Tiered Language Model framework locks private AI capabilities behind a secret key

Researchers have proposed the Tiered Language Model (TLM) framework, which splits a single neural network into public and private branches using a compact secret key that reroutes computation through a hidden sub-graph. Unlike existing approaches that either prune capabilities or restrict access via closed APIs, TLM allows one weight file to serve multiple configurations without altering underlying parameters. In experiments on 180M- and 650M-parameter models, the keyed configuration achieved perfect recall of private facts while the public version retained none of that information. The security mechanism operates on roughly 5% of the model's parameters, making it resistant to fine-tuning-based extraction, though a full key leak would expose the private branch entirely. Scaling to billion-parameter models remains unproven, but if successful, the approach could let companies release open-weight models while protecting proprietary or sensitive capabilities behind a cryptographic token.