Multi-Level Natural Language Interaction for Eliciting Implicit Preferences in Interactive Vehicle Routing

Abstract

Interactive vehicle routing enables planners to express preferences during optimization, yet existing interaction methods—such as scoring, ranking, and weight adjustment—constrain the richness of preference expression. In practice, planners naturally form multi-level judgments about routing solutions: from specific task assignments, to regional route patterns, to overall solution quality. However, current systems lack mechanisms to capture such semantically rich, multi-level feedback. This study proposes an LLM-enhanced multi-level interaction approach that enables planners to express preferences through natural language at three levels: node-level, regional-level, and solution-level. A large language model serves as a semantic bridge, interpreting natural language feedback and translating it into adjustments to a probability-based preference matrix that guides algorithmic search. The approach is illustrated through scenario-based demonstrations using a road cleaning problem, showing how natural language feedback at different levels can be interpreted and integrated into the optimization process. This work represents a step toward more intuitive and expressive human–algorithm collaboration in vehicle routing.

Keywords: Interactive Vehicle Routing, Large Language Models, Implicit Preference, Human-algorithm Collaboration, Natural Language Interaction

DOI: 10.54941/ahfe1007539