AirMoE cuts wireless bottlenecks with statistic-augmented over-the-air aggregation

Deploying Mixture of Experts (MoE) over wireless cloud-edge networks faces two coupled bottlenecks: routing activations overloads limited uplinks due to raw feature transmission, and aggregating expert outputs is hindered by channel noise. Researchers from Tsinghua University propose AirMoE (arXiv:2607.16562) to break these limits. On the routing side, each client queries a local Feature Retrieval Library (FRL) with a compact broadcast query, retrieves a prototype-induced statistic, and reports it digitally — slashing uplink traffic. The cloud then selects the most relevant clients by aligning these statistics with LM-extracted features using Jensen-Shannon divergence. On the aggregation side, selected experts simultaneously transmit outputs over a multiple-access channel, physically computing the reweighted sum via waveform superposition. The reweighting coefficients are realized through channel-aware power control, decoupling routing and aggregation both algorithmically and physically.

AirMoE provides theoretical analyses on convergence and iteration complexity. In semantic segmentation tasks, extensive experiments show AirMoE outperforms standard MoE baselines and even single large models, while ablation studies confirm each component's effectiveness. This paradigm could enable efficient large-scale AI collaboration across bandwidth-constrained edge devices, reducing latency and communication overhead.