A paper posted to arXiv describes CLEAR (Controllable and Legible End-to-end Architecture for Recommendation), a system its authors say is deployed at Meta. It shows users a ranked set of interests the system has inferred, written in natural language, and stores the interests users add or suppress as authoritative preference state. Each inferred interest is tied back to events in the user's own history.
According to the paper, each interest then drives retrieval, ranking, attribution and exposure control. A newly declared interest can enter before any behavior supports it, drawing candidates from topical retrieval, and a delivery controller tops up interests that are underexposed. The authors say CLEAR components have operated in production for over one year across four surfaces on Instagram and Facebook.
In direct user review, users rated about 88.5% of displayed interests accurate, the paper reports. In an offline benchmark, the authors' Specialized 3B generator scored 0.3967 mean-user F1@10, against 0.3633 for Gemini 2.5 Pro and 0.1911 for the prior generator. A randomized ablation measured same-day delivery losses for a new interest of 8.7 to 13.7 percentage points when individual components were removed.