Can user embeddings personalize language models more efficiently than prompts?

The standard way to personalize an LLM is to stuff a user's history into the prompt as text — but interaction data is long, noisy, and expensive to tokenize. User-LLM distills user embeddings from diverse interactions (search, media consumption, navigation, social activity) via self-supervised pretraining, capturing latent preferences and their evolution over time, then integrates them into the LLM through cross-attention and soft-prompting (with Perceiver layers to streamline). Across MovieLens, Amazon Review, and Google Local Review, it outperforms text-prompt-based contextualization especially on long-sequence tasks and tasks requiring deep user understanding, while being more computationally efficient and preserving the LLM's general knowledge.

The keeper is the representational choice: compress a user's behavioral history into a learned embedding the model attends to, rather than serializing it as prompt text — which both scales to long histories and captures temporal preference drift that flat prompt-stuffing loses.

This sits in the vault's personalization thread as the embedding-based contextualization route. It contrasts with Do user outputs outperform inputs for LLM personalization? (which keeps personalization in text but selects the right text) — User-LLM moves it out of text entirely — and connects to the broader question of whether user context belongs in tokens or in learned representations.