As foundation models grow stronger, the traditional framing of parameter-efficient fine-tuning as a mere cost-cutting measure begins to feel limiting—recent work has shown that pretraining and fine-tuning improvements decouple across different capability dimensions, and researchers have documented methods for learning individual preferences efficiently. This paper invites a different mental model: what if small, persistent adapters are not just cheaper approximations but a fundamental substrate for multiplying personal models atop shared competence? The tension becomes architectural rather than purely empirical—if agents can adapt continuously through episodic reasoning without weight updates, then where does the boundary between stateful adapters, memory systems, and fine-tuned parameters actually lie, and what does persistence mean when you can revise, version, and mix these components at scale?
Abstract
Parameter-efficient fine-tuning (PEFT) is usually treated as a cheaper alternative to full fine-tuning. We study a broader role: small trainable adapters as persistent local state on top of strong shared foundation models. In this framing, the base model provides shared competence while adapters carry instance-specific behavior such as preferences, skills, tool habits, and memory-like updates. We organize the problem around three scaling axes: Scale Up, where stronger shared priors make small local updates more useful; Scale Down, where we study how small adapters can be while remaining reliable; and Scale Out, where many persistent adapted instances coexist. MinT provides one infrastructure example for managing adapter identity, revision, provenance, evaluation, and serving residency. Together, the results suggest that PEFT can be a compact substrate for persistent personal models rather than only a budget substitute for full fine-tuning.