Transformer-based soft-output MIMO detection for coded systems

Deep learning (DL)–based multiple-input multiple-output (MIMO) detection has emerged as a promising technique for future wireless communication systems. In this paper, we propose a transformer-based DL framework for MIMO detection, specifically designed for coded MIMO systems with soft iterative decoding. To address the prohibitive computational cost associated with collecting soft-information training labels, a computationally efficient recursive label generation method is employed. The proposed framework demonstrates strong generalization capability, as it is trained at a single signal-to-noise ratio (SNR) and evaluated over a wide range of SNR values. In addition, its generalization capability and scalability are validated across various antenna configurations and modulation orders. Simulation results show that the proposed method significantly improves bit error rate (BER) performance, achieving near–maximum-likelihood (ML) performance for coded MIMO systems while maintaining excellent generalization performance.