Deep In-Context Learning (ICL) for Wireless Communications

Abstract

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 non-orthogonal 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 adopts an unfolded iterative receiver design, where a Transformer-based backbone with Virtual Width Networks (VWN), Tensor Product Attention (TPA), and Mixture-of-Experts (MoE) layers performs joint representation learning for CE and detection. A differentiable Physics-Guided Feature Construction (PGFC) bridge converts intermediate outputs into soft symbol estimates and explicit physical features, which are then fed into subsequent stages for refinement. Additionally, a novel resampling strategy is adopted during inference to improve detection reliability without requiring any network retraining.

The proposed architecture is evaluated through link-level simulations adopting standard 3rd Generation Partnership Project (3GPP) fading channels (e.g., EPA and ETU) across various MIMO-OFDM configurations and 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.

Popular Abstract

Modern wireless communication systems need to transmit more and more data through limited radio spectrum. One important challenge is that a wireless receiver must first understand how the radio channel has changed the transmitted signal before it can recover the data. To do this, current systems send known reference signals, often called pilots. These pilots help the receiver estimate the channel, but they also consume time-frequency resources that could otherwise be used to transmit user data.

This thesis studies how artificial intelligence can help reduce this pilot overhead in multiple-antenna wireless systems. Instead of treating channel estimation and data detection as two separate steps, the proposed receiver learns to perform them jointly. The pilot signals are used as contextual information, allowing the neural network to infer the channel condition and recover the transmitted data within a unified model.

The work investigates several pilot designs. In current 5G systems, pilot signals from different antennas are kept separated to avoid interference. This is reliable, but it requires more pilot resources. In the proposed in-context learning based scheme, different antennas are allowed to transmit pilots over the same resource elements. Although this creates interference, the AI receiver can learn to interpret the overlapping pilots as useful context. The results show that this can reduce the need for strictly orthogonal pilot patterns while maintaining good detection performance.

The thesis also studies superimposed pilot transmission, where pilot and data symbols are placed on the same resource elements. This removes dedicated pilot overhead, but it also makes the detection problem harder because pilots and data interfere with each other. The results show that the proposed receiver can mitigate part of this interference. In favorable signal-to-noise and coding conditions, the saved pilot resources can lead to higher throughput.

Overall, this thesis shows that combining physical communication knowledge with deep learning can make wireless receivers more flexible. The proposed receiver does not simply replace traditional signal processing with a black-box model; instead, it uses the structure of the wireless system to guide the learning process. This provides a possible direction for future wireless systems that need to use spectrum more efficiently while supporting increasingly complex transmission schemes.

• author: Fu, Zhongwang
• supervisor: Juan Vidal Alegria; Sha Hu; Jonathan Lindberg
• organization: Department of Electrical and Information Technology
• course: EITM02 20261
• year: 2026
• type: H2 - Master's Degree (Two Years)
• subject: Technology and Engineering
• keywords: In-Context Learning, AI Receiver, MIMO-OFDM, Channel Estimation, MIMO Detection, Demodulation Reference Signal, Superimposed Pilots.
• report number: LU/LTH-EIT 2026-1147
• language: English
• id: 9235508
• date added to LUP: 2026-06-15 13:05:09
• date last changed: 2026-06-15 13:05:09