Back to feed

Your Algorithm: Giving People More Ways to Personalize Facebook Recommendations

Engineering at Meta

Nov 10, 2026

11/10/2026

Decomposed Task-Specific LLMs With Independent Safety And Evaluation Improve Reliability Over A Single General Model

Your Algorithm: Giving People More Ways to Personalize Facebook Recommendations · Engineering at Meta

Science, Technology & Innovation · Nov 10, 2026

Meta finds that narrowly scoped, fine-tuned LLMs—paired with embeddings, independent safety checks, and task-specific judges—can make preference systems faster, more reliable, and easier to evaluate than a single general-purpose model.


11/10/2026

Canonical Preference State Across Surfaces Enables Free-Form User Intent To Influence Recommendations

Your Algorithm: Giving People More Ways to Personalize Facebook Recommendations · Engineering at Meta

Science, Technology & Innovation · Nov 10, 2026

Meta’s “Your algorithm” lets users express editable natural-language preferences that are validated, enriched, centrally stored, and applied consistently by Facebook Feed and Reels ranking systems, overcoming the slow and limited nuance of implicit engagement signals.


11/10/2026

Responsiveness Depends On Delivery, Cache Coherence, And Client-Side Invalidation, Not Just Ranking Models

Your Algorithm: Giving People More Ways to Personalize Facebook Recommendations · Engineering at Meta

Science, Technology & Innovation · Nov 10, 2026

Meta handles immediate personalization through delivery scheduling and client-cache invalidation as well as ranking updates: new interests trigger supplementary fetches, while disinterests invalidate cached recommendations so changes appear promptly.


11/10/2026

Controllable Recommendations Require Asymmetric Multi-Stage Controls For Positive And Negative Preferences Across Retrieval Ranking And Delivery

Your Algorithm: Giving People More Ways to Personalize Facebook Recommendations · Engineering at Meta

Science, Technology & Innovation · Nov 10, 2026

Meta treats positive and negative recommendation controls differently: “Interested In” expands candidate discovery with bounded exposure, while “Not Interested In” uses layered filtering and demotion to maximize exclusion recall across the entire delivery pipeline.