Learning to Refine LLRs: Modular Neural Augmentation for MIMO-OFDM Receivers

The growing demands for higher throughput and cost-efficient operation in wireless communications drive the need for robust multi-user MIMO-OFDM receivers that can cope with hardware impairments and non-linear environments. While classical model-based receivers and recently proposed deep neural network (DNN) architectures provide complementary benefits, they either rely on simplified linear Gaussian assumptions, require excessive training complexity, or struggle to address cross-subband interference. In this work, we propose a compact and modular DNN augmentation that refines the soft outputs of existing receivers (model-based or data-driven), thereby enhancing their resilience to non-linearities and inter-carrier interference. Our design leverages an element-wise scaled convolutional neural network tailored to perform learned interference cancellation across users and neighboring subcarriers, combined with a training algorithm that encourages accurate log-likelihood ratios for soft channel decoding. Numerical results demonstrate that the proposed augmentation consistently improves diverse receiver algorithms in challenging channel conditions while incurring minimal overhead, highlighting its potential as a practical enabler of flexible AI-aided receivers.