Proprietary large language models (LLMs) excel across many tasks, yet aligning their outputs with diverse user preferences remains challenging. Existing black‑box LLM personalization methods typically attach a user‑specific scoring head, causing the number of personalized parameters to grow linearly with the user count and requiring extra adaptation for unseen users. To overcome these issues, we introduce Mixture‑of‑Facets (MoF), a scalable personalization framework. MoF models user preferences as compositions of shared latent preference facets rather than dedicated per‑user parameters. Personalization is performed via history‑conditioned routing over shared facet heads, enabling adaptation to users not seen during training without adding new parameters. Across a range of personalization tasks, MoF achieves stronger personalization performance while being more scalable and parameter‑efficient than prior approaches. Additional analysis shows strong generalization to unseen users.
Review