Self-Evolving In-Context Learning Boosts Wireless Beamforming

The paper introduces an advanced in-context learning (ICL) framework designed to significantly improve pilot-based beamforming performance in multi-user multiple-input single-output (MU-MISO) wireless communication systems. This novel scheme integrates an ICL-Transformer backbone with dedicated pilot encoder-decoder networks (EDNs) and beamformer EDNs. A standout feature of this ICL network is its ability to handle diverse channel models without requiring explicit retraining, achieved through the dynamic construction of model-specific context datasets. To enhance both convergence and robustness, the framework incorporates three key innovations. First, a curriculum learning (CL) strategy is employed, transitioning smoothly from supervised LMMSE-labeled imitation to unsupervised sum-rate maximization. Second, a self-evolving mechanism dynamically expands and refines the context datasets for all channel models throughout the CL-based training process. Third, a mismatch-aware extension is included, which integrates various mismatches into the general ICL framework, thereby bypassing the need for explicit channel calibrations. Ablation studies confirmed the effectiveness of both the in-context architecture and the enhanced training strategies. Simulation results across a variety of communication environments demonstrate that the proposed scheme rapidly adapts to both seen and unseen channel models without gradient-based parameter updates. It also effectively mitigates mismatch issues through intelligent context constructions, consistently outperforming established beamforming schemes, including the WMMSE benchmark and recent Transformer-based methods, under pilot-based settings.