Large language models (LLMs) have shown promise in intelligent transportation systems, but their direct use in constrained route planning remains unreliable because such tasks require exact numerical consistency and strict compliance with operational constraints. This challenge is particularly important in urban freight and logistics, where routing errors can reduce efficiency and undermine sustainability. To address this issue, this study proposes a supervised fine-tuning (SFT) framework that specializes a general-purpose LLM as an orchestration agent for route planning. Instead of generating routes directly, the model translates natural-language requests into structured function calls that invoke deterministic optimization solvers for the Traveling Salesperson Problem (TSP), Capacitated Vehicle Routing Problem (CVRP), and Vehicle Routing Problem with Time Windows (VRPTW). Experiments on a controlled synthetic benchmark with thousands of routing instances show that direct generation is ineffective for constrained routing, while tool augmentation substantially improves reliability. More importantly, SFT further strengthens function-calling performance, especially on the most challenging VRPTW task, where the overall success rate of the 8B model increases from 0.408 in the zero-shot setting to 0.792 after fine-tuning. The fine-tuned 8B model also outperforms a much larger zero-shot 235B model while requiring far fewer computational resources. These findings indicate that reliable LLM-based transportation decision support is better achieved by combining compact language models with deterministic optimization tools rather than relying on larger models for direct route generation, offering a lightweight and more sustainable path for real-world logistics deployment.
Chen et al. (Fri,) studied this question.