Meta has published a technical account of 'Your algorithm,' a Facebook feature that it says lets people describe, in their own words, what they want to see in Feed and Reels. The company frames it as a response to recommendation systems that learn from implicit engagement, which it says adapt slowly and capture little nuance. [1]
According to the post, the page shows a one-sentence 'Lately Summary' of recent viewing, plus two editable lists. 'Interested In' is prefilled with interests inferred from past activity. 'Not Interested In' collects topics a user has marked, hidden or X-ed out. Meta says users can add, move or remove entries, and that changes apply to both Feed and Reels. [1]
For added interests, Meta says each one drives four dedicated retrieval sources. Candidates that pass a relevance check get a promotion in late-stage ranking. A control layer spaces out content on the same interest and caps each interest's daily contribution, so a new interest 'should show up, not take over,' in the company's words. [1]
Exclusions are enforced at several stages, Meta says. Embedding-based matching and demotion rules act during ranking, a fine-tuned LLM reranker reviews the final list, and on-device cache invalidation drops already queued content that matches. Requests about quality, such as seeing less AI-generated or sexually suggestive content, map to dedicated demotion signals. Requests naming a creator filter that creator's content directly. [1]
Meta says LLMs appear at several points. One generates the initial interest profile and summary. A smaller model validates and normalises each edit, alongside a separate integrity and safety component. The company reports that a small model fine-tuned to return only items to remove was faster and more reliable than a general-purpose model judging items one by one. It says it evaluated these calls with a separate LLM acting as a judge. [1]
The post gives no usage figures, rollout details or measured outcomes. Meta acknowledges remaining work on interpreting ambiguous input and making recommendations respond more precisely. None of the effectiveness claims have been independently verified. [1]