From fastest to safer and more experience-aware: an LLM-driven framework for personalized multimodal routing

Smart transportation systems are increasingly influencing how travelers navigate multimodal networks. However, most routing services are still optimized for time efficiency and offer limited support for safety-aware, human-centric decisions. This gap is significant because travelers’ satisfaction and adoption of multimodal travel depend on how well route guidance reflects real-world trade-offs among efficiency, crash risk exposure, and experiential qualities of the street environment. At the same time, conventional personalization methods that infer preferences from historical behavioral data often face cold-start problems, high data requirements, and limited transferability across contexts. To address these challenges, this study proposes an end-to-end large language model (LLM)–enabled personalized multimodal route planning framework that integrates risk-aware multi-objective route search with human-centric recommendation. Specifically, we first generated Pareto-optimal feasible routes using a label-correcting multi-objective algorithm that jointly considers travel time, mode-specific traffic safety risk, subjective streetscape perception, and transfer burden. Safety risk was quantified through mode-specific multivariate crash models calibrated with historical crash records, and streetscape perception is predicted from street-view imagery using a deep learning model. To facilitate practical deployment, we employed a greedy hypervolume-based sampling strategy to reduce redundancy while maintaining the representativeness of the solution set. Building on the sampled candidate routes, we developed an LLM-based multi-agent recommendation workflow that interprets traveler profiles and trip contexts, generates transparent rationales, and iteratively improves via a self-evolving memory mechanism. Using real-world multimodal network, crash, and street-view data from Daejeon, South Korea, we demonstrate that the proposed framework improves profile–route alignment and identifies alternative routes with lower route-level safety risks across commuting, leisure, and shopping scenarios. The findings suggest that risk-aware, human-centric routing can support safer route-choice decisions in smart mobility services. Furthermore, the self-evolving mechanism mitigates performance disparities across different LLM scales, highlighting a feasible pathway for cost-effective deployment of safety-oriented personalized routing systems in real-world settings.