The development

What happened

Meta has described how Your algorithm uses large language models to turn people’s stated interests into Facebook Feed and Reels recommendations. Its 10 November engineering account explains a shared preference profile, task-specific models and changes to retrieval, ranking and cached content; it does not establish a new worldwide rollout.

Key facts

  • People can edit Interested In and Not Interested In lists. An LLM generates initial interests and a recent-activity summary; smaller models validate and normalize edits.

  • Positive interests receive dedicated retrieval and ranking promotion, with spacing and daily caps to keep one topic from dominating. Preferences feed both Feed and Reels.

  • Negative topics pass through embedding filters, demotion and a fine-tuned LLM reranker. Separate signals handle content-quality requests, while named creators can be filtered directly.

  • Extra content fetches and device-cache invalidation address recommendations queued before an edit. Meta gives no measured response-time or perfect-filtering guarantee.

  • Meta uses another LLM to compare outputs offline. The article supplies no human-validated accuracy result or access-by-region timetable, and says ambiguous requests remain an improvement area.

The context

Why it matters

TideReach assessment: explicit preferences affect more than a chatbot interface: they must alter candidate retrieval, ranking and delivery to change what a person actually sees. Meta’s account makes those mechanisms clearer, but its own evaluation description does not demonstrate that unwanted topics are always removed or that recommendations match every user’s intent.