Beyond Shortest Path: Agentic Vehicular Routing with Semantic Context

TL;DR

The paper addresses the gap between traditional fastest-route optimization and user-centric, semantically rich routing in urban mobility by proposing PAVe, a hybrid system that integrates a classical multi-objective router with an LLM-driven semantic reasoning layer. It employs a candidate route set $P=\{\rho_0,\dots,\rho_k\}$ from a multi-objective graph $G=(V,E)$ and enriches options through a geospatial POI cache and task-context signals, using a ReAct-based agent to select and justify the best route. Key contributions include the PAVe architecture with a five-step pipeline (Task Classification, Candidate Route Generation, Contextual Enrichment, Agent Evaluation, Feedback Loop), an experimental evaluation across four LuST-based scenarios, and a public release of prompts and code. The results demonstrate that LLM-assisted routing can achieve high accuracy in aligning routes with complex user intents, while revealing limitations in action reliability and the single-waypoint feedback loop, guiding future work toward multi-stop planning, user profiling, and real-time context integration. Overall, the hybrid approach offers a scalable path to personalized, adaptive urban mobility that can reduce congestion and emissions by aligning routing decisions with semantic user needs and environmental considerations.

Abstract

Traditional vehicle routing systems efficiently optimize singular metrics like time or distance, and when considering multiple metrics, they need more processes to optimize . However, they lack the capability to interpret and integrate the complex, semantic, and dynamic contexts of human drivers, such as multi-step tasks, situational constraints, or urgent needs. This paper introduces and evaluates PAVe (Personalized Agentic Vehicular Routing), a hybrid agentic assistant designed to augment classical pathfinding algorithms with contextual reasoning. Our approach employs a Large Language Model (LLM) agent that operates on a candidate set of routes generated by a multi-objective (time, CO2) Dijkstra algorithm. The agent evaluates these options against user-provided tasks, preferences, and avoidance rules by leveraging a pre-processed geospatial cache of urban Points of Interest (POIs). In a benchmark of realistic urban scenarios, PAVe successfully used complex user intent into appropriate route modifications, achieving over 88% accuracy in its initial route selections with a local model. We conclude that combining classical routing algorithms with an LLM-based semantic reasoning layer is a robust and effective approach for creating personalized, adaptive, and scalable solutions for urban mobility optimization.

Figures (5)

• Figure 1: The overall of our PAVe framework. a) Where the User chooses the desired trajectory b) With the information of the map, the path algorithm chooses the best two options. c) The LLM Agentic Assistant considers the context of the User and the Environment to include the POIs of the desire in the two options. d) The LLM translates all this for the path algorithm to recalculate the routes. e) The LLM chooses the best option for the User, considering the context, and returns the output.
• Figure 2: PAVe System Overview
• Figure 3: Classification Prompt
• Figure 4: Builds Prompt
• Figure 5: Evaluation Prompt