How do Large Language Models Understand Trajectory Data? Insights from Various Trajectory Formatsand Response Strategies for Transportation Mode Detection

The effectiveness of large language models (LLMs) in transportation mode detection remains underexplored, creating a significant research gap in understanding how these models process trajectory data. This study uses the Geolife dataset to investigate the ability of pre-trained and fine-tuned LLMs to detect transportation modes across 14 trajectory formats, categorized into overview information, coordinate-based, and spatial encoding. Meanwhile , two response strategies are compared: direct answer and Chain-of-Thought (CoT) reasoning. The results show that fine-tuning significantly enhances the classification performance for all trajectory formats. Among the evaluated formats, the coordinate-based format with timestamps achieves the highest accuracy of 85.2% after fine-tuning using the direct answer strategy. The direct answer strategy proves to be more effective than the CoT strategy, reaching an average 49.0% improvement in accuracy via fine-tuning. Additionally, the model exhibits systematic misclassification patterns, reflecting challenges in distinguishing between transportation modes with similar movement characteristics. Furthermore, our analysis reveals that hallucinations are prevalent in CoT responses, particularly of the types of input-conflicting hallucinations and factual inaccuracies, which increase the likelihood of misclassification. These findings highlight the potential of LLMs in transportation mode detection while emphasizing the need for enhanced trajectory formats, improved response strategies, and strategies to mitigate hallucinations.