LoRaGeo-PSW: a prompt-aligned large language model for few-shot fingerprint geolocation in urban LoRaWAN networks

Accurate and efficient geolocation remains a critical need for low-power IoT networks, particularly in large urban environments where GNSS-based positioning is often infeasible. Fingerprint-based localization using LoRaWAN signals offers a scalable alternative, but conventional methods depend on rigid matching algorithms and static radio maps, leading to poor performance in complex cityscapes. This work proposes LoRaGeo-PSW, a novel geolocation framework based on large language models (LLMs) that aligns structured wireless signal features with prompt-driven reasoning. Built upon a GPT-2 foundation, the model encodes LoRaWAN fingerprints—including RSSI and SNR from multiple gateways—as token sequences, and interprets them in context using a structured workflow-aware architecture. By integrating signal processing workflows via lightweight adapter modules and employing low-rank adaptation (LoRA), LoRaGeo-PSW offers both interpretability and parameter efficiency. Crucially, the model enables few-shot localization by conditioning on a handful of example fingerprints, thereby adapting to new environments without retraining. Evaluated on a public LoRaWAN dataset from urban Antwerp with over 130,000 transmissions, the model achieves a median localization error of approximately 150 m, substantially surpassing classical fingerprinting and deep learning baselines. This work introduces a new paradigm for wireless localization, demonstrating that LLMs can effectively bridge structured signal reasoning and geospatial inference through prompt-driven alignment.