Separable Expert Architecture: Toward Privacy-Preserving LLM Personalization via

On April 25th, a research team quietly posted a preprint that attempts to solve one of the quietest crises in AI deployment: the fact that personalising a language model and protecting user privacy are, under most current architectures, fundamentally in tension. The paper describes what they call the Separable Expert Architecture. It has two interlocking mechanisms. The first is composable adapters — modular components that steer a model's outputs toward a specific user's preferences without touching the base model underneath, designed to be stacked and swapped like interchangeable lenses on a fixed camera body. The second is deletable user proxies — lightweight representations of individual users that can be surgically removed from the system when a user requests it, without degrading the model's performance for everyone else. The architecture's central claim is that personalisation can be structurally separated from the model's core weights, which means user data never needs to be baked permanently into a shared system. You don't have to choose between a model that knows you and a model that can forget you.

If confirmed, here is what this means. Right now, the dominant approach to LLM personalisation quietly violates what most users believe about their data — that deleting their account deletes their influence on the model. It doesn't, in any meaningful technical sense. A working implementation of this architecture would give that deletion button actual teeth. For enterprises operating under GDPR's right to erasure or CCPA's deletion provisions, it would represent a genuine compliance path rather than a legal fiction. For AI product teams, the composable adapter design offers something equally practical: the ability to maintain one clean base model while serving thousands of individualised experiences, reducing both compute overhead and the risk of catastrophic forgetting when user preferences conflict. The second-order effect is more quietly significant — if users actually trust that their data can be removed, they may share more freely in the first place, which improves model quality across the board. The privacy mechanism becomes a data flywheel in disguise.

Watch for independent replication attempts, particularly from teams with access to production-scale personalisation infrastructure, and for any response from major AI labs whose current architectures this work implicitly challenges.

NewsHive monitors these sources continuously. All signal titles above link to the original reporting.

Intelligence by NewsHive. Need help navigating what this means for your business? Contact GeekyBee →