Tiered LLM Design Hides Private Skills Behind a Secret Key

Researchers propose Tiered Language Models (TLM) that split a neural network into public and private branches controlled by a secret key. The key permutes a small fraction of parameters to route computation, enabling private capabilities while public behavior remains unchanged. In tests with 180M- and 650M-parameter models, the private branch achieved perfect recall of private facts while the public side stayed at zero.