A new research paper proposes a way to make AI chatbots remember your preferences without stacking up a mountain of per-user parameters.
Researchers describe Mixture-of-Facets (MoF), a framework for personalizing black-box large language models, the kind you access through an API rather than open weights. Today's typical approach bolts a scoring head onto the model for each user, so the system grows bigger and slower as the user base grows, and it still needs retraining to handle anyone new. MoF instead breaks preferences down into a shared set of latent "facets" and routes each user's history through them, so no new parameters get added when a new user shows up. The researchers report it outperforms prior methods on personalization tasks while staying leaner, and it generalizes well to users the model never saw during training.
Personalization is one of the harder unsolved problems in applied AI: a chatbot that remembers you want short answers feels different from one that treats everyone the same. Most existing fixes trade accuracy for scale, since they personalize well for known users but add overhead as the user base grows or shifts. If this shared-facet approach holds up outside the paper's own benchmarks, it points to a more realistic path for companies running LLMs through an API, where touching the model's weights isn't an option.
The paper's own benchmarks are the only test so far; real-world preference drift and adversarial users are harder exams still to come.