SLMs Enable Robust Closed-Loop Control with Multi-Agent Self-Correction

Achieving autonomous industrial operations requires the ability to generate and reconfigure control policies directly from natural language specifications, minimizing manual intervention. While large AI models can generate policies, their high inference latency and computational footprint often make them impractical for real-time, edge-based closed-loop control systems. This presents a significant barrier to practical deployment in industrial settings. This study investigates the potential of Small Language Models (SLMs) for this purpose. Researchers retrained a compact Qwen2.5-1.5B model using Group Relative Policy Optimization (GRPO) and embedded it within a novel architecture. This architecture includes an action agent, a symbolic/digital-twin validation layer, and a reprompting agent that iteratively refines the SLM's outputs to ensure valid actions. This multi-agent self-correction loop allows the SLM to operate effectively within defined rules. In randomized thermal-control simulations, the framework achieved an average action-alignment accuracy of 91.5% with a mean inference latency of 3.84 seconds. Despite some token-level disagreement, it maintained a 95% in-range rate for physical regulation, demonstrating robust performance. These findings suggest that SLM-based architectures, combined with validator-guided correction, offer a practical and reconfigurable path toward autonomous control at the edge, overcoming the limitations of larger models.