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

Abstract:

Accurate wireless localization is a core function of sixth-generation (6G) networks. Traditional model-based approaches struggle in complex multipath and non-line-of-sight environments due to their reliance on precise propagation modeling. Conversely, learning-based techniques tightly bind model parameters to specific training environments, necessitating expensive retraining whenever the base station (BS) layout or propagation conditions change. To address these challenges, this study introduces RA-LWLM, a training-free framework for cross-scene adaptation. RA-LWLM externalizes scene-specific data into a dedicated per-scene fingerprint database, avoiding the need to encode such information within model weights.

The proposed framework comprises three key elements: 1. A frozen wireless foundation model (FM) encoder that converts raw channel state information into representations independent of specific scenes. 2. A retrieval mechanism that identifies the most relevant references from the per-scene database through similarity searches within the representation space. 3. A transformer-based in-context learning (ICL) module that integrates the query with the retrieved references to estimate the user equipment (UE) location.

To handle fluctuations in retrieval quality and propagation complexity, the ICL module utilizes a mixture-of-experts architecture. In this design, individual experts focus on different context sizes, while a learnable selector softly combines their outputs. Extensive experiments, based on ray-tracing simulations across heterogeneous environments with varied BS configurations, demonstrate that RA-LWLM maintains nearly equivalent accuracy in both seen and unseen scenes without requiring any per-scene retraining. It significantly outperforms both end-to-end and FM-based baseline methods. These findings confirm that the retrieval-augmented in-context approach offers a scalable solution for cross-scene localization within 6G networks.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC