LLMs as Interpretable Controllers for Dynamical Systems
A recent investigation published on arXiv examines the potential of Large Language Models (LLMs) as interpretable controllers for managing dynamic thermal environments. The study assesses five LLMs of different sizes regarding their proficiency in adhering to setpoints, understanding natural-language instructions, reasoning about the effects of actuators, and utilizing existing model-based knowledge. The scenarios tested include restrictions on heater or fan operations and the use of a physics-based prediction tool. Findings indicate that the effectiveness of control is influenced by model complexity: lower and mid-scale models often misinterpret actuator dynamics or produce inconsistent reasoning, whereas more advanced models like Qwen-3~14B and GPT-4o demonstrate precise temperature regulation and reliable actuator performance.
Key facts
- Study evaluates LLMs as controllers for a dynamic thermal environment
- Five LLMs of varying scales are tested
- Scenarios include penalties on heater or fan usage
- Models have access to a physics-based prediction tool in some cases
- Low- and mid-scale models misinterpret actuator dynamics
- High-complexity models like Qwen-3~14B and GPT-4o achieve accurate temperature tracking
- Study examines ability to follow setpoints and interpret natural-language commands
- Research explores interpretable control with LLMs
Entities
Institutions
- arXiv