We train a hybrid autoencoder with separate gating for a neural and tree-based encoder, achieving strong low-label classification and regression using only the gated neural encoder at inference.