Meta Introduces Your Algorithm for Personalized Facebook Feed Controls
New AI-powered tools allow users to directly influence Facebook content recommendations using natural language.
1New User-Centric Recommendation Interface
Meta has launched Your Algorithm, a suite of tools designed to make Facebook recommendations more transparent and steerable. Users now have a single editable view of their interests that shape their Feed and Reels. This system moves beyond traditional implicit engagement signals, allowing individuals to interact directly with the ranking algorithms.
2Managing Preferences with Natural Language
The interface features a Lately Summary of recent viewing activity alongside two lists: Interested In and Not Interested In. Users can add interests using natural language, move topics between lists, or remove them entirely. Every change applies responsively to both Feed and Reels, ensuring that content updates reflect these explicit user preferences.
3System Architecture and Preference Lifecycle
The system architecture covers the full lifecycle of a preference, from initial inference to final delivery. Profile generation creates an interest profile, while preference management validates changes and saves them to a shared store. Ranking and delivery optimization components then apply these preferences throughout the content funnel to ensure consistent, personalized results.
4Balancing New Interests and Content Discovery
When a user adds a new interest, the system triggers four dedicated retrieval sources to find relevant content. To prevent flooding, a control layer spaces out recommendations and caps daily contributions for specific interests. This ensures that new topics appear in the feed without completely taking over the user's viewing experience.
5Layered Filtering for Negative Feedback
Interests marked as Not Interested In are enforced through multiple stages to ensure high precision. During ranking, embedding-based comparisons remove candidates, while demotion rules push down near matches. An LLM reranker reviews the final list, and client-side cache invalidation removes matching content that was already queued for display before the update.
6Optimizing Responsiveness for User Changes
To close the gap between setting a preference and seeing results, Meta added two delivery mechanisms. A supplementary tail load fetches fresh recommendations immediately when a new interest is added. Simultaneously, cache invalidation drops previously ranked content that matches a newly added negative preference, ensuring that changes take effect almost instantly.
7Role of Large Language Models
Large language models (LLMs) manage the preference lifecycle by generating initial profiles and validating user edits. A smaller, efficient model maps phrases to clean topic names, while a fine-tuned LLM reranker performs a final check against negative interests. This approach is faster and more reliable than using general-purpose models for every item.
8Evaluating LLM Performance and Accuracy
Evaluating LLM outputs requires a specialized approach, as standard offline metrics cannot judge the accuracy of summaries or reranking decisions. Meta uses a separate LLM as a judge to grade output against expected behavior. This provides an objective, repeatable method for comparing prompts and models before they are deployed to users.
9Future Directions for Personalization
While Your Algorithm enhances transparency, Meta continues to explore ways to interpret ambiguous user input more effectively. Future developments aim to make recommendations respond with even greater precision to specific requests. The company is also investigating additional entry points to allow users to shape their content preferences from more locations across Facebook.