The concept of shared latent-space adapters originates from the bottleneck adapter tradition in transfer learning, featuring very few parameters and low training cost.