Semantic-Driven Mathematical Modeling and Adaptive Refactoring for Logistics Optimization: A Language-Model Orchestrated Optimization Pipeline

While large language models (LLMs) offer a convenient natural language interface for logistics optimization problems, it remains challenging to directly generate reliable mathematical models and executable code from unstructured text requirements. LLMs tend to produce invalid constraints or syntactically incorrect code. In addition, traditional logistics optimization methods lack the flexibility to adjust warehouse rules or operational goals without manual expert intervention. To address these issues, we propose LOOP (a Language-Model Orchestrated Optimization Pipeline), which automatically translates natural-language requirements into optimization algorithm code while retaining the rigor of classical models and solvers. LOOP leverages task-specific agents to construct accurate mathematical models and adopts a difference-driven code generation approach. First, it synchronizes model changes into executable code via semantic mapping and ensemble difference analysis. Second, it incorporates a multi-layered verification mechanism to detect and correct pre-execution logical inconsistencies between the model and code. We evaluate LOOP on the capacitated vehicle routing problem (CVRP) using the Augerat benchmark dataset (six difficulty levels). Experimental results show that the final code achieves a 96.3% pass rate across 54 cases, and generated solutions differed by only 1.5% from expert baselines. These findings confirm that LOOP improves the agility of dynamic constraint solution development without sacrificing quality.