From Trajectories to Relations: Interpreting Overtake Manoeuvres with Large Language Models and Qualitative Spatial Representations (Short Paper)

Abstract

This paper explores whether qualitative spatial encodings make overtake manoeuvres more interpretable to a large language model (LLM). Using simulated traffic trajectories from a microscopic road-network model, we encode the same overtake event at three levels of complexity through Point-Descriptor-Precedence (PDP) inequality matrices. An LLM is then used in a staged, zero-shot API workflow in which longitudinal and lateral descriptor encodings derived from these matrices are interpreted separately and then merged into a final manoeuvre interpretation. Outputs are assessed through expert-based qualitative evaluation focusing on relational correctness, spatial precision, completeness, and interpretive adequacy. For the selected case, the results suggest that the model can produce meaningful natural-language interpretations at all three levels, but that these change as complexity increases. The 2-point representation yields the clearest focal manoeuvre, the 5-point representation offers the best balance between contextual richness and interpretability, and the 10-point representation produces the most nuanced structural reading, although in a less concise form. The paper should be read as an exploratory study of representation effects rather than as a benchmark of LLM performance.