Closed-Loop Control with Rule-Aligned Small Language Models and Multi-Agent Self-Correction
06:00 · July 14, 2026 · arXiv cs.AI RSS
A key step toward autonomous industrial operation is the ability to create and reconfigure control policies from natural-language requirement specifications, with minimal or no manual redesign. In this setting, policy generation by AI agents can be a credible path when paired with a plant-aware validator (e.g., a digital twin) that can check generated candidate actions before execution. However, practical deployment is constrained by inference latency and compute footprint: large cloud-based models are often too slow, opaque, or data-sensitive for edge closed-loop use. This work investigates whether a compact Small Language Model (SLM) can be retrained for control reasoning and embedded in a validator-guided correction loop. We use a Qwen2.5-1.5B model aligned via Group Relative Policy Optimization (GRPO), combined with (i) an action agent, (ii) a symbolic/digital-twin-style validation layer, and (iii) a reprompting agent that iteratively steers outputs toward valid actions. In randomized thermal-control simulations (30 experiments with 500 steps each), the framework achieves 91.5% average action-alignment accuracy (86.3%--100% across cases) at 3.84\,s mean inference latency. Under symbolic re-mapping, it maintains a 95% in-range rate, indicating robust physical regulation despite reduced token-level agreement. These results support SLM+validator architectures as a practical path toward reconfigurable autonomous control at the edge.
Summary
A recent study demonstrates how a compact 1.5B-parameter language model can serve as the reasoning core for closed-loop thermal control without relying on cloud-hosted LLMs. The authors align Qwen2.5-1.5B through Group Relative Policy Optimization (GRPO) so that the model internalizes numerical and logical rules for thermal regulation, then embed it in a three-agent loop. An Action Agent proposes control moves together with a chain-of-thought trace; a symbolic Validation Agent compares each proposal against an expert rule set derived from a digital-twin model; and a Reprompt Agent supplies corrective hints when discrepancies arise. Because the alignment phase never exposes the model to reprompting examples, any subsequent self-correction emerges at inference time.
In thirty randomized 500-step thermal simulations the architecture records 91.5 % average action-alignment accuracy, with per-run figures ranging from 86.3 % to 100 %, while sustaining a mean latency of 3.84 s on edge-class hardware. When the symbolic validator remaps out-of-range outputs, physical regulation remains intact, yielding a 95 % in-range rate across trials. The design therefore separates token-level agreement from functional safety, allowing the system to tolerate occasional reasoning deviations without violating thermal constraints.
The work highlights three practical contributions: a reproducible GRPO pipeline that converts a general-purpose SLM into a control-specific policy; a zero-shot validator loop that supplies corrective feedback without additional fine-tuning; and empirical evidence that the resulting controller preserves regulation performance under interface perturbations. These elements together point toward reconfigurable, plant-aware autonomy that can be deployed locally while still incorporating high-level natural-language specifications.
Why it matters
Directly applicable to Dutch industrial automation and process-control SMEs seeking low-latency, sovereign edge AI; strong technical depth and reproducibility elements suit advanced researchers.
