tuned language models can use sensor feedback to make control decisions automatically. The system is evaluated on a dual-tank process with liquid level and temperature as controlled variables, while the LLM generates actions for the pump, valve, and heater. The framework uses Llama 3.2 3B-Instruct with QLoRA fine-tuning and compares different prompting strategies together with a safety layer. The results show that LLMs can perform closed-loop process control, but prompt design has a strong effect on controller behavior. The experiments also show that predictive context does not always improve performance and that safety mechanisms remain important for reliable deployment. This work was carried out by Vidyashree Rayar as part of her Master’s thesis at the Reliable and Secure Systems Lab (RSS Lab), BTU Cottbus-Senftenberg. 💻 GitHub Repository: https://github.com/Reliable-and-Secur... 🔗 LinkedIn: https://www.linkedin.com/posts/rss-la... 🌐 RSS Lab: https://www.rss-lab.de