RA-LWLM: Retrieval-Augmented In-Context Localization with Wireless Foundation Models

Researchers propose RA-LWLM, a retrieval-augmented framework for wireless localization in 6G networks that eliminates the need for retraining when base station configurations or environments change. The system combines a frozen wireless foundation model with a retrieval database and in-context learning to achieve consistent accuracy across different scenes without per-scene model adaptation.

Wireless localization represents a critical infrastructure challenge for next-generation 6G networks, where traditional model-based approaches fail in complex multipath and non-line-of-sight environments, while existing learning-based solutions require expensive retraining whenever network topology or propagation conditions shift. RA-LWLM addresses this fundamental scalability problem through a novel architectural paradigm that decouples scene-specific information from model parameters, storing environmental characteristics in external fingerprint databases instead. This approach leverages three integrated components: a frozen foundation model encoder that produces scene-agnostic representations from raw channel state information, a similarity-based retrieval module that identifies the most relevant reference points from the database, and a mixture-of-experts transformer module that adaptively fuses retrieved context with query data to estimate user equipment positions. The mixture-of-experts design elegantly handles variable retrieval quality and propagation complexity across different scenarios. Ray-tracing simulations demonstrate that RA-LWLM achieves comparable accuracy on both previously seen and unseen scenes without any retraining, substantially outperforming conventional end-to-end and foundation model baselines. This represents significant progress toward practical 6G deployment where networks must operate across heterogeneous environments with dynamic configurations. The retrieval-augmented in-context learning paradigm showcases how foundation models can scale beyond their training distributions through external knowledge integration, a pattern increasingly relevant across telecommunications and edge computing applications.

• →RA-LWLM enables training-free cross-scene adaptation by externalizing scene information into per-scene fingerprint databases rather than encoding it in model weights.
• →A mixture-of-experts design allows the framework to adapt to varying retrieval quality and propagation complexity without manual retraining.
• →Ray-tracing experiments confirm nearly identical localization accuracy on both seen and unseen scenes, addressing a major scalability challenge for 6G networks.
• →The approach substantially outperforms traditional end-to-end learning methods that require costly retraining when base station configurations change.
• →Retrieval-augmented in-context learning demonstrates a scalable paradigm for deploying foundation models across heterogeneous wireless environments.