WiFo-MUD: Wireless Foundation Model for Heterogeneous Multi-User Demodulator

Multi-user signal demodulation is critical to wireless communications, directly impacting transmission reliability and efficiency. However, existing demodulators underperform in generic multi-user environments: classical demodulators struggle to balance accuracy and complexity, while deep learning-based methods lack adaptability under heterogeneous configurations. Although diffusion models have been introduced for demodulation, their flexibility remains limited for practical use. To address these issues, this work proposes WiFo-MUD, a universal diffusion-based foundation model for multi-user demodulation. The model aligns inter-user signal-to-noise ratio imbalance and performs conditional denoising via a customized backbone. Furthermore, a communication-aware consistency distillation method and a dynamic user-grouping strategy are devised to enhance inference. WiFo-MUD achieves state-of-the-art results on large-scale heterogeneous datasets, demonstrating efficient inference and strong generalization across varying system configurations.

💡 Research Summary

The paper introduces WiFo‑MUD, a diffusion‑based wireless foundation model designed to tackle the long‑standing trade‑off between accuracy and computational complexity in multi‑user MIMO demodulation, while also addressing the poor generalization of existing methods across heterogeneous system configurations. Traditional linear detectors (LS, LMMSE) are computationally cheap but sub‑optimal, and nonlinear approaches (sphere decoding, OAMP) achieve near‑optimal performance at prohibitive cost. Recent deep‑learning solutions (RE‑MIMO, OAMP‑Net) improve the balance but still falter when antenna counts, modulation orders, or user numbers change, especially under realistic inter‑user interference (IUI) and SNR imbalance. Diffusion models have shown promise for modeling complex channel and noise distributions, yet prior works ignore multi‑user dynamics and require many iterative refinement steps, leading to high latency.

WiFo‑MUD consists of three main components. First, a low‑complexity coarse estimator (e.g., LMMSE) provides an initial symbol estimate for each user. Second, a lightweight Multi‑User Aligner (MU‑Aligner) computes an equivalent noise power for each coarse estimate, maps it to a diffusion time‑step, and performs user‑specific noise suppression. This step aligns the SNRs across users, mitigating the detrimental effect of severe SNR disparity before feeding the data to the main network. Third, the aligned estimates are processed by a Wireless Diffusion Transformer (WiDiT), a conditional denoising diffusion model built on a Transformer backbone. WiDiT incorporates channel state information, user identifiers, and complex‑valued symbol embeddings, allowing the self‑attention mechanism to capture global inter‑user dependencies and channel correlations.

To overcome the iterative nature of diffusion sampling, the authors propose a Communication‑aware Consistency Distillation (CCD) scheme. During training, a multi‑step “teacher” model generates high‑quality outputs, while a “student” model learns to produce comparable results in a single diffusion step by minimizing a consistency loss that aligns their predictions. This dramatically reduces inference latency without sacrificing accuracy. Additionally, a dynamic user‑grouping strategy clusters users with similar SNR and channel characteristics, applying group‑wise denoising to further suppress IUI and enable scalable operation when the number of active users varies.

The authors construct a large‑scale heterogeneous dataset covering a wide range of antenna configurations (from 8×8 to 64×64), modulation schemes (QPSK to 64‑QAM), user counts (2–16), carrier frequencies, and channel models (Rayleigh, Rician, correlated fading). Experiments evaluate both full‑shot (training and testing on the same distribution) and zero‑shot (testing on unseen configurations) scenarios. WiFo‑MUD consistently outperforms classical ML detectors, sphere decoding, OAMP‑Net, RE‑MIMO, and recent diffusion‑based demodulators (ALD, predictor‑corrector) by 1.5–3 dB in SNR for the same bit‑error‑rate, and shows only marginal performance degradation (<2 %) in zero‑shot tests where competing methods suffer >10 % degradation. After CCD, single‑step inference achieves more than a five‑fold reduction in latency and a 30 % reduction in memory footprint, making the approach viable for real‑time hardware implementation.

In summary, WiFo‑MUD leverages the expressive power of diffusion models and the global modeling capability of Transformers, augments them with communication‑specific preprocessing (MU‑Aligner) and a novel consistency‑distillation technique, and demonstrates strong, scalable performance across a broad spectrum of multi‑user wireless scenarios. The work opens avenues for extending foundation‑model concepts to multi‑cell, RIS‑assisted, and mmWave systems, as well as for online continual learning to adapt to rapidly changing channel conditions.