Structure from Reasoning, Numbers from Search: On-Premise Open LLMs as Structural Priors for Coupled MIMO Controller Tuning

The paper introduces a novel approach for tuning strongly coupled MIMO controllers using on-premise open-source large language models (LLMs) as structural priors. While classical tuning methods outperform LLMs in simpler scenarios, the LLM demonstrates significant advantages in complex systems, achieving a penalized cost of J ~ 16.9 with only 18 evaluations, compared to traditional methods. This research delineates the conditions under which LLMs can effectively aid in control tuning, emphasizing their role in providing sample efficiency and interpretability rather than direct optimization.