2026-10-17

1TransX scales a behavior-sequence ranking model by component and splits serving into three stages — LinkedIn (radar)

• Split the model into a sequence encoder, behavior memory and candidate-conditioned fusion, and size each separately.
  • Behavior memory is a fixed-length compressed state of the user history, independent of the candidate.
  • Candidate-conditioned fusion is the part that uses the candidate as the query and reads the memory with cross-attention.
• Disaggregate serving, which moves expensive computation out of request time.
  • Serving disaggregation is a setup that hands computation formerly done on one server to different stages and resources according to its nature.

• The evaluation surface is a "social recommendation application" with about 20K requests per second and 0.5K–1.2K candidates per request verified.

  • The paper does not disclose which LinkedIn product this is.
  • We infer it is likely the LinkedIn feed inferred. The evidence is only the author affiliation and the phrase "social recommendation".

  "We evaluate TransX in a large-scale social recommendation application serving approximately 20K requests per second with roughly 0.5K–1.2K candidates per request." — arXiv 2610.20913 §1 (paper), 2026-10-14

• Production A/B ran separately for each variant, and the baseline differs by variant verified.

  • The core model (TransX-Base) shows CTR +6.02% and conversion +4.45% over the existing production sequence model.
  • The variant post-trained for retrieval shows CTR +2.96% over the two-tower (a structure that embeds the user and the candidate with separate towers and scores by dot product) sequence model.
  • The ranking variant shows conversion +1.33% over a DLRM (a deep learning recommendation model architecture that combines sparse feature embeddings and an MLP) learning-to-rank baseline.

  "In separate production A/B experiments, TransX-Base improves click-through rate (CTR) by 6.02% and conversion by 4.45% over the production sequential model baseline; TransX-Retrieval improves CTR by 2.96% over the sequential two-tower production model baseline; and TransX-Ranking improves conversion by 1.33% over a DLRM learning-to-rank baseline." — same source §1

• Behavior memory is independent of the candidate and finite in size, and holds recent actions at high resolution and older actions as lower-resolution summaries verified (§3.2).

  • Because it is independent of the candidate, a memory computed once can be shared by all candidates in the request.

• Serving is split into three stages: offline bulk state construction, nearline incremental update and online prediction verified (§5.3).

  • The abstract says this split cut total serving cost by nearly 80% against colocated serving (a setup that does all computation on one server) at the same traffic and SLO.

  "TransX therefore disaggregates serving into three stages: bulk state construction, nearline incremental maintenance, and online prediction." — same source §5.3

• As a result, the same total serving cost allows about 5× the FLOPs per request of an HSTU-family production baseline verified.

  • The abstract says that matching FLOPs per request gives about 4× request throughput per H200. Training infrastructure raised supervised target throughput by 84% under the same conditions verified (abstract).

• Scaling returns differ by component (anisotropic) verified.

  • With memory resolution and fusion capacity fixed, scaling encoder depth can saturate beyond moderate depth.
  • Increasing supervision shifts the optimal allocation toward fusion.
  • The memory resolution needed to preserve quality stays roughly constant as the model grows.

  "encoder scaling may saturate beyond moderate depth at fixed memory resolution and fusion capacity; increasing supervision shifts the allocation frontier toward candidate-conditioned fusion; and the memory resolution needed to preserve quality remains roughly stable as model capacity grows." — same source §1

↑ Contents

Corrections

• A secondary report saying "on 2026-10-15 Threads changed ranking to prioritize posts from accounts you follow" is judged a re-report with a wrong date, not a new announcement, and the canon was not changed inferred.
  • The evidence is that the phrase the article quotes is identical, character for character, to a 2024-11-21 post, and we found no new post with the same phrase in 2026-10. The article text could not be opened in this session.
  • This article came from the WebSearch ("Facebook Instagram Threads feed recommendation algorithm update October 2026") results.
  • The Mosseri phrase the article quotes ("We are rebalancing ranking to prioritize content from people you follow…") is the same as the 2024-11-21 Mosseri post.
  • This change is already recorded as 2024-11-21 in the cold start section of services/threads.md.

↑ Contents

Radar

• TransX (LinkedIn) covers per-component scaling and three-stage serving disaggregation, and reported CTR +6.02% in A/B. Added to "New this month" in radar.md.

↑ Contents

What to check next

↑ Contents

Lookup log

The tracked services (X, Meta, Netflix) had no new primary material.

• X: the latest commit in the public repository xai-org/x-algorithm is 92f0918 from 2026-10-15, already covered in the 2026-10-16 daily. There is no release, and no WebSearch cross-check was done.
• Meta: the new posts on engineering.fb.com are three dated 2026-10-15, all unrelated to ranking and recommendation.
  • The three posts cover the AI-based development process of the Forum app, handling World Cup traffic, and a Sapienz award. We opened the Forum post and confirmed it has no ranking statement.
  • The latest post in the ai.meta.com/blog list is from 2026-07. transparency.meta.com was not checked this time.
• Netflix: the 10 RSS items are the same as those seen in the 2026-10-16 daily, and all were reposted on 2026-10-14. WebSearch cross-checks found no new announcement with a primary source.
• arXiv cs.IR: 30 items were parsed, and 1 of them (TransX) passed.
  • The near miss is RecSoup. It is a CIKM 2026 paper that solves controllable multi-objective recommendation by model merging, but it has only public benchmarks.
• It was Saturday, so the industry blog sweep was not done.

↑ Contents