This paper proposes a new paradigm that reframes path recommendation (PR) as a natural language generation problem to overcome the inflexibility and generalization difficulties of existing PR methods. We propose PathGPT, a system that converts existing path data into natural language format, stores it, and then generates paths by inputting it into a large-scale language model (LLM) pre-trained through an information retrieval system along with user requirements. We demonstrate that this system enables adaptive zero-shot path generation without retraining in various scenarios. Experimental results using a large-scale path dataset demonstrate that it outperforms existing learning-based methods.
