Daily reports
2026-08-08 ~ 2026-10-21
- 62 reports
- 1220 min of reading
How to readEach report follows the same order: In one line → Source → Background → Canon update → Why it matters. If you are short on time, the one-line summary on the cover and the Why it matters cards give you the day's change.
2026-1015
- Meta changed the item representation in its ads ranking and retrieval to sequences of semantic tokens and disclosed that, together with GEM, it raised conversions by 0.35%.
- Sources read: 3
- Canon files updated: 2
- Radar entries: 1
- X switched three dedup rules for the For You video carousel to on by default and stopped the carousel from appearing in the same response as Who-to-Follow. The carousel itself i…
- Sources read: 3
- Canon files updated: 1
- Radar entries: 1
- The tracked services (X, Meta, Netflix) had no new primary material. LinkedIn released TransX, which scales the ranking model by component and splits serving into three stages, …
- Sources read: 1
- Canon files updated: 0
- Radar entries: 1
- X let the For You video carousel mix in the immersive-surface model's video quality view (VQV) prediction when ordering videos. The VQV weight defaults to 0.0, so default behavi…
- Sources read: 4
- Canon files updated: 3
- Radar entries: 1
- X released the training code for a re-ranker that fills one page of an immersive surface using reinforcement learning. The reward is based on dwell time, and the serving switch …
- Sources read: 2
- Canon files updated: 2
- Radar entries: 1
- X raised the like and dwell time weights and lowered the video open weight from 0.07 to 0.02 in the weight mirror of its feed service home-mixer.
- Sources read: 4
- Canon files updated: 2
- Radar entries: 1
- Pinterest deployed a structure that infers a user's long-term goals (journeys) with a single Qwen3-4B generation instead of keyword clustering and uses the resulting names as no…
- Sources read: 1
- Canon files updated: 0
- Radar entries: 1
- X added two fields to the inference request of its ranking model Phoenix that record whether a feed request is an app launch, a refresh, or a scroll down. The rest is H100 train…
- Sources read: 1
- Canon files updated: 1
- Radar entries: 0
- X built a carousel module that groups 3–5 vertical videos inside the For You feed. It is off by default.
- Sources read: 1
- Canon files updated: 1
- Radar entries: 0
- X turned on by default the global popular posts candidate source it released four days earlier. It is a source that feeds everyone the same list of popular posts as candidates, …
- Sources read: 1
- Canon files updated: 1
- Radar entries: 0
- X's ranking score formula (value model) gained, in code for the first time, three expected-dwell-time terms of the form "probability of opening × seconds spent once opened." The…
- Sources read: 2
- Canon files updated: 2
- Radar entries: 0
- X released the server crate for generative SID retrieval. The server only does prefix-range search over a prebuilt SID snapshot.
- Sources read: 1
- Canon files updated: 1
- Radar entries: 0
- X released a retrieval model that generates item semantic IDs (SID) to pick candidates, together with a For You candidate source that calls that service. The source is off by de…
- Sources read: 1
- Canon files updated: 1
- Radar entries: 0
- X's visibility-filtering service (VF) now decides, for each action-limit reason, which actions to block and what notice text to show, and its responses carry per-action text.
- Sources read: 1
- Canon files updated: 2
- Radar entries: 1
- Six cold-start parameters in the X Home feed changed. The cold-start retrieval candidate cap rose from 0 to 200, which turned this path on, and the post-age cap fell from 48 hou…
- Sources read: 3
- Canon files updated: 2
- Radar entries: 2
2026-0928
- X released the server side of the overturn-hold gate it introduced in the 2026-09-26 commit. The gate stops automation from re-applying a suspension or label that a human overtu…
- Sources read: 2
- Canon files updated: 2
- Radar entries: 1
- According to the KuaFu paper, item-level compression of user behavior history for profiling was deployed on Tencent's advertising and recommendation platform for 10 months, and …
- Sources read: 2
- Canon files updated: 1
- Radar entries: 1
- Pinterest said it deployed to production a standard-ads lightweight ranking model expanded from two-tower to 3-tower. In online A/B tests, three click metrics each rose about 1%…
- Sources read: 1
- Canon files updated: 1
- Radar entries: 1
- X added a new final precheck to its enforcement service that blocks re-applying suspensions and labels a human reviewer overturned on appeal. The gate's default mode is Off.
- Sources read: 2
- Canon files updated: 2
- Radar entries: 1
- X rewrote the data lookup stage of its visibility-filtering service (VF) in a single day. It removed the per-data-type lookup functions and moved them to a data table and a sche…
- Sources read: 2
- Canon files updated: 1
- Radar entries: 1
- X deleted the 1,255-line ranking scorer from home-mixer. Now home-mixer and vm-ranker compute the value model with the same code in a shared crate.
- Sources read: 1
- Canon files updated: 1
- Radar entries: 1
- The three slots announced yesterday were filled with real code within a day. The VM-Ranker service started computing the value model directly (4 new files, 56 parameters).
- Sources read: 1
- Canon files updated: 1
- Radar entries: 1
- X added modes that let the re-ranking service VMRanker either receive value model inputs or compute value model scores. Both parameters that turn on the new modes are false by d…
- Sources read: 2
- Canon files updated: 2
- Radar entries: 2
- Two production papers from Meta's search and discovery teams were posted to arXiv together. One describes a structure that co-serves two retrieval paths, GPU and CPU; the other …
- Sources read: 3
- Canon files updated: 0
- Radar entries: 2
- X rolled back both the experimental dwell-regret alternative scoring in home-mixer, its feed-assembly service, and the 19-feature user gate that decided entry to it. Only the we…
- Sources read: 1
- Canon files updated: 1
- Radar entries: 0
- X merged the flag set that handled delayed feedback in ads ranking training into one, and split training samples into three streams: fresh, early-relabel, and late-split.
- Sources read: 1
- Canon files updated: 1
- Radar entries: 0
- X added a holdout filter that removes popular posts from the home feed for each viewer at a ratio set by fav-count bracket. It is off by default.
- Sources read: 2
- Canon files updated: 2
- Radar entries: 1
- X connected an ads purchase-value loss to training of the For You ranking model (default weight 0) and changed the click dwell-time head to a binary classification at a 10-secon…
- Sources read: 2
- Canon files updated: 2
- Radar entries: 1
- X added observation that measures the diversity of the For You slate at three stages: before heuristics, top 10, and final. X's public code now has its first metric showing how …
- Sources read: 1
- Canon files updated: 1
- Radar entries: 0
- Pinterest deployed PinDCO, which picks ad creatives by weighing whole-page gains and losses, across its entire ads platform. In a production A/B test, ad CTR rose 3.09% in relat…
- Sources read: 2
- Canon files updated: 0
- Radar entries: 2
- X made it possible to pass a seed post into the diversity re-ranking (DPP) of the For You slate. Candidates similar to this post are more likely to be dropped from the result.
- Sources read: 1
- Canon files updated: 1
- Radar entries: 0
- X's visibility-filtering safety filter raised its backend lookup timeout to 390 ms and now evaluates author-based rules with a cached author id even when the tweet lookup fails.
- Sources read: 1
- Canon files updated: 1
- Radar entries: 1
- X's visibility filter now drops posts by an explicit action value instead of the filtering reason. Most filtering outcomes stay the same.
- Sources read: 1
- Canon files updated: 1
- Radar entries: 0
- X made the 26 engagement-head weights used in ranking scores perturbable in a fixed direction per user and per head. The default perturbation strength is 0.0, so default behavio…
- Sources read: 1
- Canon files updated: 1
- Radar entries: 0
- Belatedly found a 2026-02 blog post stating that Shopify deployed a generative recommendation model built on Meta's HSTU for product recommendation, and added it to the radar.
- Sources read: 1
- Canon files updated: 1
- Radar entries: 1
- Huawei published a CIKM 2026 paper that replaced the ranking model of an app distribution platform with 100M+ DAU, from MMoE to PTDG. PTDG builds a separate task dependency grap…
- Sources read: 1
- Canon files updated: 0
- Radar entries: 1
- Pinterest expanded its conditional embedding retrieval (CLR) for Home Feed into a unified architecture and cut p90 latency by 85%, from 80ms to 12ms.
- Sources read: 4
- Canon files updated: 0
- Radar entries: 4
- X fixed the cold-start experiment so that the control and treatment arms leave identical bucket observation records. In the treatment arm, cold-start candidates are narrowed to …
- Sources read: 1
- Canon files updated: 1
- Radar entries: 0
- X added a new reply spam verdict path to grox, its LLM annotation pipeline. It judges a chain in which one author keeps replying to their own replies as a whole thread, and the …
- Sources read: 2
- Canon files updated: 1
- Radar entries: 0
- X's feed server started requesting backbone scores from the ranking model Phoenix for all five feeds. For the first time, code shows that Phoenix has a two-layer structure of a …
- Sources read: 1
- Canon files updated: 1
- Radar entries: 0
- X merged the nine rule files of its impression verdict service into three and replaced the per-rule structs with one-line data definitions. Policy is unchanged.
- Sources read: 2
- Canon files updated: 2
- Radar entries: 2
- X named, in code for the first time, the 16 countries where it enforces age gating on NSFW media. South Korea is on the list.
- Sources read: 1
- Canon files updated: 1
- Radar entries: 1
- Alibaba published a paper extending SID (Semantic ID) generative retrieval to Pailitao, Taobao's image search surface. It released only an offline +13.8% figure and did not disc…
- Sources read: 1
- Canon files updated: 0
- Radar entries: 1
2026-0819
- X refactored 14 files in the post visibility-verdict rule layer with no policy change. CI now includes a 90-case golden corpus and a test that requires every wired rule to have …
- Sources read: 2
- Canon files updated: 2
- Radar entries: 0
- arXiv cs.IR: nothing new (no new submissions since 2026-08-27, weekend)
- X added a new ads blender that decides ad slots by looking at both post risk and ad tier. It is not yet the default blender and sits behind an experiment gate.
- Sources read: 2
- Canon files updated: 2
- Radar entries: 1
- X also made it possible to cap host memory usage when saving checkpoints of its recommendation model Phoenix. On the restore side, X added an approach that loads only one batch …
- Sources read: 5
- Canon files updated: 2
- Radar entries: 3
- X opened a general path in its account-action engine that lets scorers request actions directly. A request runs only if it passes both the action kinds fixed in code and a dynam…
- Sources read: 4
- Canon files updated: 2 (services/x.md, radar.md)
- Radar entries: 2
- X made the staging deployment of its visibility verdict service (VF) compare each request's verdict against the reference VF and log only the cases where the results differ.
- Sources read: 4
- Canon files updated: 2 (services/x.md, radar.md)
- Radar entries: 3
- No new sources for Meta and Netflix
- There was no new material for the tracked X, Meta, and Netflix. Instead, we added Xiaohongshu's production CVR ranker GateDiffInt, which combines diffusion and LLM intent distil…
- Sources read: 1
- Canon files updated: 1 (radar)
- Radar entries: 1
- The X For You ranking stage can now consume the re-ranking slate context that the Phoenix ranking model computes at serving time. The gate that turns on this path is false by de…
- Sources read: 2
- Canon files updated: 2 (services/x.md, radar.md)
- Radar entries: 1
- X's ranking model Phoenix switched its training from Adam to the Muon optimizer. This is the first case where the code confirms Muon adoption in a production recsys training pip…
- Sources read: 2
- Canon files updated: 2 (services/x.md, services/netflix.md)
- Radar entries: 0
- X removed the ads brand safety V2 migration logic that the canon recorded yesterday and redesigned the verdict within a day. A verdict now uses V2 if any V2 label is present and…
- Sources read: 1
- Canon files updated: 1 (x.md, 11 sections)
- Radar entries: 0
- X added a V2 path alongside the ad brand safety verdict and made conflicting adult content verdicts go back to Grok 4.5. If the judge disagrees, a hard verdict is lowered to sof…
- Sources read: 2 (1 X commit (Gen4+2), 1 of 30 arXiv triage items passed (SDF))
- Canon files updated: 1 (x.md, 7 sections updated)
- Radar entries: 1 (Google Discover SDF newly registered)
- Meta engineering blog posts: 0 (last relevant post 2026-08-05, already folded into the canon)
- Netflix tech blog posts: 0 (last post 2026-08-07, unrelated to algorithms)
- In a follow-up commit the day after the Gen4 release, X stated in a code comment the rule for reading weights: they multiply predicted probabilities, not action counts.
- Sources read: 1 (1 X Gen4+1 commit, 0 new Meta/Netflix blog posts, 1 of 30 arXiv triage items passed [DrEM])
- Canon files updated: 1 (x.md, 9 sections updated, 6 sources added)
- Diagrams updated: 1 (x-pipeline.svg filters 17→18)
- Radar entries updated: 1 (DrEM affiliation confirmed, A/B figures folded in)
- ⚠ Claims changed in bulk. In the Gen4 code X released on 2026-08-13, the ranking weights appeared with actual values for the first time.
- Sources read: 2 (X Gen4 release: 1
- arXiv triage: 1 watching of 30)
- Canon files updated: 1 (x.md heavily revised)
- Diagrams updated: 1 (x-pipeline.svg)
- Radar entries updated: 1
- ⚠ Claim changed. Meta's first statement that it put an LLM in the online serving path of an organic feed was published in the ConnectionMind paper. Llama 3.1-8B handles requests…
- Sources read: 3 (2 canon updates, 1 radar)
- Canon files updated: 3 (netflix.md
- facebook.md
- llm-in-recsys.md)
- Radar entries: 2
- There were no new announcements from the tracked X, Meta, or Netflix. Instead, four recommendation papers from Alibaba and Kuaishou, published within the same three days and con…
- Sources read: 4 (all radar candidates)
- Canon files updated: 0
- Radar entries: 2 new, 2 open threads updated
- There was no new announcement from the tracked services (X, Meta, Netflix) that would change the canon. Instead, we added two arXiv papers with production claims to the radar.
- Sources read: 3 (all radar candidates)
- Canon files updated: 0
- Radar entries: 2 new and 1 open thread update
- Netflix disclosed the serving infrastructure of its LLM ranker GenRec, and Meta disclosed the training infrastructure of its ads foundation model GEM. Both explain not what the …
- Sources read: 4
- Canon files updated: 4
- Radar entries: 2
- Quote defects found in yesterday's entries: 3