In-Context Radio Map Estimation via Ripple Autoregressive Modeling

Accurate radio map estimation is critical for wireless applications such as coverage planning, localization, and network deployment. However, most existing methods follow a supervised learning mindset, designing various U-Net-based model architectures or loss functions that rely on costly labeled data and delicate model training. Inspired by the remarkable generalization ability of large language models, we are the first to formulate radio map estimation as an in-context learning (ICL) problem, leveraging a pretrained large autoregressive vision model (LAVM) to predict the radio map for a new transmitter (Tx) position, prompted by a few input-output demonstrations without requiring model updates. We propose Ripple, a novel ICL framework that integrates visual tokenization with a ripple autoregressive modeling strategy to explicitly capture the causal structure of wireless signal propagation from the Tx outward. Furthermore, we introduce a two-stage generation strategy for coarse-to-fine prediction to better model non-line-of-sight (NLoS) propagation effects. Extensive experiments demonstrate that Ripple outperforms ICL baselines, highlighting its effectiveness and generalizability in radio map estimation.