Abstract
Large language models (LLMs) possess significant potential for customized AI; yet, the successful integration of user history for tailored answers continues to pose challenges. Current methodologies frequently transform user timelines into extensive textual descriptions, resulting in significant computing expenses and the risk of losing subtle information. Motivated by the effective amalgamation of LLMs with other modalities, including pictures, we provide USER-LLM, an innovative framework that considers user timelines as a separate modality and utilizes user embeddings for optimal contextualization of LLMs. User embeddings, produced by a pretrained user encoder, encapsulate latent user behaviors and interests derived from various interaction data. By incorporating these embeddings with LLMs via cross-attention, USER-LLM allows LLMs to dynamically tailor their responses to specific user histories.
Our assessment across three distinct datasets (MovieLens, Amazon Review, and Google Local Review) reveals that User-LLM attains a significant decrease in computation (up to 78.1X) relative to text-prompt-based approaches, while maintaining performance integrity. User-LLM consistently sustains or enhances performance on tasks necessitating profound user comprehension, especially with extensive user histories, underscoring its efficacy in adeptly acquiring and using user knowledge for tailored answers.
Publisher: ACM Digital Library [Access the full article there.]
Author: Lin Ning, Luyang Liu, Jiaxing Wu, Neo Wu, Devora Berlowitz, Sushant Prakash, Bradley Green, Shawn O’Banion, Jun Xie | Google DeepMind, United States of America.
Citation: Ning, L., Liu, L., Wu, J., Wu, N., Berlowitz, D., Prakash, S., Green, B., O’Banion, S., & Xie, J. (2025). User-LLM: Efficient LLM Contextualization with User Embeddings. Companion Proceedings of the ACM on Web Conference 2025, 1219–1223. https://doi.org/10.1145/3701716.3715463