Personalized large language models increasingly rely on profile-conditioned inference: they pair a public request with rich, user-specific facts to tailor the output. The paper argues that this paradigm exposes the user profile itself, along with the user's private data, during inference.
To address this, the authors introduce Profile-DP, a differential privacy framework designed for user-controlled profile release. The approach aims to let users decide how much of their profile can be shared when a personalized model answers a request.
The significance is in shifting control back to the user: instead of the profile being an implicit part of the inference pipeline, its release becomes a privacy decision. Since this is a single source, there are no conflicting findings to compare; the contribution is the proposed framework itself.