The MoE architecture typically introduces auxiliary loss functions to force the gating network to distribute requests evenly, preventing routing to only a small number of popular experts.