Industrial mechanism-aware multi-agent collaboration of large language models for job shop scheduling optimization

As manufacturing systems transition toward mass customization and high-mix, low-volume production, operational decision-making must increasingly adapt to heterogeneous constraints, dynamic resource states, and evolving objectives. Conventional optimization models, typically engineered for fixed and well-structured scenarios, struggle to maintain robustness under such structural variability. Although large language models (LLMs) demonstrate strong reasoning and generalization capabilities, their deployment in mechanism-constrained physical systems remains limited by insufficient feasibility alignment and controllability. This limitation becomes particularly pronounced in job shop scheduling, a highly constrained combinatorial optimization problem requiring strict adherence to precedence relations, resource exclusivity, and temporal feasibility under dynamic operating conditions. Here, we propose an industrial mechanism-aware multi-agent LLM framework (IMA-LLM) that embeds domain structural constraints directly into a collaborative generation–verification architecture. The framework integrates mechanism-level analysis, constraint-aligned structural decomposition, and solver synthesis within a closed-loop feasibility-aware refinement process, enabling natural-language task descriptions to be systematically transformed into executable and constraint-compliant optimization solutions. By explicitly aligning algorithmic structure with physical scheduling mechanisms, the proposed approach mitigates hallucination-induced infeasibility and improves structural reliability. Extensive experiments on benchmark job shop and flexible job shop scheduling instances demonstrate that IMA-LLM significantly improves validation accuracy, solution stability, and robustness under increasing constraint complexity compared with state-of-the-art baselines. These results establish a scalable paradigm for integrating language intelligence with structured combinatorial decision-making, advancing the reliable deployment of LLM-driven optimization in real-world manufacturing systems.