What happened
In an Engineering at Meta article posted on November 10, 2026, eight authors describe the system behind “Your algorithm,” a Facebook feature that lets people steer recommendations using their own words. Meta says recommendation systems have traditionally learned from implicit engagement signals: a scalable approach, but one that can adapt slowly and miss nuance. The new system adds explicit preferences alongside behavior, with a single editable view of interests that shape both Feed and Reels.
The interface has three parts. A “Lately Summary” describes recent viewing activity. “Interested In” contains topics inferred from past activity, and “Not Interested In” includes topics users have explicitly rejected, hidden or marked as unwanted. People can add interests in natural language, move them between lists or remove them. Meta says changes apply responsively to Feed and Reels and continue to update as people give feedback on individual posts.
Behind the page, profile generation creates an initial interest profile; preference management validates edits, enriches them with signals needed for ranking and stores them; ranking applies the two lists through the recommendation funnel; and delivery optimization helps changes take effect quickly. Both Feed and Reels read from the same preference store, so Meta says the preferences visible to users are the ones used in ranking.
For a new “Interested In” topic, four retrieval sources seek candidates: direct search, clusters of related topics, the content graph and item-to-item similarity. A relevance check compares each candidate with the interest, and qualifying candidates receive a promotion in late-stage ranking. A separate control layer spaces out recommendations for that interest and caps its daily contribution so it appears without taking over. Each request retrieves candidates for only one interest, giving interests turns; new interests get extra weight for a limited period before the system learns from engagement.
“Not Interested In” is applied through several layers. Embedding comparisons remove close matches during ranking, demotion rules push near matches down, and an LLM reranker reviews the final list. On-device cache invalidation removes matching items that were queued before the preference changed. Meta says non-topic requests, such as reducing AI-generated or sexually suggestive content, map to dedicated demotion signals; a named creator’s content is filtered directly. To speed up changes, a new positive preference triggers an extra fetch outside the normal cycle, while a negative preference prompts the client to obtain topic clusters and drop matching cached content.
LLMs appear at several points. They generate readable inferred interests and the Lately Summary, validate edits by checking topic coherence and duplicates, and provide the final check against unwanted content. Meta says the edit-validation model is smaller and efficient to serve at scale. The final reranker is fine-tuned to return only items to remove; Meta reports this was faster and more reliable than asking a general-purpose model to judge each item individually. A separate LLM judge grades outputs against expected behavior to compare prompts and models before launch.
What the archive adds
New Runtime’s earlier coverage on October 17, 2026, introduced the same “Your algorithm” feature as an editable space in Facebook’s post menu for managing Feed and Reels topics. That earlier note covered the user-facing control; Meta’s engineering account adds implementation detail about retrieval, ranking, filtering and delivery. This is a new fact on a known story, not a separate product announcement.
Trend signal
This case does not yet fit a pattern New Runtime tracks. Meta describes a substantial system for translating explicit preference into ranked recommendations, but the article does not establish a broader shift in agent capabilities, interoperability or work organization.
Why it matters
The engineering account makes clear that a preference control is not just a prompt field attached to a recommender. A positive preference can fail if the relevant items never enter the candidate pool; a negative preference can fail if the system filters only exact matches or leaves already-ranked items in a device cache. Meta’s design spreads the work across retrieval, ranking, reranking and delivery, with distinct mechanisms for each failure mode.
For teams building recommendation products, this is a useful architectural checklist: translate user intent into a normalized representation, use it both to retrieve and rank candidates, protect overall balance, and make downstream caches respond to edits. The separation of a smaller validation model from a fine-tuned removal model also illustrates why one general-purpose model call may not be the best fit for every stage. These are design choices, not evidence that Meta’s system improves satisfaction or recommendation quality; the article provides no outcome figures.
Limits and uncertainty
Meta does not report user counts, measured changes in recommendation relevance, latency, cost, error rates or how often the system misinterprets a preference. The article describes system behavior and design choices, but does not independently establish that users’ stated preferences reliably produce the intended experience. It also does not quantify how daily caps and temporary extra weighting are set, or how the system handles conflicts between explicit preferences and engagement signals. Meta says it is still working on ambiguous input and more precise responses, and is exploring additional entry points for shaping preferences.
Skill to build now
If you are designing preference controls for a recommendation product, map a user edit across the whole delivery path this week: candidate retrieval, ranking, final filtering and any cached results. Test a new positive preference that is absent from ordinary retrieval, and a negative preference applied after results have already been queued. The cases Meta describes show why testing only the ranking model can miss failures in the rest of the system.
Reader takeaway
The main lesson is architectural: a natural-language preference becomes a real product control only when every stage that selects and delivers recommendations can act on it. Meta’s account shows one way to coordinate that change across retrieval, ranking, filtering and caches, but leaves the user-level effectiveness of the approach unmeasured.