Wireless Signals as Tokens: Repurposing Large Language Models for Indoor Localization: (Invited Paper)

Abstract

Indoor positioning based on Wi-Fi signals remains a long-standing challenge due to environmental dynamics, multipath propagation, and hardware heterogeneity across devices and vendors. Existing fingerprinting and model-based methods rely on labor-intensive calibration and retraining, leading to significant performance degradation when network conditions or deployment environments change. To overcome these limitations, we present RATLoc, a radio-agnostic and transferable indoor localization framework built upon a large language model (LLM). RATLoc treats each access point measurement as a token, enabling direct ingestion of unstructured, variable-length, and multi-modal Wi-Fi radio without manual feature engineering. By leveraging parameter-efficient fine-tuning through low-rank adaptation (LoRA), RATLoc rapidly adapts to new environments and radio modalities while keeping the pretrained backbone frozen via few-shot learning, achieving efficient fewshot learning. A deterministic regression head further enables continuous two-dimensional coordinate prediction with consistent outputs suitable for real-world applications. Extensive experiments demonstrate that RATLoc achieves state-of-the-art accuracy while eliminating the need for vendor-specific calibration or fixed radio inputs. Moreover, the model generalizes effectively to unseen environments with only a few calibration samples.