The architecture uses separate LSTM encoders for each modality, then fuses their outputs through a fully connected network for binary classification. Result: 54.03% test accuracy (marginally above ...