Meta Introduces Your Algorithm for Personalized Facebook Feed Controls
New AI-powered tools allow users to directly influence Facebook Feed and Reels recommendations.
1New User-Driven Personalization
Meta has launched Your Algorithm, a suite of tools designed to make Facebook recommendations more transparent. Users can now explicitly state their preferences in natural language to shape their Feed and Reels content. This system moves beyond traditional implicit engagement signals, allowing for more direct and immediate control over the content displayed.
2Interface and Preference Management
The interface features a Lately Summary of viewing activity alongside two lists: Interested In and Not Interested In. Users can manually add, move, or remove topics within these lists. Every modification is applied responsively across both Feed and Reels, ensuring that user feedback directly influences the ranking and delivery of future content.
3System Architecture Overview
The system architecture manages the full lifecycle of a preference, from initial inference to final delivery. Key components include profile generation, preference validation, and ranking optimization. By utilizing a shared preference store, the system ensures that the interests displayed on the user interface are consistently applied throughout the content ranking funnel.
4Retrieval and Content Balance
To integrate new interests without overwhelming the user, the system employs four dedicated retrieval sources. These include direct search, topic clusters, and content graph analysis. A control layer in the reranking stage spaces out recommendations to ensure balance, capping the daily contribution of any single interest to maintain a diverse feed experience.
5Layered Filtering for Precision
The Not Interested In list is enforced through multi-stage filtering to ensure high precision. During ranking, embedding-based comparisons remove irrelevant candidates, while demotion rules push down near-matches. Additionally, an LLM reranker reviews the final list to catch any content that might have bypassed earlier filters, providing a robust layer of content control.
6Optimizing Delivery Responsiveness
Meta implemented two delivery mechanisms to ensure changes take effect immediately. A supplementary tail load fetches fresh recommendations outside the normal cycle when a new interest is added. Conversely, cache invalidation removes previously ranked content from the device that matches a newly added Not Interested In topic, preventing unwanted items from appearing.
7Role of Large Language Models
LLMs are integrated across the preference lifecycle to generate profiles, validate user edits, and perform final checks. A smaller, efficient model handles real-time preference validation, while a fine-tuned LLM reranker reviews the final ranked list. This approach ensures that the system remains both scalable and highly accurate in interpreting user intent.
8Evaluating Model Performance
Evaluating LLM outputs requires a specialized approach since standard offline metrics are insufficient. Meta uses a separate LLM as a judge to grade the output of each call against expected behavior. This objective, repeatable method allows engineers to compare different prompts and models effectively before deploying them to the live production environment.
9Future Development Opportunities
While the current system significantly improves transparency, Meta continues to explore ways to interpret ambiguous user input more effectively. Future iterations aim to increase the precision of recommendations and expand the number of entry points for preference settings. These ongoing efforts seek to provide users with even greater agency over their digital experience.