LLM agents let users control personalization across all platforms

A team of 18 researchers from multiple institutions, led by Jiacheng Lin, published a paper arguing for a paradigm shift in personalization: from platform-centric to user-governed. Currently, platforms like Google, Amazon, and Netflix build user profiles from the behavioral fragments they observe, but no platform can construct a complete picture due to competitive incentives, legal constraints, and privacy concerns. The paper posits that only users themselves can aggregate their cross-platform and offline information, and large language model (LLM) agents make this integration practically feasible for the first time. The agents can reason over heterogeneous personal data—such as email histories, social media activity, and browsing logs—to transform fragmented contexts into actionable personalization capabilities.

The researchers provide proof-of-concept evidence: users who export their data from multiple platforms and use an off-the-shelf LLM agent achieve better personalization results than any single-platform baseline. For example, the agent can recommend products, articles, or music by understanding a user's full context across work, social, and entertainment apps. The paper concludes by outlining a research agenda for building scalable user-governed personalization systems, addressing challenges like data privacy, agent reliability, and interoperability. This work could fundamentally change how personalization happens, putting control back in the hands of users and reducing reliance on walled gardens.