Meta gives Facebook users direct control over Feed and Reels recommendations

A single place to manage recommendations

Meta describes Your algorithm as a suite of AI-powered tools designed to make Facebook recommendations more transparent and steerable. The system combines a user-facing page with recommendation infrastructure that stores preferences and applies them during ranking and delivery.

The page begins with a one-sentence “Lately Summary” describing recent viewing activity. It also includes an “Interested In” list, prefilled with interests inferred from past activity, and a “Not Interested In” list containing topics that a person has rejected, hidden, or marked as not interesting.

People can add interests in their own words, move them between the two lists, or remove them. According to Meta, each change affects Feed and Reels responsively, with further updates based on feedback on individual posts.

How preferences reach the recommendation funnel

Meta says traditional recommendation systems rely heavily on implicit engagement signals. That approach can scale, but the company says it adapts slowly and captures less nuance than direct input from users.

For an interest added to “Interested In,” Your algorithm uses four retrieval sources. Two search for matching content directly, while two extend discovery through the content graph and item-to-item similarity. Relevant candidates receive a promotion during late-stage ranking so they can compete with other recommendations.

The system is designed to balance visibility with variety. A new interest should appear without taking over the feed, according to Meta. A control layer spaces out recommendations for the same interest and limits each interest’s daily contribution. New interests receive extra weight for a limited period before the system learns from how people engage with the resulting content.

“Not Interested In” preferences are enforced at several stages. Embedding-based comparisons can remove close matches during ranking, demotion rules push down near matches, and an LLM reranker reviews the final list. Requests involving quality, such as reducing AI-generated or sexually suggestive content, use dedicated demotion signals. A creator named in the list is filtered directly.

LLMs manage preferences and speed up changes

Large language models are used throughout the preference lifecycle. They generate the initial interest profile and Lately Summary, validate edits, identify duplicates, map phrases to clean topic names, and perform a final check against “Not Interested In” interests.

Meta says a smaller model handles validation efficiently at scale. The approved preference is enriched with an embedding and search keywords before being saved to a shared store used by both Feed and Reels. Separate integrity and safety checks also form part of the process.

The company has added delivery mechanisms for changes that might otherwise be delayed by recommendation and device caches. Adding an interest triggers a supplementary tail load, an extra fetch request outside the normal cycle. Removing an interest can invalidate cached content by obtaining topic clusters and dropping matching candidates before display.

For users, the system offers a more direct way to shape Facebook recommendations than relying only on viewing and engagement behavior. Meta’s description also shows how the company is combining editable controls, ranking logic, device-side handling, and task-specific LLMs to make those preferences affect both discovery and filtering.