Meta Engineering Blog Details "Your Algorithm": How LLMs of Different Sizes Collaborate to Turn a User's Sentence into Recommendation Results
Meta has publicly disclosed the technical details of its "Your algorithm" system in an engineering blog. This tool allows users to tell Facebook what they want to see and what they do not want to see using natural language; the system then translates these preferences into actual recommendation changes through different stages and large language models (LLMs) of varying scales. A key finding is that, in the filtering stage, a fine-tuned small model responsible solely for "returning items to be deleted" is faster and more reliable than a general-purpose model evaluating each item individually.
The engineering team at the US technology company Meta published a technical blog post on November 10, 2026, detailing the internal architecture of the "Your algorithm" system. This tool was previously launched to users in October, allowing people to write in natural language what they want or do not want to see in their Facebook Feed and Reels. The blog's focus is not on feature introduction, but rather on explaining an engineering challenge: how to transform a casual sentence spoken by a user into a sorting signal that actually takes effect within the recommendation system.
The Meta team points out that traditional recommendation systems rely on implicit behavioral signals (such as clicks and dwell time). While these can run at scale, they adapt slowly and fail to capture subtle intentions. "Your algorithm" attempts to fill this gap: when a user directly states "I want to see more content about urban gardening" or "I do not want to see AI-generated images anymore," the system must immediately understand this statement, store it, and execute it in subsequent recommendations. This presents three specific challenges—new interests must appear quickly without overwhelming other content; exclusion items must filter reliably even when topics are vague; and previously cached old recommendations must be updated.
On the "Interested In" side, each interest added by a user drives four independent retrieval paths: two direct matches (via search and via topic clustering) and two extended coverage paths (via the content graph and item-to-item similarity). The purpose is to ensure that content for brand-new interests does not completely miss the recall stage. Candidates that pass the relevance check receive a boost in the late-stage ranking, after which the control mechanism of the reranking layer determines where they appear. Meta emphasizes "balance over quantity": the control layer spaces out recommendations for the same interest and sets a daily cap for each interest; during the retrieval stage, each request fetches candidates for only one interest, allowing all interests to take turns getting opportunities.
On the "Not Interested In" side, filtering is distributed across multiple stages. During the ranking phase, embedding comparisons identical to those used in the relevance check are employed to remove candidates that are too close to the excluded interest, while demotion rules push down near matches. After ranking, an LLM reranker reviews the final list. On the device end, cache invalidation mechanisms discard content that has been queued for display but matches new exclusions. Notably, not all exclusions are topic-based—requests to reduce AI-generated content or sexually suggestive content map to dedicated demotion signals rather than topic matching; if a user specifically names a creator, that creator's content is filtered out directly.
LLMs appear in three different positions within this system, each undertaking different tasks and corresponding to models of different scales. Upon a user's first visit, an LLM is responsible for generating the initial profile: it converts continuously refreshed interest signals into readable topic names, populates the "Interested In" list, and writes a "Lately" summary. Whenever a user edits their preferences, another smaller LLM handles validation—checking whether the phrase constitutes a coherent topic, detecting duplicates, and mapping them to standardized names. Meta explains that this step uses a small model to serve efficiently at scale. After ranking is completed, a fine-tuned LLM reranker performs a final check.
Meta shared a key engineering discovery in the blog: in the reranking filtering stage, a small model fine-tuned to "only return items to be deleted" is both faster and more reliable than having a general-purpose model judge each candidate individually for relevance. This conclusion directly influenced model selection—the approach is not to use a larger general-purpose model for more granular individual reasoning, but rather to use a smaller specialized model for batch exclusion.
How to evaluate these LLM calls themselves is also a question. Meta notes that standard offline metrics cannot determine whether a "Lately" summary is accurate, nor can they judge whether the reranker removed the correct content. Their solution involves using another independent LLM as a judge to score the output of each call against expected behaviors, thereby providing a repeatable comparative baseline before launch.
Meta acknowledges unresolved issues at the end of the article: the ability to interpret ambiguous inputs, making recommendations respond more precisely to each specific request, and collecting user preferences from more entry points. The value of this blog lies in demonstrating a complete engineering pipeline from natural language to sorting signals, and highlighting the differentiated requirements for model scale, output format, and evaluation methods across different tasks within this pipeline.