The Durable Asset Is the Data, Not the Adapter: A Position on Portable Personalization for an Era of Commoditized, On-Device Models
Parameter-efficient fine-tuning made model customization cheap, and the surrounding tooling is now commodity. But a customization compiled against one base model does not survive a change of base. As foundation models commoditize and migrate to the edge, base models will turn over frequently and silently, and every turnover orphans the fine-tuning built on top of it. This position paper argues that the adapter is being managed as the durable asset when it is in fact a disposable, base-coupled compilation artifact. It makes three claims: that low-rank adapter weights are non-transferable across independently trained bases for a fundamental reason (the absence of a shared coordinate frame); that the durable, portable asset is the personalization source — a versioned preference store and a frozen per-user evaluation set — from which an adapter can be recompiled against any base; and that the resulting platform primitive (automatic carry-forward of personalization across base-model upgrades, proven against each user's own acceptance test, executed inside the user's trust boundary) is unbuilt today. A falsifiable experimental protocol is specified for the central recompile-versus-transport claim.
Authors
- Colby Philbin
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-05-27
- DOI
- https://doi.org/10.5281/zenodo.20412406
- Primary Topic
- Data Visualization and Analytics
- Type
- preprint