On the Scaling of PEFT: Towards Million Personal Models of Trillion Parameters

The paper reframes parameter-efficient fine-tuning (PEFT) as infrastructure for personalized models, not just cheap fine-tuning.

It organizes the scaling challenge into three parts: scaling up shared priors, scaling down adapter size, and scaling out many adapted instances.

PEFT is presented as persistent local state for many custom models built on top of large foundation models.

MinT is highlighted as one example of a management layer for handling this kind of scalable personalization.