ARC-LLM: Control Theory Makes Multi-Agent AI Safe and Auditable for Process Control
NTNU researchers map industrial control loops to LLM agents running on a $1,000 GPU.
Large language models (LLMs) struggle with narrow domain tasks because they lack bounded context and task constraints. Researchers from NTNU propose a systematic solution derived from Advanced Regulatory Control (ARC) theory: each feedback loop in a process-control chain is mapped to a specialized LLM operator agent. The agent carries only the loop's control-theoretic context—controlled variable, setpoint, chain priority, and selector type. A single orchestrator agent (either deterministic or Claude-based) manages interaction logic including MIN/MAX selectors and override paths. This architecture ensures that every constraint conflict is resolved deterministically, regardless of LLM output, inheriting the safety properties of traditional ARC chains.
The system was evaluated on a real-world dairy barn ventilation case over a four-day mixed-season scenario. Using Qwen 2.5 7B Instruct as operator agents running locally on a 24 GB consumer GPU at a 5-minute cadence, the multi-agent setup produced auditable control trajectories. Each decision was paired with an operator-voice rationale, creating a campaign logbook suitable for regulatory compliance. The approach demonstrates that safe, explainable LLM-driven process control is feasible with modest hardware, opening the door for wider industrial adoption without sacrificing safety or auditability.
- Each feedback loop maps to a single LLM operator agent with narrow, control-theoretic context (CV, setpoint, priority).
- A deterministic orchestrator resolves conflicts via MIN/MAX selectors, ensuring safety regardless of LLM output.
- Runs offline on a $1,000 24GB consumer GPU at 5-minute cadence for real-world dairy barn ventilation control.
Why It Matters
Brings safe, auditable LLM-based process control to modest hardware, enabling industrial automation without expensive infrastructure.