Replace high-dimensional weights with a compact trainable latent vector — 200×–500× fewer trainable params ...