Qwen2.5-1.5B multi-agent system achieves 91.5% autonomous control accuracy

A key challenge in autonomous industrial operations is generating and reconfiguring control policies from natural-language requirements without manual redesign. Large cloud-based models are often too slow, opaque, or data-sensitive for edge closed-loop use. To address this, researchers at an undisclosed institution (paper accepted at IEEE CCTA 2026) propose a framework that combines a compact Small Language Model (SLM) with a validator-guided correction loop. They use the Qwen2.5-1.5B model aligned via Group Relative Policy Optimization (GRPO), integrated with three agents: an action agent that produces candidate actions, a symbolic/digital-twin validation layer that checks physical plausibility, and a reprompting agent that iteratively corrects invalid outputs.

In randomized thermal-control simulations across 30 experiments with 500 steps each, the framework achieved 91.5% average action-alignment accuracy (ranging from 86.3% to 100% across cases) at a mean inference latency of just 3.84 seconds. Under symbolic re-mapping tests, it maintained a 95% in-range rate, indicating robust physical regulation even when token-level agreement dropped. These results suggest that SLM+validator architectures can provide a practical, reconfigurable path for autonomous control at the edge, balancing accuracy, speed, and compute efficiency without relying on large cloud models.