POLAR: Permutation-Oriented DeepSets Learning for Adaptive Beamforming in 6G AI-Native Networks

Beamforming (BF) is a vital technique for interference management and spatial multiplexing in modern wireless networks. While conventional BF approaches depend on model-based optimization, recent studies have investigated deep learning (DL) to directly derive BF policies from data. However, current learning-based approaches generally represent interferer information as ordered feature vectors, even though interferers inherently constitute an unordered set in dynamic wireless environments. This inconsistency may result in policies whose functionality is dependent on the particular ordering of interferer inputs, thus reducing their structural resilience. In this paper, the multi-interferer BF problem is reformulated as a permutationinvariant set-to-vector mapping task, and a DeepSets-based neural architecture is proposed in which interferers are explicitly regarded as elements of an unordered set. The resulting BF policy is designed for integration within artificial intelligence (AI)-native radio access network (RAN) architectures, where inference is executed as a network application (NwApp) in the near realtime RAN intelligent controller (Near-RT RIC). The proposed approach is evaluated using physically meaningful BF metrics, including desired-direction array gain, interference leakage, and proxy signal-to-interference-plus-noise ratio (pSINR). Results indicate that the suggested design achieves performance similar to a conventional feedforward neural network (FFNN) baseline while preserving structural integrity despite variations in interferer input ordering. These findings highlight the importance of incorporating permutation-invariant inductive biases when designing learning-based BF policies for dynamic RAN environments.