2026-10-16

Other changes reviewed today:

• X changed the ads author brand-safety fallback to on by default in the same commit.
• Facebook announced "Your algorithm", which adjusts Feed and Reels recommendations with free text.
• Netflix introduced in a blog post a measurement study finding that the current recommendations raise engagement by 4% over matrix factorization.

1The video carousel can now also pick videos using the immersive-surface model's predictions — X

• The carousel's sort score is now the probability of opening the video and the VQV probability predicted by the immersive-surface model, each multiplied by its weight and summed verified — home-mixer/util/video_carousel.rs:104-123.

  • The formula is VideoCarouselVideoOpenWeight × video_open + VideoCarouselImmersiveVqvWeight × immersive_vqv.
  • A video with no immersive score has that term counted as 0.

• The defaults are a video_open weight of 1.0 and a VQV weight of 0.0 verified — home-mixer/params/param.rs:1083-1094. Default behavior is therefore the same as before.

  • The carousel itself (EnableVideoCarousel, lines 1053-1058) is also still off by default.

• The two components below run only when all three conditions hold verified: the VQV weight is not 0, the request is a For You request, and the carousel is on.

  • The three conditions are checked in ranks_by_immersive_scores (video_carousel.rs:17-21).

• The first component fetches the user behavior sequence once more with a different aggregation approach. That approach (IMMERSIVE_LONG_DWELL) got only a name in the proto in the 2026-09-10 commit (Gen4+19), and this is the first home-mixer code to use it verified — query_hydrators/immersive_scoring_sequence_query_hydrator.rs, param.rs:163-168, 1101-1106.

  • The component is named ImmersiveScoringSequenceQueryHydrator (75 new lines).
  • The default aggregation the For You ranker uses is DENSE_WITH_LONG_DWELL, and this component fetches with IMMERSIVE_LONG_DWELL.

• The second component collects only the carousel-eligible videos among the candidates and requests predictions from a different Phoenix cluster on an immersive-surface basis. In the request, the video length field is filled with the minimum video length verified — scorers/immersive_video_scorer.rs, util/phoenix_request.rs:228-252.

  • The component is named ImmersiveVideoScorer (65 new lines).
  • Carousel-eligible videos are videos that are not replies and have a video_open score.
  • The requested cluster is Experiment3Fou, and the surface value is ProductSurface::GalleryPage.

• The second component runs between the Phoenix scorer and VMRanker (a scorer that reorders the candidate set to be shown in the feed) verified — candidate_pipeline/phoenix_candidate_pipeline.rs:462-471.

  • The order is PhoenixScorer → ImmersiveVideoScorer → VMRanker.

• The immersive prediction is stored only in the candidate's immersive_scores field, and this diff has no code that reads this field in main-feed score computation verified.

• The same weighted sum is used in both stages that pick carousel videos verified — selectors/top_k_score_selector.rs:16-25, filters/video_carousel_filter.rs.

  1. When picking 30 more carousel candidates from outside the top 50 by score
  2. When picking the final 3–5 after the filter

• A new test shows the expected behavior. With a VQV weight of 1.0 and a video_open weight of 0.0, videos with VQV of 0.9, 0.5 and 0.1 are picked in that order verified.

  let result = TopKScoreSelector.select(&carousel_query("0.0", "1.0"), candidates);
  ...
  assert_eq!(tweet_ids(extras), vec![101, 106, 103]);
  

• The ads author brand-safety fallback switch changed from false to true verified — param.rs:807-812.
  • The switch is named EnableAdsAuthorBrandSafetyFallback.
  • The sync time of the cron mirror (a file that syncs production config values every day and publishes them) changed from 2026-10-13T16:00:30Z → 2026-10-14T16:00:44Z.
• This switch substitutes an author-level label when a post-level Grok safety label does not yet exist. It was introduced off by default in the 2026-10-03 commit (Gen4+35). The V2 verdict switch (Gen4+2, introduced 2026-08-17), which the same commit changed to true by default, is also true, so both switches are now on verified.
  • The author-level labels substituted are GrokSfa/GrokNsfaLimited/GrokNsfa.
  • The V2 verdict switch is EnableAdsBrandSafetyVerdictV2 (lines 801-805).

Visibility-filtering (VF) is a service that decides, post by post, whether "this viewer may be shown this". Each SafetyLevel (a bundle of per-surface policies) has a rule list, and a rule has the form "if the condition is true, take this action (drop, blur, limit engagement, etc.)".

• The way rule conditions are written changed from a flat array (an AND of all conditions) to a Boolean tree that combines named conditions verified — visibility-filtering/rules/rule_spec/rule.rs (305 new lines).
  • The tree is built with all!/any!/! and named conditions.
  • For example, the two conditions "the request came from an age-verification country and the viewer has not completed age verification", which were repeated in about 25 rules, are now combined into the single VIEWER_IS_NOT_AGE_VERIFIED_IN_AGE_VERIFICATION_COUNTRIES.
• The rule lists, order, actions and country lists of the five SafetyLevels are unchanged verified — rules/levels.rs.
  • A test adapter (rules/legacy_table.rs) was added that re-outputs the new tree as the old flat table and compares it with the existing golden files, and the golden files did not change in this commit. The test therefore guarantees that there is no policy change.
• Evaluation, policy-catalog document generation and per-node verdict traces all come from one and the same rule definition verified. The serving path short-circuits without a trace.
  • Catalog generation is handled by rules/catalog.rs and is test-only.
  • The per-node trace is produced by rules/rule_spec/observed.rs (275 new lines).
• The debug tool explain.rs (+854 lines) exposes this trace, so we read this as a direction toward explaining "why was this post hidden" rule by rule inferred.
• Behavior may change in two viewer lookups outside the rules.
  • The Gizmoduck (X's user information service) lookup that fetches viewer information changed from a call without context to a lookup context that includes deleted, deactivated and soft users verified — clients/gizmoduck_client.rs:23-29, 55-62. The test name changed from viewer_lookups_pass_no_context to viewer_lookups_include_every_user_state_tweetypie_reads, so we infer the aim is to align the lookup scope with Tweetypie, the reference service for comparison inferred.
  • If a found viewer has no age information, it is now decoded as "age not stated" (NotStated) regardless of state verified — hydration/decode/viewer.rs:76-79. When the two changes combine, a viewer in a state that previously was not looked up could become a target of the sensitive-media drop rule for viewers with no stated age inferred. The diff does not show the scale.
• Distribution metrics by viewer state and by age (vf_viewer_state, vf_viewer_age) were introduced verified. We read them as instrumentation to observe the effect of the two changes above inferred.

• The assert that was required together when turning on seqpack (packing that concatenates several user sequences into one batch row) was removed from the Phoenix training code verified — phoenix/xrex/models/recsys_model.py. Only the constraint was lifted and there is no config change, so we read it as a change that permits experiments inferred.
  • The removed assert was the one requiring right_anchored_rope.

• We infer that the carousel is treated as "the entrance to the immersive viewer" and that the aim is to pick the entrance's videos with the viewer's model as well inferred. There are two grounds.
  • GalleryPage, Experiment3Fou and IMMERSIVE_LONG_DWELL are all names from the immersive surface side.
  • The reward of the training config immersive_page_rl, released the day before, counts VQV completion separately.
• Following the video_open weight cut (0.07 → 0.02, Gen4+40) and the dwell-time reward (Gen4+41), this commit points the same way as the shift of the video-side objective from "opening" to "time actually watched and video quality views" inferred.
• Because the VQV weight defaults to 0, it is unknown whether this is actually served.

↑ Contents

2Announced "Your algorithm" for adjusting recommendations with free text — Facebook

• In the "Your algorithm" space inside the three-dot menu, users can directly add and manage the topics that drive Feed and Reels recommendations verified (existence of the feature).
• Users can write what they want more and less of in free text, and a follow-up question may appear after pressing Interested/Not interested verified.
• The scope is Facebook's Feed and Reels. There is no mention of Instagram or Threads and no information on countries or launch schedule verified.
• The statement that an LLM interprets natural language and that recommendations change within the same session is the newsroom's claim, and there is no model structure or figure.
• We infer that an LLM converts the free text into topics or embeddings and feeds them into the existing ranking as conditions inferred. The basis is that the post says only that it is used "together with what it already knows" and does not mention replacing the ranker.

"LLMs can incorporate the interests, context, and intent behind natural language, unlocking an algorithm you can actually talk to." — Your Algorithm, Your Words, 2026-10-13 (newsroom)

"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." — Same post

↑ Contents

3Measured the value of personalized recommendations with a choice model — Netflix

• It estimates members' daily viewing choices with a low-rank choice model verified (blog).
  • The model learns title representations and a member state that summarizes viewing history. A separate "recommendation effect" term is added for recommended titles.
  • Variation in whether a title is recommended comes from natural variation such as algorithm exploration verified (blog).
• The validation experiment is a randomized A/B that slightly pushes recommendations toward, or pulls them away from, specific categories. The correlation between the model-predicted diversion ratio and the experimental value was 0.86, with an R² of 0.73 verified (blog).
  • The diversion ratio indicates, when some set of titles is shown less, how much of the reduced viewing moves to other titles and how much leaves Netflix.
• In the counterfactual comparison, engagement relative to the current recommendations falls as follows verified (blog, abstract).
  • Random recommendations: −16%
  • Popularity-based recommendations: −12%
  • Basic matrix factorization (MF): −4%
• Popularity-based recommendations and MF concentrated viewing on fewer titles, while random recommendations have slightly higher diversity but a large engagement loss verified (blog).
• Splitting the viewing difference produced by recommendations into three paths, mechanical exposure accounts for about 7% verified (blog).
  • Most of the rest is selection (showing to people who would have watched anyway) and targeting (showing to people whose decisions recommendations change).
  • The targeting effect is about 7× the exposure effect and is largest for mid-popularity titles verified (blog).

"Relative to the current system, random recommendations reduce engagement by 16%, popularity-based recommendations reduce engagement by 12%, and matrix factorization reduces engagement by 4%." — Measuring the Impact of Personalized Recommendations (blog)

"Exposure is not the main driver of differences in consumption, as mechanical exposure explains only about 7% of the gap under the current recommender system, while selection and targeting account for the vast majority of the difference and the targeting effect is nearly seven times larger than the exposure effect." — Same post

↑ Contents

4A paper disputes the claim that generative retrieval beats dense retrieval — Radar (Shopify)

• It argues that most of the gap by which GR appears ahead on public data (Amazon Toys) comes from popularity. Mixing the same static item value once into the DR score reverses the gap, and recall also stays higher than GR verified.
• It also reports that on an internal large-scale e-commerce dataset, GR did worse than a small DR encoder verified. There is no online A/B or deployment claim.

"Finally, on an industrial-scale internal e-commerce dataset with LLM-judged relevance, we show that a 14B SID-LLM still underperforms a $47\times$ smaller dense encoder." — arXiv 2610.19065 abstract, 2026-10-13

↑ Contents

Corrections

• The VF summary in the safety section of x.md stated the Gen4 (2026-08-13) configuration, "3 SafetyLevels, 54 recommendation rules", in the present tense.
  • While analyzing VF, we confirmed that there are currently five levels (Home Hydration and Immersive Expanded Recommendations were added) and that the constant entry counts are 1·39·65·87·47 verified — rules/levels.rs.
  • We kept the original text (services/x.md) as a record of the Gen4 point in time and added a correction paragraph directly below it.

↑ Contents

What to check next

↑ Contents

Coverage

• X: One commit has been pushed to the X algorithm repository since 2026-10-15.
  • The one commit confirmed in the xai-org/x-algorithm commit list is 92f0918.
  • The WebSearch cross-check ("X For You algorithm video carousel October 2026") turned up no new announcement with a primary source.
• Meta: The latest posts on engineering.fb.com are from 2026-10-12 and 2026-10-13, and the recommendation-related post (GEM) was already covered in the 2026-10-14 daily.
  • The ai.meta.com/blog list has no recommendation or ranking posts.
  • The ranking explanation index on transparency.meta.com is still "UPDATED DEC 31, 2023", and we did not open the individual system cards.
  • We found the 2026-10-13 newsroom (about.fb.com) post that had been missed and folded it into section 2 above.
• Netflix: All 10 items in the RSS were republished at 2026-10-14 19:47–20:02 UTC, so the dates cannot be trusted.
  • Of these, the only post related to feed and recommendation is the one in section 3 above. The rest are infrastructure posts (DataJunction, TimeSeries store, ads event pipeline, Conductor, Spark, Stratum, GPU compute), the FM integration post already covered, and the LLM-as-a-Judge post.
  • The latest item in the research.netflix.com archive is 2026-02-24.
• arXiv cs.IR: We parsed 30 items and 1 passed (Shopify GR vs. DR).
  • Two came close to passing. AdCoEvo automates ad delivery operations and is far from feed ranking, and Dual-Branch Graph Distillation for Cold-Start has only public benchmarks.
• Because it is Friday, we did not skim industry blogs.

↑ Contents