Reproduce “Toward Open Weight Models Without Risks: Separating Public and Private Capabilities in LLMs” on OpenResearch

About this paper

Charbel El Feghali, Arkil Patel, Nicholas Meade, Spandana Gella, Verna Dankers, Siva Reddy

Mila and McGill University researchers developed Tiered Language Models (TLMs), a framework that enables a single open-weight LLM to offer distinct public and private capabilities by using a compact permutation key to reconfigure a small subset of its parameters. This approach allows authorized users to unlock private functionalities without needing additional learned parameters or compromising the public model's behavior, demonstrating strong separation of capabilities and robustness against adversarial attacks.

Key insights

• Capability separation can be achieved by encoding authorization directly into the model's parameter configuration through permutations, moving beyond input-space authentication or distributed learned components.
• An asymmetric gradient flow during pretraining allows the public configuration to remain functional even when a subset of parameters is primarily shaped by the private configuration's objectives.
• A compact permutation key provides a more robust and efficient access control mechanism compared to learned adapter weights or input-based 'passwords,' offering resistance to common adversarial extraction techniques.

Summary by alphaXiv.