HybridNEFT: An Asymmetric CNN-Transformer Architecture for Efficient Near-Field CSI Feedback in 6G XL-MIMO Systems

Extremely large-scale multiple-input multiple-output (XL-MIMO) systems operate in the near-field regime due to their massive antenna arrays, creating substantial challenges for accurate channel state information (CSI) feedback. While deep learning has advanced CSI feedback by enabling compact channel representation, current approaches remain inadequate for modeling the intricate structure of near-field CSI and impose computational costs that exceed the capabilities of practical user equipment. To address these challenges, we propose HybridNEFT, a near-field CSI feedback framework that integrates a lightweight attention-free convolutional encoder with a hierarchical global-attention Transformer decoder, which leverages progressive token reduction and global attention to achieve accurate reconstruction under substantially reduced computational overhead. Extensive simulations show that HybridNEFT achieves an 8-16 dB improvement in normalized mean-squared error over state-of-the-art methods and maintains over 99% cosine similarity across all compression ratios. Moreover, it requires far fewer total FLOPs than near-field SOTA models, while keeping encoder-side complexity comparable to lightweight CNN-based baselines. These results demonstrate that HybridNEFT provides a highly favorable accuracy-complexity trade-off and is well suited for practical near-field CSI feedback in 6G XL-MIMO systems.