Facebook

Facebook's source situation is the opposite of X's. There is no public code. Instead, the Transparency Center's AI system cards (Meta's per-surface documents explaining ranking) list prediction head names and individual feature sentences verbatim. A prediction head is the prediction output a model produces separately for each action. Earnings call (IR) transcripts give the online results of ranking changes in numbers every quarter. Regulatory filings (the EU Digital Services Act, DSA) and patents contain parts that are more specific than the blog posts.

The model and infrastructure lineage (DLRM → DHEN → Wukong → HSTU → GEM) is not specific to Facebook, so it moved to topics/. This file covers only Facebook as a surface.

Summary

The latest public canon for Facebook Feed ranking is the 2021 structure: a friend-graph lookup, candidate reduction by a lightweight model in pass 0, a multi-task linear weighted sum in pass 1, and contextual diversity in pass 2 verified (inventory, 3-pass structure). On top of that, the share of unconnected (recommended) content grew from 11.7% in 2021 to 41.0% in 2025 Q4 verified (surface composition measurement). With this increase, the ranking problem shifted from a social graph problem to a pure recommendation problem.

↑ Contents

Validity period of system card quotes — a caution for the whole repository

The Transparency Center system card quotes in this repository are the 2026-10-06 version (rendered 2026-10-07) for fb-feed, and the other cards were read directly on 2026-08-09 observed 2026-08-09, 2026-10-07. Other files only link to this section. Here, the whole repository means services/ and topics/.

• Re-verification requires browser rendering. The card pages are client-rendered, so curl returns no body, and they intermittently block automated fetches.
  • On the 2026-08-09 retry, curl got HTTP 400 and WebFetch got HTTP 429.
• Meta revises cards without a changelog. Only the updated date changes, and nothing records what changed.
• Therefore every claim based on a card is a snapshot as of the date it was read. When quoting, give both the card's updated date and this repository's read date.
  • The 10 prediction heads, the candidate counts per stage, the 50 signals, and the absence counts inside the card fall under this.

↑ Contents

Feed surfaces

Facebook's connected and unconnected content are handled by two separate AI systems, and the final feed is the result of inserting ads into the outputs of the two systems verified Transparency Center — Our approach to explaining ranking (2023-12-31).

"a balanced combination of the outputs of both AI systems ... in addition to advertisements and additional product offerings such as groups and reels"

⚠ This sentence is not in the body of the Nick Clegg 2023 post. We retrieved the full text of that post on 2026-08-09 and checked. That post must not be used as the source of this quote observed 2026-08-09. The corpus checked is 1 item: the body at that URL.

Meta publishes a separate system card for each surface, and this is the primary source for the list of surfaces verified Our approach to explaining ranking, 2023 onward.

• Facebook Feed covers only connected content from friends, followed Pages, and joined groups. The card states this scope itself verified.

  "Your feed might also include suggested content and advertisements. That content is not powered by the AI we describe in this system card"

• Facebook Feed Recommendations covers unconnected content. It has its own card, and its set of prediction heads is separate too.
• Facebook Reels is a surface where content is consumed one item at a time on a vertical screen.
• In 2024, Facebook Video unified Reels, long-form video, and Live into a single full-screen player verified Zuckerberg Q2 2024 IR.

  "This quarter we rolled out our full-screen video player and unified video recommendation service across Facebook -- bringing Reels, longer videos, and Live into a single experience."

• Facebook Feed Ranked Comments is a separate card covering comment ranking.
• The Friends tab launched on 2025-03-27 and shows only friends' posts, without recommendations inferred. Zuckerberg's own post is the primary source, but we could not obtain the original URL and confirmed it only through news reports.

Surface composition measurement — Widely Viewed Content Report

As of 2025 Q4, 41.0% of US organic Feed views are unconnected (recommended) content, and friends' content is 20.1% verified. The Widely Viewed Content Report (WVCR) is a disclosure of a regulatory nature, not marketing copy. The table below shows the share of US organic Feed views as of 2025 Q4 verified WVCR.

• Views that include an external link are 1.5% verified.

  "98.5% ... did not include a link to a source outside of Facebook"

• The unconnected share grew from 11.7% in 2021 Q3, through 15.2% in 2022 Q3, to 41.0% in 2025 Q4 verified.
• As of 2025 Q4, "news from friends" is now 1/5 of Feed views. This one table explains where the whole Meta ranking literature focuses.

AI recommendation share — a metric Meta stopped talking about

In 2023 and Q1 2024, Meta gave the share of feed content recommended by AI as a percentage verified. The Q2 2024–Q1 2026 earnings call transcripts do not mention this percentage metric again, and the performance metric changed to time-spent increments observed 2026-08-09.

• The 2023-06-29 post gave the AI recommendation share of Facebook and Instagram feeds verified — ai.meta.com.

  "more than 20 percent of content in a person's Facebook and Instagram feeds is now recommended by AI"

• Zuckerberg said in Q1 2024 that the AI recommendation system delivers about 30% of Facebook feed posts verified.

  "about 30% of the posts on Facebook feed are delivered by our AI recommendation system. That's up 2x over the last couple of years. And for the first time ever, more than 50% of the content people see on Instagram is now AI recommended."

• This percentage metric has not been announced again since Q1 2024 observed 2026-08-09. All later metrics were replaced by time-spent increments.
  • The corpus checked is all 8 earnings call transcripts from Q2 2024 to Q1 2026, and the percentage wording appears 0 times in them.
• So the figure circulating in 2026, along the lines of "30% of the Facebook feed is AI-recommended," is 2 years old. The WVCR's 41.0% (2025 Q4) is more recent.
  • The two figures also differ in definition. WVCR is based on views, while the IR figure is based on deployment. They must not be mixed.

↑ Contents

Candidate generation (retrieval)

Connected content has no ANN (approximate nearest neighbor search) inferred. The corpus is already restricted to the user's connections, so embedding search is not needed inferred. All embedding retrieval is concentrated on the unconnected side inferred. The evidence is that the inventory definition below and the pass 0 description consist only of a graph lookup and candidate reduction by a lightweight model.

Inventory

The inventory of Facebook Feed is the non-deleted posts shared by friends, groups, and Pages connected to the user, limited to those posted since the last login verified How machine learning powers Facebook's News Feed ranking algorithm, 2021. The definition explicitly states the time boundary since the last login.

"any non-deleted post shared with Juan by a friend, Group, or Page that he is connected to that was made since his last login"

The same post gives the scale as follows verified.

"more than 2 billion people (more than 1,000 posts per user, per day, on average)"

Candidate counts per stage — the system card describes them differently for each surface

The candidate counts per stage are written differently on each surface's card, and the Feed Recommendations card, which covers unconnected content, has no figures verified each system card.

• ⚠ Feed does not shrink from 500 → 500, while Video narrows in 2 stages from 1,000 → 200. Meta's documents themselves describe a different funnel shape for each surface.
• Reels' 10–100 is an order of magnitude smaller than Feed's 500. This appears to be because content is consumed one item at a time by vertical swiping, so the slate (the bundle of items shown together at once) itself is small inferred. Meta did not state a reason.
• The 2021 blog post describes pass 0 as the stage where a lightweight model picks about 500 of the most relevant posts eligible for ranking verified. This lets later passes rank fewer stories with high recall verified.

  "a lightweight model is run to select approximately 500 of the most relevant posts for Juan that are eligible for ranking"

  "This helps us rank fewer stories with high recall in later passes"

Unconnected content retrieval

Retrieval for unconnected content narrows billions of items to thousands and then to a few hundred in hundredths of a second verified — ai.meta.com, 2023.

"retrieval systems that take just hundredths of a second to narrow billions of pieces of content down to thousands and then to a few hundred"

• The same post contains the following expression: "a novel hierarchical deep neural retrieval architecture". This one line is the start of a lineage that runs through Andromeda in 2024 to HILL/MoNN in 2026. Details are in topics/embedding-retrieval.md.
• ⚠ The card and the blog post conflict. The unconnected system card gives no candidate counts at all, while the blog post describing the same system writes "billions → thousands → a few hundred". The two documents describe it at different levels of detail.

We could not find a primary source for the per-source breakdown of candidate counts in Facebook Feed retrieval (no primary source found) inferred. The corpus checked is 15 Transparency Center system cards, ranking-related posts on engineering.fb.com, ai.meta.com, and about.fb.com, and earnings call transcripts. The ranking-related posts are in the sources list. This means there is no breakdown within this corpus, not evidence that undisclosed documents lack one. This contrasts with X, whose candidate sources can be enumerated from code (x.md candidate generation section).

↑ Contents

Ranking

According to the 2021 engineering blog post, a Facebook Feed score is the sum of per-action predictions multiplied by personalized weights, and the weights are set based on surveys verified (the only published formula, weights section). The 2026-10-06 updated Facebook Feed card lists 10 prediction heads and 50 signals verified (prediction heads section). A separate page, Our Approach to Facebook Feed Ranking (updated 2025-06-11), splits predictions into three tiers of usage frequency and publishes only this frequency hierarchy instead of weights verified (same section). As of Q3 2025 remarks, Facebook, Instagram, and ads recommendation run on three separate Transformers, and merging them into one is an ongoing goal verified (unified model section).

The only published formula

The 2021 engineering blog post contains the following two formulas verbatim verified engineering.fb.com, 2021. The first line computes per-action predictions from features, and the second takes the weighted sum of those predictions to produce the score.

Yijt = f(xijt1; xijt2; … xijtC)
Vijt = wijt1·Yijt1 + wijt2·Yijt2 + … + wijtk·Yijtk

• The subscript i is the post, j is the viewer (user), t is the point in time, and k is the prediction task verified. The original says: "for each post i, we estimate Yijt", "the characteristics of the post X_it toward viewer j at time t".
  • Because the weight w is indexed by (i,j,t), it is personalized per post, per user, and per point in time.
• The text states directly why it chose a linear combination verified same post. Actions people rarely take get low predictions and so automatically play only a small role in ranking.

  "Any action a person rarely engages in (for instance, a like prediction that's very close to 0) automatically gets a minimal role in ranking, as Yijtk for that event is very low."

• The formula has no numeric weights at all. It is written only in symbols verified same post.

How the weight w is set = surveys

The weight w is set based on the actions that people say in surveys are more meaningful and worth spending time on verified 2021 blog post.

"the way we take each prediction into account for Juan is based on the actions that people tell us (via surveys) are more meaningful and worth their time"

"People with higher correlation gain more value from that specific event, as long as we make this method incremental and control for potential confounding variables."

The survey signals themselves are a cross-service technique, so they moved to topics/survey-signals.md.

3-pass structure

As of the 2021 blog post, Facebook Feed ranking consists of three passes verified. A lightweight model reduces the candidates, a multi-task neural network scores each post, and finally contextual features such as diversity rules are added verified. Documents differ on which stage handles integrity processing.

1. In pass 0, a lightweight model picks about 500 from the inventory. The goal is high recall verified.
2. Pass 1 is the main scoring stage that scores each post independently, and it uses a multi-task neural network verified. It is pointwise (scoring posts one by one separately).

   "the main scoring pass, where each story is scored independently"

   "we score each post using multitask neural nets"

3. Pass 2 is the contextual stage. Here, contextual features such as content-type diversity rules are added to help diversify the feed verified.

   "the contextual pass. Here, contextual features, such as content-type diversity rules, are added to help diversify Juan's News Feed."

• The 2021 blog post says integrity processing is applied to every post at a position before ranking verified: "certain integrity processes are applied to every post".
• ⚠ The stage at which integrity is applied differs by document.
  • The 2021 blog post says before ranking verified.
  • The fb-video card says after ranking verified: "Certain integrity processes are applied to all results".
  • The fb-feed card (2026-10-06 updated version) mentions integrity in two places observed 2026-10-07. We confirmed this by rendering the English card in headless Chrome.
    • According to the "Gather inventory" stage, posts flagged for violations are already dropped at the inventory collection stage.

      "gathers all potential posts shared by friends, Pages you follow and groups you've joined, which excludes posts flagged for going against our Community Standards."

    • In the "Leverage signals" stage, distribution-reducing integrity is in the same paragraph as pass 0. Pass 0 is the stage where a lightweight model picks about 500.

      "A lightweight model is run to select approximately 500 of the most relevant posts. The system also applies certain integrity processes to help reduce the distribution of problematic content."

  • So we read fb-feed as "removal-type at the inventory, score-reduction-type near pass 0" inferred. The evidence is that the card puts the two sentences under different stage titles. However, the card does not say whether score-reduction-type integrity enters the pass 1 score.
  • In the version read on 2026-08-09 (updated 2026-06-24), we recorded only one integrity sentence. We cannot determine whether the exclusion sentence at the inventory stage was newly added in the 10-06 update or was missed then. We did not keep the original text of the earlier version to compare, and Wayback did not respond on 2026-10-07.
• We could not find a primary source for the multi-task neural network structure of pass 1 (no primary source found) inferred. We cannot tell whether it is in the MMoE/PLE family or how many towers it has. The corpus checked is the Meta blog posts in the sources list and the 15 system cards.

Prediction heads — Facebook Feed system card

The Facebook Feed card lists 10 prediction heads, with 5 signals attached to each head, 50 in all verified fb-feed (last updated 2026-10-06, previous version 2026-06-24). The number and order of heads are the same as in the previous version observed 2026-10-07 rendering. Rendered in order, the gist of the card text is:

1. Whether the user scrolls past
2. Whether the user scrolls past, reflecting format preference. A separate model handles the same task.
3. Whether the user skips a Story In Feed (SIF)
4. Whether the user skips the video channel
5. 'Interested' click
   • Card text: "How likely you are to click 'Interested' if the interested/not interested options are presented to you."
   • The signals are the list of authors and posts on which the user pressed Interested in the last 180 days, the current time, the number of times the user watched for 30 seconds or more in the last 30 days, and the video length.
   • This is the current name of 'Show more'. The rename is in the 2025-03-11 entry of the user controls table.
6. Entering Story full screen
7. Clicking the like/reaction list
8. Deep/strong intent toward a connection's post
9. Clicking an external link and dwell time on that site. Even dwell outside Facebook is a prediction target.
10. Strong intent to interact with a connection

Negative predictions explicitly exist. The Transparency Center text states directly that predictions of hide/snooze/unsubscribe, report, and angry reaction are used in the direction of "lowering distribution" verified Our Approach to Facebook Feed Ranking (last updated 2025-06-11).

The same page splits predictions into 3 tiers of usage frequency: Used most frequently / occasionally / less frequently. This is Meta's counterpart to X's ModelWeights list, and it is a form that publishes only the frequency tiers instead of weights verified.

The unconnected card's prediction heads are different

The Feed Recommendations card, which covers unconnected content, uses prediction heads different from the Facebook Feed card verified fb-feed-recommendations. The heads and signals include screenshots, private shares on other platforms, and Messenger and WhatsApp share history verified.

The prediction heads are:

• Dwell time prediction
• share
• 'X' click
• Dwell time increase
• Group join
• like
• Private share on other platforms
• Screenshot
• RSVP click
• 'Show more'

Among the signals, the following stand out:

• Screenshots are an explicit positive verified same card.

  "List of the owners the user has screenshot their post in the last 180 days"

• Cross-app signals are included verified same card.

  "How many times a user has shared posts on Messenger in the past 1 month"

  "Owners whose post the user has shared on whatsapp in the last 180 days"

• There are location signals verified same card.

  "Your approximate distance from the event in a post"

Actual feature names (fb-feed card excerpts, all verified)

The fb-feed card publishes ranking features as sentences verified fb-feed card. They include the post's position, the number of times the user saw the same type of content, predictions based on friends' reactions, and time windows of several lengths verified.

• The position at which the post sits in the feed is a feature verified: "At which position is the post located on your News Feed?"
• The number of times the user saw posts of the same content type in the last hour enters as a feature verified. Diversity is a re-ranking rule and also a ranking feature.

  "Number of times you have seen posts with the same content type (e.g. video, photo, etc) as the current post in the past hour"

• The number of times the user quickly scrolled past the same friend's post is a feature verified: "Number of times you quickly scroll past the same friend's post in the past 7 days"
• There is a feature that is a weighted count of the number of times a post was viewed verified: "A weighted count of the number of times a post has been viewed in 2 hours."
• There is a feature that quickly predicts the probability of interaction based on how much friends reacted to similar content verified. This means a structure that feeds model output back in as a feature, that is, a cascade.

  "Quickly predicts how likely you are to interact with a post ... based on how much your friends have engaged with similar content"

• The lookback windows (the past period a feature aggregates) are tiered as 1 hour / 2 hours / 1 day / 7 days / 14 days / 28 days / 30 days / 180 days inferred. The evidence is the window lengths that repeatedly appear in the 2026-10-06 version's feature list.
  • The "3 months" window recorded in the version read on 2026-08-09 is not in the current version observed 2026-10-07.
• ⚠ In the version read on 2026-08-09, the following sentence was the card's only embedding feature: "An embedding to represent the viewer". The 2026-10-06 version does not have this sentence observed 2026-10-07.
  • "embedding" appears 0 times in the full text rendered that day. Most of the 50 signals in the current version are aggregate counts, ratios, and time-window features, and the post-position feature above and the friend-reaction-based prediction feature are also in this version observed 2026-10-19 re-render.
  • We infer it was dropped in the 10-06 update inferred. This is the only time the card's date has changed. However, since we did not keep the earlier version's original text, we could not compare directly.
  • A sentence being dropped from the card does not mean the model stopped using a viewer embedding. The card lists only some of the "significant predictions".
• The "How to customize" section of the 2026-10-06 version has an Instagram app explanation paragraph mixed in observed 2026-10-07.

  "You can click the Instagram logo on your home screen and then click on either "Following" or "Favorites" on mobile app. ... you won't see recommended content from instagram."

  • This appears to be an editing error in which Instagram card text was copied into the Facebook card inferred. Facebook's "See newest content first" item follows separately right after.

Online results from IR (all verified, regulatory disclosure)

In all four quarterly earnings releases of 2025, the company gave Facebook time spent or view growth as numbers verified.

Because freshness (the same-day share) was reported as a separate metric for 3 consecutive quarters, we infer that recency was promoted to an explicit ranking objective inferred.

ConnectionMind — an LLM-based ranking layer that reasons over graph paths (2026-08 paper, production A/B)

ConnectionMind is a graph-path reasoner attached at the candidate scoring and re-ranking role verified §6.1. For the top 5–10% of heavy users, an LLM reasons over paths online; the other 90–95% are served by a Student GNN trained on meta-paths (path patterns linking node types) that the same LLM mined offline verified §6.1. The paper of 2026-08-10 reports this structure and the production A/B results verified ConnectionMind: Leveraging Social Networks and Large Language Models for Personalized Recommendation at Meta (arXiv 2608.10187).

⚠ This is a candidate to change an existing claim. It is the first first-party statement from Meta that an LLM enters the online serving path of organic feed ranking. Details are covered in the case section of topics/llm-in-recsys.md; here only the surface and pipeline position are noted briefly.

• The paper has 6 authors. Haoyu Han is a Michigan State intern, and the other 5 are affiliated with Meta Platforms (Menlo Park) verified. The affiliation footnotes give no Meta internal team name (Core Ranking / GenAI, etc.).
• The paper does not name the surface.
  • §3 writes: "large-scale personalized video recommendation in a social platform setting".
  • §6 writes only this: "short-form video platform".
• It is most likely Facebook (Reels or the unified Video surface) inferred. There are three reasons.
  • The graph schema is User/Page/Item with User–Group, Group–Item, Page–Item and Item–Item (Co-watch), so Groups and Pages are 1st-class relations. Instagram and Threads have no Groups or Pages schema.
  • The main metrics are watch time, video sessions and exposure, so it is a video feed.
  • Since Q2 2024, Facebook Reels, long-form video and Live have been bundled into a unified video recommendation service, and this unified surface takes graph signals as they are.
• However, the paper does not confirm the surface, so it must not be cited by a service name. It should be cited as "Meta short-form video platform".
• Serving is hybrid verified §6.1. For the top 5–10% of heavy users, Llama 3.1-8B-Instruct reasons over graph paths online. The other 90–95% are served by a Student GNN.
  • The Student GNN is trained on meta-paths that the same LLM mined offline verified §6.1.

    "For this cohort, requests are routed to the full ConnectionMind LLM"

    "we employ an offline distillation pipeline. The LLM processes historical contexts to discover effective meta-paths"

  • It is neither pure retrieval nor a pure ranker but a path reasoner attached at the candidate scoring and re-ranking role verified §6.1.
• The model is an LLM that outputs graph paths as JSON verified §4.1.
  • The state s_d is a prompt that serializes the current paths, candidate next-hop edges and short textual node profiles: "current paths, candidate next-hop edges, and compact textual node profiles".
  • The LLM outputs a structured action with two components, path expansions and item–path pairs: "structured action with two components: path expansions and surfaced item–path pairs".
  • Beam search and constrained decoding are not used explicitly; outputs are verified after the fact with a format reward (§4.3).
  • Training goes from SFT (BFS shortest-path teacher-forcing) to GRPO.
  • The GRPO stage uses a format reward of −1/+1, an F1-based recommendation reward and step-wise shaping, with weights α_rec=0.5·α_step=0.3·α_fmt=0.2.
• The online A/B ran for several weeks on tens of millions of users verified §6.2.
  • The relative lifts over production given by Table 2 are Watch Time +0.43% ± 0.14%, Exposure +0.33% ± 0.08% and Video Sessions +0.22% ± 0.13%.
• The offline benchmarks are Delicious and Foursquare, split 60/20/20 verified §5.1. They are public benchmarks from the social recommendation literature, not Meta data.
  • The baselines are only social-family models such as GraphRec, DiffNet++, RecDIFF and BIGRec; there is no LightGCN, PinSage, SASRec or HSTU.
• There are 0 citations of HSTU, GEM or LLaTTE verified reference list.
  • Meta internal systems (HSTU, GEM, MetaCLIP, LLaTTE, MARM) do not appear in the references.
  • Of the literature by Meta-affiliated authors, it cites only Fan et al. 2019 (GraphRec).
• The paper does not state how it relates to the company's other ranking lines. It may be a parallel experiment by a separate team or may stem from internal politics, but there is no evidence to decide.
• Undisclosed items were moved to open questions. The items that are not public are below.
  • Graph size (number of nodes and edges)
  • Training data scale
  • Hot-path latency, QPS and GPU cost of the 8B LLM
  • A/B guardrails (ads revenue, integrity, negative effects)
  • Per-cohort/country breakdown
  • Details of the base GNN architecture

⚠ Meta's own documents need to be checked for consistency. None of the fb-feed card, fb-video card or fb-reels card lists a "graph-path reasoning LLM" as an explicit prediction head or processing stage observed 2026-10-19. On that day the three cards were re-rendered in headless Chrome; the editions are fb-feed 2026-10-06, fb-video 2026-06-29 and fb-reels 2026-06-19.

Unified model — not yet

Per the Q3 2025 statement, Facebook, Instagram and ads recommendation run on three separate Transformers, and merging them into one is an ongoing goal verified. No timeline was given. This is widely misunderstood, so it is stated explicitly.

• Zuckerberg said in Q3 2025 that three giant Transformers run Facebook, Instagram and ads recommendation verified. The three are separate, and unifying them is an ongoing goal. No timeline was given.

  "there are three giant transformers that run Facebook, Instagram, and ads recommendations. ... we're also working on combining these three major AI systems into a single unified AI system"

• The origin of this goal is a Q2 2024 statement verified.

  "I'd like to see us move towards a single, unified recommendation system ... We're not there"

• ⚠ This is the most confusing part. Susan Li said in Q4 2025 that per-surface model consolidation together with back-end improvements raised ad quality by 12% verified.

  "After seeing strong success with the consolidation of Facebook Feed and video models in the first half of 2025, in Q4 we consolidated models for Facebook Stories and other surfaces into the overall Facebook model. This, along with a series of back-end improvements, drove a 12% increase in ad quality."

  • The sentence mentions "Facebook Feed", but the outcome metric is ad quality. This is a consolidation of ads ranking models across surfaces, not a consolidation of organic Feed ranking.
  • The 12% is not the effect of model consolidation alone. The original text also cites back-end improvements as a cause.
• As of the Q4 2025 statement, sharing a platform between organic content and ads is still at the validation stage verified.

  "We're also going to start validating the use of ads signals and organic content recommendations as we continue to work towards having a more shared platform for organic and ads recommendations over time."

• The name of the organic-side research appears in the Q2 2025 statement verified: "cross-surface foundation recommendation models".

↑ Contents

Re-ranking, diversity, and filters

The re-ranking information Facebook has published is extremely thin compared with X: the algorithm name (MMR/DPP), the decay function and the threshold N are all missing. The difference in disclosure level is itself worth recording. Diversity is applied in a separate pass after pointwise scoring (2021 blog post), and the only concrete rule in the fb-feed card is to mix content types evenly verified. As of the page updated 2025-07-02, only 4 demotions remain in the list verified (Demotions section).

• Diversity is applied at pass 2, a separate pass after pointwise scoring verified 2021. It is structurally the same as X's post-selection placement inferred.
• The only concrete rule is to mix content types evenly. For example, it does not show several video posts in a row verified fb-feed card.

  "tries to ensure that your Feed has a balanced mix of content types. That means, for example, you wouldn't see multiple video posts in a row"

• A feature corresponding to this rule actually exists: "same content type ... in the past hour" in the actual feature names list.
• Listwise (looking at the whole item list at once) prediction exists verified ai.meta.com, 2023. There is also a personalized method for controlling delivery frequency to optimize long-term user value verified same post.

  "pointwise and listwise predictions"

  "personalized methodology for delivery frequency control to optimize for long-term user value"

• No primary source found for author-diversity rules, dedup rules or the consecutive limit N inferred. The corpus checked was 15 system cards and the engineering.fb.com/ai.meta.com ranking posts. Details are in topics/feed-diversity.md.

A detailed comparison is in topics/feed-diversity.md.

Demotions — largely deleted in 2025

Facebook first published its demotion criteria on 2021-09-23, and the current "Types of content we demote" page (last updated 2025-07-02) keeps only 4 verified. The cause is the policy change of 2025-01 verified. It is a textbook case of this repository's rule ("there is no guarantee that a 2019 paper is today's production").

• First published on 2021-09-23, grouped under 3 value criteria verified Sharing Our Content Distribution Guidelines.
  • The three criteria are direct feedback, incentives for high quality and accuracy, and a safe community.
• The count "about 24" is inferred. The announcement gives no count; only the 3 group names exist. Re-verified on 2026-08-09.
  • This figure circulates as counted from the CDG (Content Distribution Guidelines) list page at the time, and needs to be rechecked against an archive capture.
• The current "Types of content we demote" page (last updated 2025-07-02) keeps only 4 verified.
  • Clickbait Links
  • Engagement Bait
  • Fact-Checked Misinformation
  • Content Likely Violating Our Community Standards
• The individual pages and the index page returned 404 observed 2026-08-09. This observation is limited to 4 URLs requested directly.
  • The individual pages borderline-content/, low-quality-comments/, links-to-domains-and-pages-with-high-click-gap/, and the index content-distribution-guidelines/ itself returned 404.
• The earlier record that "engagement-bait/ and clickbait-links/ are alive" did not reproduce inferred. On 2026-08-09 both paths below returned 404.
  • /features/approach-to-ranking/types-of-content-we-demote/{engagement-bait,clickbait-links}/
  • /features/approach-to-ranking/{engagement-bait,clickbait-links}/
  • The two items appear to exist now only as sections in the body of types-of-content-we-demote/ (200).
• The cause is a policy change verified — More Speech, Fewer Mistakes. On 2025-01-07 Joel Kaplan wrote that too much content is demoted because systems predict it may violate standards, and that most of these demotions are being removed. He also said demotion of fact-checked content would stop.

  "we demote too much content that our systems predict might violate our standards. We are in the process of getting rid of most of these demotions."

  "stop demoting fact checked content"

• ⚠ Meta's own documents still disagree verified. The 4 pages updated 2025-07-02 and the 7 pages updated 2025-04-25 publish different lists at the same time.
  • The 4 pages still say they demote "Fact-Checked Misinformation", while Kaplan said that would stop.
• Disappearing from the documents does not mean disappearing from the system inferred. Low-quality comment demotion was dropped from the CDG, but it is still alive as a prediction task in the fb-feed-ranked-comments card (2026-04-02).

  "How likely a comment appears spammy, abusive, or policy-violating"

  "How likely the comment is high-quality and contributes positively to the conversation"

borderline content — curve and intervention

Zuckerberg's 2018 post wrote that the closer content gets to the line of what is allowed, the more people engage with it on average, and described an intervention to reduce the distribution of borderline content inferred. The current Transparency Center wording, as of the version updated 2025-04-25, is much weaker and does not mention the curve verified.

• The source is Zuckerberg's 2018-11-15 post "A Blueprint for Content Governance and Enforcement" inferred. The original URL (facebook.com/notes/…) is currently not accessible, and it was confirmed only through reproductions of contemporaneous quotations.
  • It wrote that, even if people later say they dislike the content, the closer it gets to the allowed line, the more it is engaged with on average.

    "no matter where we draw the lines for what is allowed, as a piece of content gets close to that line, people will engage with it more on average — even when they tell us afterwards they don't like the content"

  • The chart is named "Natural Engagement Pattern", and the intervention is to penalize borderline content to reduce its distribution: "penalizing borderline content so it gets less distribution".
  • It is suppression, not removal, and it flips the distribution curve so that it falls near the line.
• The current Transparency Center wording is much weaker and does not mention the curve verified (updated 2025-04-25).

  "If content ... doesn't violate the Community Standards, but might still be problematic or otherwise low-quality, Meta may reduce its distribution"

Click-Gap — mechanism clear, current status unknown

Click-Gap finds domains that get disproportionately many outbound clicks from Facebook relative to their place in the web graph verified Remove, Reduce, Inform, 2019. As of 2025, it is no longer in the public guidelines. Whether operation stopped is unknown, so it stays an open question.

• The original definition is below verified same post.

  "Click-Gap looks for domains with a disproportionate number of outbound Facebook clicks compared to their place in the web graph"

  • In the web graph, domains with many in/out links sit at the center and those with few sit at the periphery verified same post.
• The targets are domains that succeed on News Feed in a way that does not match the authority they built outside Facebook verified same post: "succeeding on News Feed in a way that doesn't reflect the authority they've built outside it".

clickbait / engagement bait — two definitions still in use

Clickbait Links and Engagement Bait are the items remaining on the demotion page updated 2025-07-02, and both have published definitions verified (Demotions section).

• Clickbait Links are links that induce clicks by creating false expectations verified: "lure people into clicking ... by creating misleading expectations".
  • There are 3 tactics.
    • withholding information
    • sensationalist phrasing ("You won't believe…")
    • punctuation (all caps, excessive exclamation marks)
• Engagement Bait is a post that explicitly asks for engagement such as votes, shares and comments for a purpose other than a specific call to action verified.

  "Posts that explicitly request engagement (such as votes, shares, comments, tags, likes, or other reactions) for purposes other than a specific call to action"

• The 2017 detection approach was to train ML on data from hundreds of thousands of posts that teams reviewed and categorized verified Fighting Engagement Bait, 2017.

  "Teams ... reviewed and categorized hundreds of thousands of posts"

  • It applied both demotion of individual posts and Page-level demotion of repeat publishers verified same post.

Political content — absorbing the suppression layer into ranking signals

On 2025-01-07 Meta said it would treat civic content from people and Pages you follow more like other content in the feed, and would rank it by explicit and implicit signals verified. It also said it would recommend more political content using these personalization signals.

"we're going to start treating civic content from people and Pages you follow on Facebook more like any other content in your feed, and we will start ranking ... based on explicit signals (for example, liking a piece of content) and implicit signals (like viewing posts)"

"we will start recommending more political content based on these personalized signals."

This is a shift from the 2021 "blanket suppression of civic content" to personalized ranking. The Threads-side timeline of the same shift is organized by date in services/threads.md.

↑ Contents

Ads blending

For ads blending, the patent is far more concrete than the official blog posts. The patent is the only mechanical primary source in this area. The patent (filed 2015, granted 2019) says that scores of organic stories and ads are converted to a common unit and sorted in one ranking, and that ad spacing is limited by count or pixel height verified. The widely circulated Total Value formula has no Meta primary source inferred.

The underlying patent is US 10,345,993 B2 "Selecting content items for presentation in a feed based on heights associated with the content items" verified. It is a Meta Platforms patent, filed 2015 / granted 2019, and the sibling patent is US 9,729,495 verified.

• Unified ranking: scores of organic stories and ads are converted to a common unit of measurement and then sorted together in one ranking verified patent.

  "Scores associated with organic news feed stories and scores associated with advertisements are converted into a common unit of measurement, and advertisements and organic news feed stories are together ranked in a single ranking."

• Top buffer: before the first ad, the sum of the heights of organic stories must exceed a threshold distance verified patent.
• Spacing between ads: between ads there must be at least a certain number of organic stories, or the sum of their heights must exceed the gap distance threshold verified patent. It is expressed both by count and by pixel height.

  "at least a threshold number of organic news feed stories are presented between advertisements"

  "a combination of heights of the organic feed stories exceeds a gap distance threshold"

• position discount: a value predicting the drop in interaction by position, and this discount varies with the sum of the predicted heights of the candidates placed above verified patent.

  "a predicted decrease in user interaction with a content item based on the position"

• The height prediction itself is an ML model verified patent. It predicts the height at display time from content type/language/comment count and client screen size/OS/resolution/app version: "the likely height of the content item when presented".
• In sum, Meta's ad interleaving is based on pixel distance, not slot index. No other public source describes it this way.

The Marketing API documentation says the effective bid is computed from the bid strategy, the bid amount and the probability of achieving the optimization goal verified Marketing API: Bidding Overview. It is the only mechanical statement on the official documentation side.

"Facebook evaluates your bid_strategy, bid_amount, and the probability of acquiring your optimization_goal to calculate an effective bid"

⚠ The widely circulated Total Value formula has no Meta primary source inferred. The formula has the form Total Value = Bid × Estimated Action Rate × Ad Quality. The corpus checked was the bidding and delivery pages of the Marketing API developer docs and the ads-related Transparency Center pages. Secondary sources write multiplication and addition differently. It must not be quoted as a formula.

⚠ "Ads Allocation in Feed via Constrained Optimization" (KDD 2020) is a LinkedIn paper. It cannot be used as Meta evidence.

Ad load (the share of the feed occupied by ads) is not published as product documentation inferred. "ad-load optimization" is repeatedly mentioned on earnings calls as a contributor to impression growth, but there is no numeric breakdown inferred.

↑ Contents

Cold start

Public material on Facebook cold start is far more concrete on the new-content side. The Epinet paper on Facebook Reels defines cold start by the number of users reached, approximates uncertainty at the retrieval stage and uses Thompson Sampling verified (New content). We found no public Meta material on new-user cold start inferred (New users).

New content — much more concrete here

The 2024 paper Epinet for Content Cold Start by Meta and Stanford addresses new-content cold start on Facebook Reels verified arXiv:2412.04484.

• The target is Facebook Reels, with about 120M users per day verified same paper.
• The cold-start definition is given as a number verified: "videos which have been shown to fewer than 10,000 users".
• The method approximates uncertainty with an epinet (epistemic NN) and uses Thompson Sampling verified.
  • The structure is a user tower and an item tower connected by an overarch, and the overarch consists of a base MLP and an epinet verified.
• The position is the retrieval stage, ahead of ranking and blending verified: "at the proposal stage of recommendation".
• A 5-day online A/B gave impression +17% aggregated over all impression ranges, and like/impression and completion/impression improved in low-impression ranges verified.
• The paper claims to be the first application of epinet to an online recommender system verified.

Side mechanisms that help cold start are also public.

• Meta Interest Learner matches new content to audiences with "few-shot learning" verified (2023). It adds that this works even with very little engagement: "even when there are very few engagements".
• SilverTorch's streaming index update is effectively new-content cold-start infrastructure verified. It decouples the model publish cadence from freshness and updates only specific tensors of the in-memory model in place: "targeted updates in-place to the specific tensors in the in-memory model". Details are in topics/model-freshness.md.

New users

No Meta primary source was found on cold start for new Facebook users (no primary source found) inferred. The corpus checked was 15 system cards, the ranking posts on engineering.fb.com, ai.meta.com and about.fb.com, and the Epinet paper. For Instagram, one-hop/two-hop expansion is public; see services/instagram.md for detail.

↑ Contents

Feedback signals

Facebook introduced dwell time signals in 2015 and 2016 and also published how they are normalized verified (dwell time section). On 2026-10-13, Meta announced Your algorithm, which lets users manage the topics in Feed and Reels recommendations directly and write what they want in free text verified (user controls). However, the widely circulated MSI (meaningful social interactions) point table has no Meta primary source inferred (MSI section).

dwell time — normalization was published too

Facebook introduced story dwell time in 2015 and external article dwell time prediction in 2016, and stated how it corrects for connection speed, loading time and article length bias verified.

• Introduced on 2015-06-12, with the bias correction stated explicitly verified Taking Into Account Time Spent on Stories.

  "Some people may spend ten seconds on a story because they really enjoy it, while others ... because they have a slow internet connection"

  • The fix is a within-person relative comparison: whether the user dwells on a given story much longer than on most other stories they view: "if people spend significantly more time on a particular story than the majority of other stories they look at".
• Introduced dwell time prediction for external articles on 2016-04-21 verified More Articles You Want to Spend Time Viewing.
  • There are two corrections. Loading time is not counted, and only dwell time within a cap is considered so that longer articles are not favored.

    "We will not be counting loading time"

    "looking at the time spent within a threshold so as not to accidentally treat longer articles preferentially"

  • Length bias is handled by clipping at a cap.
  • It also states that, before this, clicking an article and returning right away was already used as a clickbait signal: "clicked on an article and came straight back".

MSI (2018) — the announcement has no numbers

The 2018 MSI overhaul announcement contains only one number, and its effect was disclosed 3 years later verified. The circulating MSI point table comes from secondary reporting on leaked documents and has no Meta primary source inferred.

• Mosseri announced on 2018-01-11 that Facebook would prioritize posts that spark conversations and meaningful interactions verified Bringing People Closer Together.

  "we will prioritize posts that spark conversations and meaningful interactions"

  "inspire back-and-forth discussion in the comments"

  • Meta also warned of a side effect itself: Pages could see reductions in reach, video watch time and referral traffic verified.

    "Pages may see their reach, video watch time and referral traffic decrease"

• The announcement contains only one number verified: "live videos on average get six times as many interactions as regular videos".
• The effect was disclosed 3 years later, stating that the change reduced the time spent on Facebook per day verified Clegg, 2021.

  "the change led to a decrease of 50 million hours' worth of time spent on Facebook per day"

• The MSI point table has no Meta primary source inferred. It is a secondary quote from press coverage of the leaked Haugen documents. CNN reported it on 2021-10-27, WaPo on 2021-10-26, and WSJ on 2021-09-15.
  • The reported table values are like 1 / reaction·reshare 5 / RSVP 15 / significant comment·share 30, same group 0.5, stranger 0.3, angry 1.5→0.
  • When quoting, it must be stated that this is "secondary reporting on leaked documents."
• However, the direction matches Meta's own documents inferred. The Transparency Center states directly that "predicting an angry reaction lowers distribution."

User controls

Show more/less is described only as temporarily raising or lowering the ranking score, and the multiplier and duration have never been published inferred. On 2026-10-13, direct topic management and free-text feedback were announced verified.

• Neither the multiplier nor the duration has ever been published. Only "temporarily" appears, with no time unit inferred.

  • The corpus checked is the fb-feed card body, the 2022-10-05 newsroom post (version updated 2025-03-11), and three ai.meta.com Show more/less posts. They were viewed on 2026-08-09.
  • This is not evidence that undisclosed documents lack them.

• However, the method for generalizing sparse signals is public verified The new AI-powered feature designed to improve Feed, 2022.

  • It builds user/post embeddings with deep learning and trains Show More/Less by transfer learning from Like data.
  • The model resolves ambiguity, such as whether the intent is "see less of this person" or "see less about cars."

• "Your algorithm" was announced on 2026-10-13. It does not mention Instagram or Threads, and it gives no country or schedule information verified. The announced features include the following verified (existence of the features):

  • A space inside the three-dot menu to add and manage the topics that drive Feed and Reels recommendations
  • Free text for describing what the user wants and does not want
  • Follow-up questions that may appear after Interested/Not interested

  "LLMs can incorporate the interests, context, and intent behind natural language, unlocking an algorithm you can actually talk to." "Tell it what you want more or less of, and your recommendations respond in real time, adapting within the same session to reflect what matters to you right now." — Your Algorithm, Your Words, 2026-10-13 (newsroom)

• Real-time in-session adjustment and LLM use are newsroom claims, and there is no evidence for the mechanism.

• We infer that an LLM turns the free text into topics or embeddings and feeds them to the existing ranking as conditions inferred. The post says only "freeform text together with everything we already understand" and does not mention replacing the ranker.

• This is the first Meta source this repository has read that explicitly states same-session adjustment verified (wording comparison in the table above). The earlier Show more/less said only "temporarily".

Survey signals

Facebook introduced the "Is this post worth your time?" survey in 2019 and added three kinds in 2021-04: inspirational / topic interest / negative (posts with many angry reactions). Surveys are the area Meta has disclosed most deeply, and they span services. So the detail moved to topics/survey-signals.md.

↑ Contents

EU DSA — three facts missing from the US page

The EU DSA response documents contain three facts that the US page lacks inferred. Of these, the non-profiling option is confirmed by the Meta newsroom verified. The per-topic signal differentiation and the system card count are unverified leads from a PDF hosted by a third party inferred.

⚠ Items 1 and 2 below are unverified leads. We do not delete them; we only lower their grade. inferred The source is a PDF hosted by a third party (panoptykon.org), not Meta's own site, and we could not reach the original text on a Meta domain as of 2026-08-09. Also, this repository does not trust figures and quotes extracted from PDFs. PDF summarization tools have a history of inventing numbers that are not there (same caution in the sources section of services/threads.md). Until re-verified, do not cite items 1 and 2 as verified.

The supporting document is pp. 27–28 of the Facebook DSA Systemic Risk Assessment inferred. The assessment period is 2023-09–2024-08, and it was published in 2024-11/12.

1. Signal weights differ by topic. The report says it reduces the role of shares and comments in distribution for sensitive topics inferred. If true, this is the only Meta first-party statement that MSI-family signals are used differently by topic inferred.

   "We reduce this risk by limiting the role of shares and comments in the distribution of sensitive topics."

   • We noted that the same sentence also appears in the Instagram report, but that report has no direct document link either. See services/instagram.md.
2. The number of recommender system cards the report states differs from the 22 in the 2023 announcement and the 30 in the current index (FB 15 + IG 10 + RL 5) inferred.

   "we now have 15 recommender 'System Cards'"

   • 15 appears to be a count limited to DSA recommender systems inferred.
3. A non-profiling option exists verified DSA response announcement, 2023-08-22. Stories and Reels can be viewed in chronological order from followed people only, and search returns results based only on the words entered.

   "users will have the option to view Stories and Reels only from people they follow, ranked in chronological order"

   "based only on the words they enter"

   • ⚠ It lists only Stories / Reels / Search and does not name the Facebook Feed itself. DSA Art. 38 requires at least one non-profiling option.

We could not find a dedicated EU document for the "main parameters" that DSA Art. 27 requires (no distinct primary source found) inferred. Meta presents the system card set itself as compliance with Art. 27 inferred.

There is a dispute over how effective the implementation is inferred. On 2025-10-02, the Amsterdam District Court ruled that returning users to the algorithmic feed every time they launch the app is a dark pattern prohibited by the DSA inferred. It set the penalty at €100,000 per day (cap €5M) inferred. The appeal court upheld it in 2025-12, and Meta withdrew most of its appeal in 2026-01 inferred. We could not confirm the original ruling; this summary is based on secondary reporting.

↑ Contents

Generalizations from this service

As of the 2021 engineering blog, the Facebook Feed score is a weighted sum that adds per-action predictions verified (the only published formula). As of the patent, ads are converted to a common unit with organic stories and enter a single ranking verified (ads blending). Diversity rules are not public at the algorithm level inferred (re-ranking, diversity, filters).

• topics/multi-task-ranking.md: contrasts V = Σ w·Y (additive) with the multiplicative form of Instagram Suggested Posts. Within the same company there are 4 combination forms (counting additive with negative terms separately).
• topics/candidate-allocation.md: organizes the pass 0 candidate-count table. Ads are not a separate slot; they are converted to a common unit and then subject to a single ranking and pixel-spacing constraints.
• topics/feed-diversity.md: records the fact itself that diversity is not disclosed at the algorithm level.
• topics/embedding-retrieval.md: covers EBR 2020 hybrid inverted-index NN, SilverTorch, and RankGraph.
• topics/survey-signals.md: covers how surveys are used as ranking signals.
• topics/model-freshness.md: covers retraining cadence per surface.
• topics/sequence-transformer-ranking.md: covers HSTU / ULTRA-HSTU / LLaTTE.

↑ Contents

Open questions

The questions below are not yet answered by Facebook sources.

1. What is the structure of the pass 1 multi-task neural network? We cannot tell whether it is MMoE/PLE-family or how many towers it has. There is no primary source.
2. What form do the pass 2 diversity rules take? We cannot tell whether they are hard slot constraints or score penalties. There is no primary source.
3. What are the actual scale and sign range of the value model weight w? Only the direction is public: "hide/report/angry lower distribution."
4. By what rule are the 500 connected results and the several hundred unconnected results interleaved? Meta says only that the two systems are separate.
5. Has HSTU been applied to Facebook Feed? The paper anonymizes the surface ("a large internet platform"). The best current wording is the narrower one from ULTRA-HSTU (2026): "a large-scale production video serving platform that reaches billions of users daily". ⚠ "video recommendation service" is Zuckerberg's Q2 2024 IR wording, not ULTRA-HSTU's.
6. How many demotions have actually been removed since 2025-01? Two Meta pages contradict each other (4 vs 7).
7. Does the Facebook Feed itself have a DSA non-profiling option? If it does, we also do not know how it was implemented.
8. How does Facebook handle cold start for new users?
9. Is Click-Gap still operated today?
10. We have not yet read the 2026-03 Rewarding Original Creators on Facebook post. The current form of the originality demotion needs to be confirmed.
11. On which surface is ConnectionMind actually deployed? The paper does not disclose the surface. Judging from the graph schema, Facebook Reels or Video is likely, but we could not confirm it.
12. What are ConnectionMind's serving cost and latency? There is a statement that an 8B LLM enters the heavy-user hot path, but there is no QPS, p99 latency, GPU cost, or SLO at all.
13. How does ConnectionMind relate to the HSTU/GEM/LLaTTE lineage? The paper has 0 citations to them. There is no basis to judge whether it is the same team or a separate line.
14. What is the exact diff of the 2026-10-06 update to the fb-feed card? The original text of the earlier version is not available. Whether the viewer embedding sentence was deleted and whether a sentence excluding integrity at the inventory stage was added will be checked once Wayback is recovered.

↑ Contents

Sources

Meta official — engineering/newsroom

• How machine learning powers Facebook's News Feed ranking algorithm — 2021, blog post. The only primary source for the formula and the 3-pass design. ⚠ The same post is also reachable under a /core-infra/ slug
• News Feed Ranking in Three Minutes Flat — 2018, blog post. ⚠ The engineering.fb.com/2018/05/22/… path returns 404
• Bringing People Closer Together (MSI) — 2018
• Taking Into Account Time Spent on Stories — 2015
• More Articles You Want to Spend Time Viewing — 2016
• Fighting Engagement Bait on Facebook — 2017
• Remove, Reduce, Inform (Click-Gap) — 2019
• Sharing Our Content Distribution Guidelines — 2021
• Incorporating More Feedback Into News Feed Ranking — 2021
• New Ways to Customize Your Facebook Feed — 2022 (updated 2025-03-11)
• The new AI-powered feature designed to improve Feed — 2022
• How AI Influences What You See on Facebook and Instagram — 2023
• The AI behind unconnected content recommendations — 2023
• More Speech, Fewer Mistakes — 2025. Basis for the reduction in demotions
• Your Algorithm, Your Words — 2026-10-13, newsroom. Topic management and free-text feedback. It has no description of the mechanism
• You and the Algorithm: It Takes Two to Tango — 2021. ⚠ It has no mechanical content at all. It cannot be used as evidence for signals or weights

Transparency Center (primary regulatory documents)

• Our Approach to Facebook Feed Ranking — updated 2025-06-11
• Our approach to explaining ranking (system card index)
• Facebook Feed (read from the version updated 2026-10-06, rendered on 2026-10-07) · Feed Recommendations · Reels · Video · Ranked Comments
• Types of content we demote — updated 2025-07-02
• Reducing the distribution of problematic content — updated 2025-04-25
• Widely Viewed Content Report
• Regulatory and Other Transparency Reports
• ⚠ EU DSA Systemic Risk Assessment — Facebook — 2024, regulatory filing. It is a PDF hosted by a third party (panoptykon.org), and we could not reach the original text on a Meta domain. This repository does not trust figures extracted from PDFs, so every claim based on this document is inferred.

IR (regulatory disclosures — primary source for ranking performance figures)

• Q1 2024 · Q2 2024 · Q1 2025 · Q2 2025 · Q3 2025 · Q4 2025 · Q1 2026

Patents and papers

• US 10,345,993 B2 — Selecting content items ... based on heights — granted 2019, patent. The only mechanical source for ad interleaving
• Epinet for Content Cold Start — 2024, paper (Meta and Stanford). Ran a 5-day online A/B test (experiment deployment) on Facebook Reels
• ConnectionMind: Leveraging Social Networks and Large Language Models for Personalized Recommendation at Meta (arXiv 2608.10187) — 2026-08-10, paper. Meta's first statement about an LLM in the online serving path of an organic feed. Llama 3.1-8B handles the hot path for 5–10% of heavy users, and a Student GNN handles 90–95%. Surface undisclosed ("short-form video platform")

Secondary (grade downgrade required)

• CNN via KVIA — MSI score table — 2021, reporting on leaked documents
• Washington Post — angry emoji weighting — 2021
• TechCrunch — reproduction of the quoted Zuckerberg 2018 borderline note — 2018

↑ Contents