Problem
Recommenders that learn only from implicit engagement adapt slowly and capture little nuance. A user who says "more of X" or "no more Y" sees no immediate effect, because the model reacts only to implicit signals. Three things go wrong. A new interest may never enter the candidate set. A negative topic is fuzzy, so any single filter misses matches. Content already ranked and cached on the device ignores the change.
Example from the article: on Facebook, a person adds an interest in plain words and expects Feed and Reels to change right away.
Solution in brief
Keep each stated interest in one preference store that Feed and Reels both read, and validate every edit with a small LLM. Wanted interests get dedicated retrieval, a relevance check and pacing. Rejected interests are filtered at several stages, ending with a fine-tuned LLM reranker. Delivery is forced: an extra fetch for new interests, cache invalidation for rejected ones.
Questions this case answers
- How to turn free-text user preferences into ranking signals that act at every stage of a multi-stage recommendation funnel
- How to make an explicit preference change take effect immediately when ranked candidates are already cached on the client
- How to enforce a negative preference with high precision and recall when similarity filters alone miss near matches
- How to let users steer a recommendation feed with preferences written in natural language?
Tools and methods
- Preference store — one shared record of both lists, used by ranking and the settings page
- Small LLM validator — checks the phrase is a coherent topic, finds duplicates, normalizes the name
- Embedding similarity — relevance check for promotion, close-match removal for rejected topics
- Fine-tuned LLM reranker — last check of the ranked list against rejected interests
- LLM-as-a-judge — offline grading of LLM outputs before launch
Google or a one-time email code. No password.