Master thesis presentation: Deep In-Context Learning (ICL) for Wireless Communications

The thesis was carried out at Huawei Technologies with Sha Hu and Jonathan Lindberg as industrial supervisors, and Juan Alegria Vidal as academic supervisor. Michael Letmaier is the examiner.

In multiple-input multiple-output orthogonal frequency-division multiplexing (MIMO-OFDM) systems, the spectral overhead required by orthogonal Demodulation Reference Signals (DMRS) limits overall system capacity. Traditional linear estimators experience severe degradation under high pilot sparsity or nonorthogonal superposition. To address these limitations, this thesis proposes a Physics-Guided In-Context Learning (ICL) receiver that jointly performs channel estimation (CE) and MIMO detection. By formulating reference signals as contextual prompts through a tokenization scheme, the proposed artificial intelligence (AI) receiver circumvents the strict requirement of classical spatial orthogonality.

The architecture employs a deep unfolded iterative cascade, coupling explicit communication-theoretic models with a Transformer-based backbone incorporating Virtual Width Networks (VWN), Tensor Product Attention (TPA), and Mixture-of-Experts (MoE) layers. A differentiable Physics-Guided Feature Construction (PGFC) bridge computes localized physical residuals to refine latent representations across stages, facilitating progressive interference cancellation (IC). Additionally, a novel resampling strategy is adopted during inference to improve detection reliability without requiring any network retraining.

The proposed architecture is evaluated using a 3GPP 5G-NR link simulator under 2x2 and 4x4 MIMO-OFDM configurations across different modulations. Our results indicate that the proposed architecture approaches Maximum Likelihood Detection (MLD) in baseline configurations. The receiver maintains detection capability under sparse pilot conditions where standard 5G-NR baselines become intractable, and it mitigates pilot-data self-interference in superimposed DMRS scenarios. By achieving reliable detection with reduced pilot overhead, the framework translates DMRS overhead savings into throughput improvements, presenting a scalable AI-receiver architecture for spectrally efficient future wireless networks.