Field radar — papers and trends outside the tracked services

This document is a watchlist of papers and trends not yet folded into the canon of the tracked services (X, Meta, Netflix). Entries are chosen by relevance to production and written as 2–4 line pointers. Their lifespan is about 3 months, after which they are promoted or graduated. The detailed rules are in CLAUDE.md.

A daily batch handles collection. New arXiv cs.IR submissions are scanned and selected every day, and industry blogs (Google Research, Pinterest, Spotify, LinkedIn, etc.) and conference accepted lists are checked 1 time a week.

Watching (open threads)

Spread of LLM-native ranking — who follows after GenRec

• What: A trend of replacing one role in the recommendation pipeline with an LLM. As of 2026-08, five first-party statements have reached production deployment, and their mechanisms differ verified. → llm-in-recsys.md
  • Details: The five and their mechanisms are as follows verified.
    • Netflix GenRec: scoring function
    • Yandex Gryphon-v2/Sona: unifying generative retrieval and rank
    • Alibaba MetaStrategy: offline policy compiler
    • Meta ConnectionMind: graph path reasoner
    • Meta AMBER: input tokenization
  • Details: Of these, only AMBER did not disclose online A/B figures, and the other four stated A/B figures.
  • Details: When the thread was opened, only the scoring function in the serving path was considered. Statements that came in afterward touched different roles, which widened the scope. When a new statement arrives it is added to this list, and the body does not count the order.
• Why notable: Whether X (which uses an LLM for annotation) and Meta (which uses an LLM as a feature encoder) would take the same road was the biggest fork of the next 1 year. However, Meta entered with ConnectionMind in 2026-08 (⚠ the earlier proposition was half overturned).
• Watch for: We await first-party statements that other companies (Google/YouTube, TikTok) deployed an LLM in the serving path, follow-up news on the full GenRec rollout, and follow-up posts on ConnectionMind's surface confirmation and serving cost.
• 2026-08-12 additional evidence: Taobao MetaStrategy (arXiv 2608.09440, 2026-08-10) is a third angle, using the LLM as an offline policy compiler rather than placing it in the serving path. In a 7-day A/B on the "Guess You Like" surface it gained click PV +2.11%, IPV +3.12%, and GMV +2.83% verified.
  • Details: An LLM generates typed JSON (objective weights, category preferences, position policy, guardrails) nearline. A validator and compiler use it to instantiate a deterministic Generator, which competes atomically with existing strategies inside the GE (Generator-Evaluator) architecture.
  • Details: Because LLM inference is outside sync ranking, the RT impact is 0; in other words, there is no latency increase on the synchronous path.
  • Details: Unlike GenRec (scoring function) or Gryphon-v2 (generative retrieval and rank), this mechanism keeps the LLM outside the synchronous ranking path.
• 2026-08-13 additional evidence (canon update, not radar): In Meta ConnectionMind (arXiv 2608.10187, 2026-08-10), Llama 3.1-8B is used as a path reasoner over a graph. Netflix GenRec is a "scoring function" and ConnectionMind is a "graph path reasoner", so a new point to watch arose: the substance of role 3 (using an LLM as a ranker) differs by company. → services/facebook.md, topics/llm-in-recsys.md
  • Details: This LLM serves the online hot path for heavy users, who make up 5–10% of users, and the other 90–95% are handled by a Student GNN trained on meta-paths that the LLM discovered offline.
  • Details: The paper does not disclose the surface and writes "short-form video platform", but the graph schema suggests it is likely Facebook. Watch time rose +0.43%.
• 2026-08-28 additional evidence: Meta AMBER (arXiv 2608.25546, 2026-08-26) is an input tokenization mechanism that compresses each interaction event into one Event Token and feeds it into the LLM input. It serves billions of events per day on the Facebook surface with online encoding verified.
  • Details: It was written by 15 authors affiliated with "AI at Meta". Under the angle of Event Tokenization, it compresses the entire temporal snapshot of user, item, context, and outcome signals at each event into one compact Event Token and introduces it as a new token type in the LLM input.
  • Details: It improved the compute-quality Pareto frontier on industry-scale ranking and retrieval benchmarks, and it can be ported as a feature to non-LLM rankers (statistically significant). Online A/B figures are not disclosed in the abstract.
  • Details: It touches a different role from GenRec (scoring), ConnectionMind (graph reasoning), and MetaStrategy (offline policy compiler). That is, even within Meta, LLM-based recsys approaches have split into at least two axes (graph reasoner and input encoder).
• 2026-09-07 additional evidence: Spotify GLIDE and LinkedIn Follows both count as placing fine-tuned LLM embeddings in a candidate generation channel. With these two sources the thread grows to six companies and seven statements (Netflix, Yandex, Alibaba, Meta (2), Spotify, LinkedIn). Detailed analysis: daily/2026-09-07.md
  • Details: Spotify GLIDE (From Models to Products: LLMs for Recommendation at Spotify Scale, 2026-08-03 blog) is a candidate generator based on a 1B LLM and Semantic ID. Semantic ID (SID) is an identifier that represents an item as a sequence of semantic tokens. GLIDE was deployed on the Spotify Home podcast discovery surface and supplies candidates to the downstream ranking pipeline.
  • Details: In a 21-day A/B at ~20M impression scale, GLIDE gained non-habitual podcast listening +5.4% and new-show discovery +14.3%, while keeping latency and engagement requirements verified.
  • Details: LinkedIn Follows is sourced from Rebuilding LinkedIn's Follows Recommendations with LLM-Based Semantic Retrieval and Ranking (2026-08-24 blog). A fine-tuned bi-encoder LLM rewrites profiles as natural-language sentences (narrative prompting) and then embeds them. Creator embeddings are built in a nightly FAISS batch, and viewer embeddings are inferred in real time with Proxima at onboarding.
  • Details: Follows was deployed on two surfaces, the MyNetwork tab and Home Feed, and in the production A/B the follow rate rose statistically significantly, with the largest lift in the new member segment verified. The specific uplift percentage is not disclosed.
  • Details: The two sources complement Pinterest CLR (an approach that conditions the user tower of the Home Feed candidate generator on retrieval context). Pinterest approaches by conditioning, LinkedIn by verbalizing profiles, and Spotify by SID generation.
• 2026-10-12 additional evidence: In Pinterest User Journeys (From Activity to Intent, 2026-10-09 blog), an LLM generates user-level natural-language queries, and it is on the same side as MetaStrategy in that it is outside the synchronous path. Compared with clustering, the A/B showed email CTR +1.1% and push open +1.3% verified. Detailed analysis: daily/2026-10-12.md
  • Details: Qwen3-4B, distilled from a frontier teacher, reads 360-day activity logs and generates lists of journey names in batch. Each name becomes an embedding-based retrieval query on the notification surface.
• 2026-10-19 additional evidence: The Spotify and NVIDIA TensorRT LLM collaboration (Advancing Generative Recommendation Inference, 2026-10-16 blog) covers the serving cost of podcast LLM recommendation, which has been in production since 2026-01. It writes that, on Qwen3 1.7B, beam search cost fell by up to 54% and scoring cost by up to 7% verified (the comparison baseline is not clear in the text). Detailed analysis: daily/2026-10-19.md
  • Details: It separates a generation workload, which uses beam search to pull top Semantic ID sequences and convert them to catalog items, from a workload that scores multiple candidates in one request with candidate masking verified.
  • Details: It personalizes by projecting an existing user embedding into a single soft token verified.
  • Details: This statement treats an LLM in the serving path not as "does it work" but as "at what cost does it work."

Recommendation model scaling laws — independent confirmations from companies are accumulating

• What: Cross-checks the claim that recommendation models too "improve predictably like LLMs as data and parameters grow". The evidence so far comes in three strands: → sequence-transformer-ranking.md
  • Meta HSTU (a sequence model in the organic feed lineage) and Wukong verified
  • Netflix FM (a Foundation Model that builds member representations centrally) with its "consistent improvements" and the 2M→1B shadow study verified
  • Meta ads LLaTTE ("semantic features are a prerequisite") verified
• Why notable: If scaling holds, recommendation also becomes an arms race of "bigger models + longer sequences". However, the LLaTTE and HSTU claims are opposite on whether semantic features are a prerequisite, and this issue is not yet resolved.
• Watch for: We await counterexamples and cases that disclose parameter counts.
  • Details: Statements outside Meta and Netflix, in order of arXiv submission date, are ByteDance TM20K (2026-08-07), Ant Densing Law (2026-08-24), and Alibaba TransRetrieval (2026-08-26).
  • Details: After that came ByteDance ReST (2026-09-01) and HELIX (2026-09-29, [inferred] TikTok).
• 2026-08-28 additional evidence: Alibaba/Alimama TransRetrieval (arXiv 2608.25528, 2026-08-26) is the company that stated a scaling law after Meta HSTU, Wukong, and LLaTTE, and the first statement on the retrieval side. A one-month production A/B on the Alibaba display ads platform gave revenue +2.53% (95% CI [2.22, 2.70], p<0.0001) and RPM +1.28% verified.
  • Details: It was written jointly by 6 authors from Taobao & Tmall Group and 4 from RUC. It is an ads retrieval framework that suppresses token-norm divergence caused by field heterogeneity with weighted average aggregation, so that the Transformer scales.
  • Details: Target token compression cut compute per candidate by 85%, and position-style domain embedding unified multiple domains.
  • Details: The same A/B kept latency unchanged (P99 ~40ms), and split into 4 domains the gains were A +1.72% / B +2.51% / C +5.39% / D +0.87% verified.
  • Details: LLaTTE holds that "semantic features are needed" and HSTU holds that "scaling alone suffices". A third answer is added: "scaling is possible if field heterogeneity is handled explicitly."
• 2026-09-09 additional evidence (late discovery): Shopify Foundational Generative Recommender (Shopify Engineering, 2026-02-25) deployed to production for product recommendations an autoregressive sequence Transformer, based on HSTU and adapted to e-commerce. It is the first statement from the e-commerce domain confirming that the HSTU license is accessible enough for other companies to adopt in production. Detailed analysis: daily/2026-09-09.md
  • Details: The deployment surfaces are product recommendations in the Shop app and merchant storefronts.
  • Details: However, it does not disclose the identity or parameter scale of the baseline model, so the strength it adds to this thread is weak.
  • Details: It uses raw buyer event sequences (search, view, cart, purchase) directly as input. It combines RoPE-based rotary encoding, relative attention bias, batch-shared negatives, positive-aware hard negatives, and boosting-based recall optimization.
  • Details: In the production A/B it gained Shop orders +0.94% relative, high-quality CTR +5% relative, conversion rate +0.71% relative, and final product recall@4 +4.8% relative verified.
  • Details: CUDA kernel optimization made the training pipeline 7.3x faster.
• 2026-10-14 additional evidence (canon update, not radar): Meta GEM 2026 generation (blog, 2026-10-13) writes that it observed an LLM-like scaling law at its largest pretraining scale verified. A claim of 2× the performance of the previous version at the same compute is attached, so the question of this thread is shifting from "does scaling work" to "which architecture bends the curve better" inferred. → sequence-transformer-ranking.md, detailed analysis: daily/2026-10-14.md
  • Details: The sole basis of the inference above is that Meta presented the architecture comparison at the same compute.
  • Details: It swapped the backbone from InterFormer to Kunlun (arXiv 2602.10016) verified. The blog writes: "We observed an LLM-like scaling law at our largest pre-training scale with 2.5x model complexity growth from earlier versions of GEM." verified
• 2026-10-17 additional evidence: LinkedIn TransX (arXiv 2610.20913, 2026-10-14) reports that scaling returns differ by component (anisotropic). It makes a claim in the same direction as HELIX's (a ranking architecture that grows feature interaction and long behavior sequences together) "both axes must grow together", but with a different decomposition (encoder, memory, fusion).
  • Details: If memory resolution and fusion capacity are fixed, scaling encoder depth can saturate at moderate depth verified.
  • Details: As supervision increases, the optimal allocation shifts toward candidate-conditioned fusion verified.
  • Details: The required memory resolution stays roughly constant even as the model grows verified.
  • Details: See the TransX entry under "New this month" below.

Generative retrieval / semantic ID

• What: A family that represents items as sequences of semantic tokens rather than ID embeddings and replaces search with autoregressive generation. We confirmed only that Spotify gave a semantic tokenization talk at PRS 2025 verified.
  • Details: This talk is commonly misattributed to Netflix. See the open question in netflix.md.
• Why notable: If this approach holds, it replaces the entire embedding store and ANN (approximate nearest neighbor search) stack. Netflix GenPage (page generation) and GenRec (verbalization) are adjacent trends verified.
• Watch for: We await production deployment statements from follow-up work on Google/YouTube-family TIGER inferred. The original TIGER paper (RecSys 2023) has not yet been confirmed as a primary source in this repository, and we plan to check it at the next collection.
• 2026-08-11 additional evidence: HD-Rec (arXiv 2608.06997, 2026-08-07) is research on extending SID cross-domain, with Kuaishou's senior line (Ruiming Tang, Guorui Zhou, Han Li) participating. It has only public benchmark results and no production statement yet, so we record it only as a signal that the Kuaishou line has joined the SID family.
  • Details: To extend SID cross-domain it uses a hierarchical codebook (global coarse and domain-adaptive fine) together with a domain sparse MoE.
• 2026-08-12 additional evidence: Kuaishou PushDualGen (arXiv 2608.07989, 2026-08-08) is the next generation in the line from OneRec (Kuaishou's generative recommendation model) to OneRec-Thinking. It was deployed to production in Kuaishou push recommendation, and in the online A/B test the effective play rate rose +8.50% relative, the dissatisfaction rate fell −37.70% relative, and long-tail impressions also improved verified.
  • Details: To avoid the serving cost of CoT, it uses a 2-stage generation structure that first emits the SID and then attaches skippable explanation text (copy).
  • Details: Following HD-Rec, the Kuaishou SID line passed through the radar two days in a row. It is a signal that the SID line keeps evolving at the company level. It is also the second company, besides Netflix GenRec, to accumulate large-scale A/B statements.
• 2026-09-01 additional evidence: Alibaba PailitaoGR (arXiv 2608.26658, 2026-08-27 v1) is an attempt to extend SID generative retrieval from product search to image search, and is the first statement of the image-search modality in the SID line. It stated an average +13.8% improvement over baseline offline but did not disclose production A/B figures verified.
  • Details: All 9 authors are affiliated with Alibaba Group Hangzhou (we confirmed affiliations in the HTML version). Pailitao is Taobao's image search surface.
  • Details: Real photographs contain the search target, supporting evidence (e.g., text inside the photo), and irrelevant visual elements together. The paper addresses this problem along two axes.
  • Details: The target enhancer emphasizes only the search target tokens on the full image without cropping, implementing "Zooming without Cropping". The auxiliary enhancer selectively extracts relevant text evidence inside the image without OCR, implementing "Reading without OCR".
  • Details: Training combines on-policy distillation and a contrastive strategy. It states that the training and validation sets were sampled from real Pailitao image search logs.
  • Details: After Kuaishou HD-Rec/PushDualGen (short video, push) and Yandex Music Gryphon-v2/Sona (music), this is the fourth company on the SID line. With no production metrics, the strength it adds to the thread is weak.
• 2026-09-03 additional evidence: Tencent PCG TGR (arXiv 2609.00986, 2026-09-01 v1) is a 4-in-1 generative recommendation framework combining ranking, retrieval, slate generation, and reason tokens. It is a production claim counterpart to the Kuaishou OneRec family, the case with the second-largest company scale, and the first source to name HSTU and OneRec each as replacement targets within one frame. Detailed analysis: daily/2026-09-03.md
  • Details: A team of 13 TGR Team authors from Tencent Platform and Content Group (PCG) states A/B tests and production deployment on five Tencent PCG surfaces. The five surfaces are news feeds, short videos, long videos, music, and web novels, and the paper labels them as anonymous Scenario 1–5.
  • Details: CCFormer (Transformer ranking, replacing HSTU/DLRM) was fully launched in Scenario 1 and 2 verified. It gained CTR +3.57% and ad revenue up 1.71%, with half the GFLOPs of HSTU and training 2.21x faster verified.
  • Details: BARGE (hierarchical semantic ID generative retrieval) beat OneRec by 10.2–16.9% on offline Hit@5 verified. It was fully deployed as a generative retrieval channel (fully rolled out as a generative retrieval channel), with CTR +0.60% and reading time up 1.70% verified.
  • Details: HiGR (whole-slate generation) has slate quality +15.9–21.3% over OneRec and 5x inference speed. In the three-scenario A/B it gained watch time +1.22% and video views up 1.73%, but it does not yet state a production launch.
  • Details: In TGR-Reason, a Think model trained with LatentRec generates semantic-ID reason tokens offline and injects them as prompts at serving time. Cold-start new-user Hit@1 is +477.8%, and online effective consumption rose +1.75% and new-user conversion rose 13.09%.
  • Details: Items needing a deep dive are BARGE's OSQ-VAE codebook structure, the substantive difference between HiGR's PCRQ-VAE and OneRec SID, the Think model backbone and reason token corpus scale, and the Scenario→product mapping.
• 2026-09-07 additional evidence: Spotify GLIDE is a candidate generator in which a 1B LLM generates Semantic ID token sequences and maps them to catalog items, and SPDD is a controller that intervenes only at decoding time. If Kuaishou PushDualGen avoided serving cost with SID and skippable copy, SPDD is an alternative angle that leaves the serving model as is and changes only the decoding policy. Detailed analysis: daily/2026-09-07.md
  • Details: GLIDE (From Models to Products, 2026-08-03 blog) was deployed to production on the Spotify Home podcast surface verified. In a 21-day A/B at ~20M impression scale it gained non-habitual listening +5.4% and new-show discovery +14.3% verified.
  • Details: SPDD stands for Stochastic Primal-Dual Decoding (Balancing Multiple Objectives in Generative Recommendations with Adaptive Decoding, 2026-08-07 blog). It is a lightweight controller that intervenes only at the decoding (inference) time of a pretrained generative recommender.
  • Details: SPDD dynamically adjusts the weights of auxiliary objectives (e.g., relevance and diversity, relevance and novelty) as it progresses, and does not require retraining.
  • Details: In a production A/B on Spotify SASRec playlist ranking (~1 million users × 600,000 playlists), auxiliary stream-share was +5.44%, higher than fixed-weight's +3.60%, and user consumption did not decline verified.
• 2026-09-09 additional evidence (late discovery): Shopify Foundational Generative Recommender (Shopify Engineering, 2026-02-25) deployed an HSTU-based generative recommender to production with a large embedding table approach and no SID yet, and then stated semantic ID explicitly as the next step verified. Detailed analysis: daily/2026-09-09.md
  • Details: The reason given for naming it the next step is "products are modeled independently…very large embedding tables…bottleneck" verified.
  • Details: Kuaishou PushDualGen, Spotify GLIDE, and Tencent BARGE are follow-up improvements made after SID was already deployed. HD-Rec and PailitaoGR have no production deployment statement. Shopify, by contrast, made a statement from the stage just before SID adoption, which fits the observation that adoption proceeds sequentially among companies that feel the embedding table scale problem in practice.
• 2026-10-15 additional evidence: WeChat WeGR (arXiv 2610.16979, 2026-10-12 v1) is generative retrieval for WeChat e-commerce search, and per the abstract it gained orders +2.539% in a 2-month online A/B verified (abstract). Because it is a search surface it is distant from feeds, but following Tlow (a flow-based item tokenizer) in "Earlier additions (2026-08–09)" below, it is the same company's second SID production statement.
  • Details: We did not read the main text. In the same A/B, CTR was +1.508%, conversion rate +1.381%, and CPU and GPU usage −25.5% and −26.1%, and it was deployed on several WeChat search surfaces verified (abstract).
  • Details: It combines an SID tokenizer (CoCo-VAE) that injects collaborative signals and prevents codebook collapse with RL (DHPO) that solves the sparsity of exact-match SID rewards by widening the reward range per SID layer. This source adds the RL design axis of "partial rewards per SID layer" to this thread.
• 2026-10-16 counter-evidence: Shopify's Revisiting Generative vs. Dense Retrieval (arXiv 2610.19065, 2026-10-13 v1) is the first platform company source in this thread to directly rebut the superiority claim of generative retrieval (GR). It reports that most of the gap by which GR looks better than dense retrieval (DR) on Amazon Toys public data comes from popularity, and argues that mixing the same static item value once into DR scores reverses the gap verified.
  • Details: 3 authors are affiliated with Shopify (1 also holds an appointment at University of Toronto, which we confirmed in the HTML version). There is no online A/B; it is an offline comparison.
  • Details: On Shopify's in-house large-scale e-commerce data too, the comparison result is written as follows: "a 14B SID-LLM still underperforms a 47× smaller dense encoder" verified
  • Details: Gen4+N is the number these notes attach in sequence to X commits after the Gen4 release. X also runs SID retrieval alongside other channels at Gen4+34, so it is evidence for coexistence rather than replacement inferred. The basis is that in X code the SID channel was added as one source without replacing the existing two-tower (a retrieval model that embeds users and items separately).
• 2026-10-21 additional evidence: Alibaba Taobao and Tmall's Don't Rewrite the SID Map (arXiv 2610.26105, 2026-10-19 v1) is the first operational statement to quantify that reassigning SID breaks the decoding path of an already deployed generative retrieval. A warm-start update that freezes existing item SIDs and inserts only new items gave GMV +1.16% in a 3-week A/B verified. Detailed analysis: daily/2026-10-21.md
  • Details: All 7 authors are affiliated with Taobao & Tmall Group of Alibaba (confirmed in the HTML version). The surface is e-commerce search.
  • Details: It reports that in a 67-day log experiment, with beam 1 the mapping update dropped HitRate by 41.78–62.05%, while with beam 1000 it dropped only 8.33–10.85%, so a large beam masks the degradation verified (abstract).
  • Details: In the same A/B, UCTCVR was +0.29% verified (main text). The improvements of +9% for existing items and +109% for new items are offline figures.
• 2026-10-21 additional evidence (canon update, not radar): Meta GenRecs Tokenization (blog, 2026-10-19) replaced item representations in ads ranking and retrieval with token sequences of a "stable semantic vocabulary", and writes that together with GEM it gave conversion +0.35% verified. Together with the Alibaba paper the same day, SID vocabulary stability has risen to an operational issue inferred. → llm-in-recsys.md, detailed analysis: daily/2026-10-21.md
  • Details: The basis of the inference above is that each of the two sources puts "stable vocabulary" and "topology preservation" forward as a core design principle. However, Meta's "stable" is said in contrast to item IDs that disappear quickly, and it does not directly address the SID reassignment problem Alibaba covered.

Learned index (index-as-model) — does it move to organic feeds?

• What: A trend of making the index a model tensor and a training target rather than a static data structure. Meta Andromeda/HILL is ads literature, and the blog post describes SilverTorch as a UGC feed and video retrieval system verified. → embedding-retrieval.md
• Why notable: There are many precedents of stacks verified in ads moving to organic feeds. There is no first-party statement that a learned index (Andromeda/HILL) was applied to an organic feed inferred (this means no statement exists; the corpus examined refers to topics).
• Watch for: Whether a statement appears that the same stack was applied to the organic feed, and whether a company other than Meta adopts a learned index.

New this month

TransX (LinkedIn) — separately scaled sequence encoder, behavior memory, and fusion; split serving gives ~5× per-request FLOPs vs. HSTU family

• What: TransX is a framework that splits a ranking model into three independently scalable components and also separates serving into three stages.
  • Details: The three components are:
    • User behavior sequence encoding
    • Candidate-independent compressed behavior memory: recent behavior is held at high resolution and past behavior as a low-resolution summary.
    • Candidate-conditioned cross-attention fusion
  • Details: The three serving stages are offline bulk state construction, nearline incremental updates, and online prediction.
• Why notable: In an A/B on a "social recommendation application" with ~20K requests per second and 0.5K–1.2K candidates per request, Base recorded CTR +6.02% and conversion +4.45% over the sequence-model baseline verified. At the same total serving cost, it uses about 5× the FLOPs of the HSTU-family baseline verified.
  • Details: By variant, the retrieval variant recorded CTR +2.96% over two-tower, and the ranking variant recorded conversion +1.33% over DLRM verified.
  • Details: The body does not state explicitly whether the surface is the LinkedIn feed inferred. The only evidence is the affiliation and the phrase "social recommendation".
• Evidence: verified arXiv 2610.20913, 2026-10-14 v1. There are 8 authors, and apart from the corresponding author the email domain is @linkedin.com (confirmed in the HTML version).
• Status: Watching (deep dive needed). LinkedIn is not a tracked service. What to confirm is whether a full rollout happened, the A/B duration, the memory KV size, and the nearline update interval. The detailed analysis is in daily/2026-10-17.md.

Pinterest User Journeys — Qwen3-4B reads activity logs and generates journey names, deployed as notification retrieval queries

• What: Pinterest infers a user's long-term goal (journey) with a single generation from a fine-tuned Qwen3-4B instead of 5-stage keyword clustering. The output name becomes a query for embedding-based retrieval (EBR) on the notification surface, and the existing Pin ranking orders the results.
  • Details: Serving handles ~775 rps on ~100 GPUs, with p95 at ~3.0 seconds.
• Why notable: In a US·Canada A/B, it recorded email CTR +1.1% and push open +1.3% over clustering, and the changes in sitewide engagement and retention were also significant verified. The multilingual rollout showed the same direction verified.
  • Details: The post states that the prompt replaced the diversifier and the negative filter.
• Evidence: verified From Activity to Intent, 2026-10-09, Pinterest Engineering (blog post).
• Status: Watching. Pinterest is not a tracked service. SID input and feedback-as-edit preference optimization are under experiment, and a follow-up post has been announced. The detailed analysis is in daily/2026-10-12.md.

GEAR (Douyin / ByteDance) — end-to-end generative ad retrieval serving hundreds of millions of DAU on Douyin Ads; fixes representation collapse and item collision

• What: GEAR is a generative ad retrieval framework. It optimizes the tokenizer, generator, and reranker together end to end, targeting both bottlenecks of industrial-scale generative retrieval at once.
  • Details: Generative retrieval represents an item as a sequence of semantic tokens (Semantic ID) rather than an ID embedding, and replaces search with autoregressive generation. At industrial scale there are two known bottlenecks.
    • Representation collapse: the tokenizer converges degenerately under distribution shift, which makes end-to-end adaptation unstable.
    • Item collision: in a huge candidate pool, different items share the same token sequence, which lowers retrieval precision.
  • Details: The two problems are coupled, so enlarging the codebook capacity to mitigate collisions worsens the collapse.
  • Details: GEAR handles the two bottlenecks with three mechanisms.
    • BasisVQ reparameterizes the codebook with an orthogonal basis. This enables global gradient sharing and rigid spatial rotation of the latent space, which stabilizes the gradient dynamics.
    • It extends this to a prefix-aware BasisRQ, which increases the codebook's expressiveness within the same asymptotic time complexity.
    • A context-conditioned reranking head is placed inside the generation process and separates colliding items with minimal overhead.
• Why notable: It is served in production to hundreds of millions of DAU on Douyin Ads verified. Specific metrics are not disclosed in the abstract.
  • Details: Direct quote from the paper: "It currently serves hundreds of millions of daily active users on Douyin Ads, yielding substantial empirical improvements in extensive online A/B tests".
  • Details: In this repository's corpus, the cases of generative retrieval landing on ad surfaces so far were two Alibaba sources.
    • Display ads TransRetrieval (2026-08-28, Transformer scaling): it was a scaling axis, not SID generation.
    • PailitaoGR (2026-09-01, image search): production A/B figures were not disclosed.
  • Details: GEAR is an explicit statement that SID generative retrieval has landed on an ad surface at Douyin, the fourth company. It is also the first source to address the two structural problems, representation collapse and item collision, together.
  • Details: In the 2026-09 industrial generative retrieval angle, this is the fourth statement. The earlier three are:
    • Snap GRP v0.1 (2026-09-29, encoder-decoder unification)
    • Tencent PCG TGR (2026-09-01, 4-in-1 generative recsys)
    • OneTrans-V2 (2026-09-23, 3-stage cascade unification)
• Evidence: verified arXiv 2609.39327, 2026-09-30 v1. All 11 authors are affiliated with Douyin (ByteDance), and the affiliations were confirmed in the HTML version.
• Status: Watching (deep dive needed). Douyin (ByteDance) is not a tracked service, so we do not create a services canon. The detailed analysis is in daily/2026-10-02.md.
  • Details: There are five deep-dive targets:
    • The formula for BasisVQ's orthogonal basis reparameterization
    • Mathematical details of the prefix-aware BasisRQ extension
    • How the context-conditioned reranking head is integrated into the generation process
    • Specific A/B metrics (the abstract has no figures)
    • The joint training objective of the three components: tokenizer, generator, and reranker
  • Details: The checkpoint to watch is whether a second company besides ByteDance handles generative retrieval on an ad surface with end-to-end tokenizer optimization rather than batch.

HELIX ([inferred] TikTok) — unified ranking architecture that scales feature interaction and long user-behavior sequences together, deployed to TikTok e-commerce production with GMV per user ~+6% in the A/B test

• What: HELIX is a "purified and unified" architecture that scales an industrial recommendation ranking model along two axes together. The paper observes that scaling only one axis gives a suboptimal scaling-law slope, and proposes a framework that expands both axes together.
  • Details: Industrial rankers usually scale along the following two axes:
    • heterogeneous feature interaction: combinations of user, item, context, and cross features
    • long/multi-type user-behavior sequence modeling
• Why notable: It was deployed in TikTok e-commerce production and recorded GMV per user +~6% in an A/B verified.
  • Details: Direct quote from the paper: "In online A/B tests, it achieves an approximately 6% increase in e-commerce video GMV per user".
  • Details: Offline, CTR AUC, CVR AUC, and other ranking metrics improved consistently.
  • Details: The paper's core is a view suggesting that the two existing scaling families are each suboptimal. The two families are the sequence-only scaling of the Meta HSTU / GEM family and the feature-interaction scaling of the DCN·xDeepFM family.
  • Details: It is a scaling-law statement that follows Meta HSTU, Meta LLaTTE, Alibaba TransRetrieval, and others, and the first case to explicitly call for expanding both axes at once.
  • Details: In TikTok e-commerce, two families have been released so far. One is the EGA family on ad auctions, and the other is the GreenLightning family, a single inference service. This is the first time a ranking architecture unifying feature interaction and long sequences has appeared with a formal name.
• Evidence: verified (abstract wording) arXiv 2609.37183, 2026-09-29 v1, 21 authors. The v1 HTML rendering returns 404, so we judge the author affiliation to be TikTok (ByteDance) based on the abstract's phrase "Deployed in TikTok's e-commerce recommendation system" inferred.
• Status: Watching (deep dive needed, awaiting affiliation confirmation). TikTok/ByteDance is not a tracked service, so we do not create a services canon. If v2, a restored HTML rendering, or a company blog announcement appears, promote [inferred] to [verified]. The detailed analysis is in daily/2026-10-01.md.
  • Details: The deep-dive targets are the detailed architecture and the interaction analysis of the two axes. For the detailed architecture, we look at which interaction structure and which sequence encoder are used, and what the measured scaling-law slopes of the two axes are.
  • Details: The checkpoints to watch are how TikTok e-commerce was built on top of GreenLightning, and whether the next company (Kuaishou, Meta, Alibaba) states simultaneous scaling of both axes.

GRP v0.1 (Snap) — single encoder-decoder unifying retrieval, ranking, and reward; production A/B view time +0.82%, shares +2.56%, serving latency −69%

• What: GRP (presumably Generative Recommendation Platform) is a generative recommendation framework. It unifies the multi-stage retrieval → pre-rank → fine-rank → reward cascade of industrial recommendation into a single encoder-decoder model. The same model generates candidates, scores them, and provides the reward signal for RL training.
  • Details: The framework has two pillars:
    • multimodal Semantic ID generation: candidate items are represented as discrete SID token sequences, which an autoregressive decoder generates.
    • joint Multi-Head Prediction (MHP): it handles candidate scoring and also serves as the frozen reward for RL post-training. This RL post-training is mGRPO (reference-anchored margin GRPO).
• Why notable: In a Snap production A/B, the combination of bypass and source replacement recorded view time +0.82% and shares +2.56%, and platform-level guardrails were neutral verified. Serving optimization brought end-to-end retrieval latency −69%.
  • Details: Direct quote from the paper: "0.82% in view time and 2.56% in shares, with neutral platform-level guardrails".
  • Details: The retrieval-only combination gave view time +0.46% and shares +0.77% verified. The retrieval bypass combination gave "+0.24% views, −0.24% view time, +2.30% shares (p<0.05)".
  • Details: This is the third statement of an encoder-decoder unified generative recsys and the first from Snap. The earlier two statements are:
    • OneTrans-V2: unified the retrieval, pre-rank, and fine-rank cascade into a single Transformer and recorded GMV +9.74% (2026-09-26 radar).
    • Tencent PCG TGR: handles ranking, retrieval, slate generation, and reason tokens in one frame, and was deployed on 5 surfaces.
  • Details: The mGRPO joint MHP axis, which handles even the reward inside the same model, and the multimodal SID generation axis were absent from the earlier two statements.
  • Details: Serving latency −69% is a large reduction versus the multi-stage cascade. It signals that whether the encoder-decoder unification also justifies the latency budget is the key criterion for adopting industrial-scale generative recsys.
• Evidence: verified arXiv 2609.36688, 2026-09-29 v1, 22 authors. The v1 HTML rendered the organization name "Snap Inc.", and the mapping of individual authors was not rendered.
• Status: Watching (deep dive needed). Snap is not a tracked service, so we do not create a services canon. The detailed analysis is in daily/2026-10-01.md.
  • Details: The deep dive reads the original text of 26 pages · 11 tables to confirm the following:
    • The exact computation of the mGRPO reference anchor
    • How the joint MHP is used as the frozen reward
    • The multimodal SID codebook structure
    • Deployment conditions of the three A/B combinations: retrieval-only vs. bypass vs. combined
  • Details: The checkpoints to watch are how Snap extended TIGER-family SID to multimodal, and which axes differ architecturally from OneTrans-V2 and TGR.

Earlier additions (2026-08~09)

PEAR (anonymous e-commerce, [inferred] Alibaba line) — progressive-verifier autoresearch for industrial search, 4-layer verifier ladder; Main Order/DAU +2.7336% · +3.2957% in two A/B tests

• What: PEAR is a research orchestration framework that automatically explores candidate policies (candidate modification) in industrial search systems and verifies them with A/B tests. The name is short for Progressive Evidence-Based AutoResearch.
  • Details: The paper claims to solve two problems of existing approaches along two axes:
    • Under non-stationary traffic, the keep-if-better rule mistakes transient gains for persistent gains.
    • Evaluation that relies on individual signals cannot integrate evidence to guide exploration.
  • Details: The first axis is Evidence-driven AutoResearch. Each strategy task has its own independent hypothesis-guided research state within a predefined objective and intervention scope. This state evolves through a Plan → Execute → Evaluate → Update loop and updates hypotheses by interpreting experiment results in a context-aware way.
  • Details: The second axis is the Confidence-Gated Verifier Ladder. It splits evaluation into 4 layers in order of trust, and a unified confidence-based gate promotes to the next layer only when a statistically significant positive is secured. The layers are, in order, Offline Replay → Shadow-Traffic Evaluation → Rapid Online Evaluation → Decision-Grade Online Evaluation.
• Why notable: In two A/B tests on a real industrial search system, Main Order/DAU was recorded at +2.7336% and +3.2957% verified.
  • Details: The precision given down to a two-digit position, and the metric "Main Order/DAU" (orders / DAU), are standard KPIs of e-commerce surfaces.
  • Details: This is the third source in 2026-09 where the research automation angle appeared with an explicit name. The three sources have different roles:
    • Meta VDD EvoPilot: search retrieval angle
    • Kuaishou presumably AgentX-Model: the whole recommender research loop angle
    • PEAR: verifier layer angle of industrial search automation
  • Details: While AgentX-Model decomposed research orchestration into dual agents (Research and Model), PEAR decomposed evaluation confidence into layers. The two axes are not mutually exclusive, so we watch how far production statements in this angle extend.
• Evidence: verified arXiv 2609.35031, 2026-09-28 v1. The affiliation of the 9 authors is given only as "Global E-Commerce Agentic Search Team", with no company name. The team branding and the author list (Yu Gong and others) match the AliExpress/Lazada line of Alibaba International Digital Commerce, so we take it as this line inferred.
• Status: Watching (deep dive needed, awaiting affiliation confirmation). Alibaba/JD International domains are not tracked services, so we do not create a services canon. The detailed analysis is in daily/2026-09-30.md.
  • Details: If v2, a company blog, or an official counterpart post on the Alibaba/JD Tech Blog appears, promote inferred to verified.
  • Details: The deep-dive targets are the per-stage traffic size, gate threshold, and statistical method of the 4-layer verifier ladder. The next checkpoints are an axis comparison with AgentX-Model, and whether the verifier trust-layer decomposition also appears at other companies.

KuaFu (presumably Tencent) — two-axis per-item behavior compression replaces the profiling pipeline; 10 months at 1 billion weekly users and 100K QPM, GMV +1.37% (95% CI [0.71%, 2.03%]), 190 GPUs saved

• What: KuaFu is a representation layer that compresses raw behavior sequences for the user profiling pipeline of industrial-scale recommender systems. Instead of the practice where each task model trims sequences separately, it takes one behavior item as the minimal compression unit and compresses each item independently. Downstream profiling tasks share this cache.
  • Details: A two-axis projector shrinks each item embedding to 2–4 tokens × width 128–256. As a result, the per-item cache drops from 10 KB → 0.5 KB (10× on the token axis, 20× on the width axis).
  • Details: Training stacks in 4 stages:
    • understanding: reconstruction
    • post-training: layered-difficulty data of 42 task types
    • co-training: handles compression and generation together, and uses a sequence-length curriculum for reconstruction.
    • hallucination-aware RL: the compression and projector are frozen, and only the decoder is updated with DAPO.
  • Details: The RL reward applies harm weights to four hallucination types. The weights are fabrication 0.45, date misattribution 0.24, broken logic 0.24, and missed detection 0.07.
• Why notable: It was deployed in a profiling pipeline at 1 billion weekly users and aggregate 100K QPM. An additional 10 months of observation gave GMV +1.37% (95% CI [0.71%, 2.03%]). Offline, it recorded per-GPU throughput +37–350% and 190 GPUs saved verified.
  • Details: The A/B was by unique visitor. After a 0% → 100% rollout within 4 weeks, a 5% holdout was kept to continue observation, and both arms were under the same GPU and storage budget.
  • Details: Offline, it matched or exceeded the uncompressed single-task production model on all 5 headline metrics of the 4 production profiling tasks verified.
  • Details: On public benchmarks, it recorded up to +17.7 EM over existing compressors on MRQA out-of-domain. On RecBench, KuaFu 4B scored 1.90 points higher than 8B.
  • Details: Hallucination-aware RL lowered the fabrication hit rate from 33.5 → 4.5 (~29 pp decrease) and raised the fully-correct rate by +34.0 pp.
  • Details: In this repository's corpus, ChronicleRec (Tencent PCG) focuses on the ranker's input cache (target-independent Chronicle Token). KuaFu has a different role because it rebuilds the profiling pipeline itself on a shared compression.
  • Details: It is also the first large-scale production A/B statement on the profiling axis, in contrast to Meta AMBER, which addressed input tokenization, and Comcast LLM profiling, an offline factorial comparison.
• Evidence: arXiv 2609.31045, 2026-09-25 v1. The affiliations of the 11 authors are not rendered in the HTML version. The body states "deployed on the Tencent advertising and recommendation platform for ten months", and author Jie Jiang matches the author list of ChronicleRec (Tencent PCG), so we take it as Tencent's ads and recommendation line inferred.
  • Details: The authors are Jiahao Hui, Lin Zhu, Yishen Hu, Jingdong Shu, Zetai Jiang, Xining Ran, Ben Tan, Yeshou Cai, Gong Chen, Haijie Gu, and Jie Jiang.
• Status: Watching (deep dive needed, awaiting affiliation confirmation). Tencent is not a tracked service, so we do not create a services canon. If v2 or a company blog announcement appears, promote [inferred] to [verified]. The detailed analysis is in daily/2026-09-29.md.
  • Details: There are three deep-dive targets:
    • The exact definitions of the other 3 profiling tasks besides Content-interest summarization
    • The data size and loss function of each of the 4 training stages
    • How hallucination-aware RL balances against refusal drift
  • Details: Tencent's ads and recommendation line is pushing lifelong user modeling with ChronicleRec in the ranker role and KuaFu in the profiling role. We watch whether this direction expands further.

Pinterest 3-Tower Engagement Co-Train — ads lightweight ranking beyond two-tower; CTR, gCTR30, oCTR up ~1% each, CPC down ~1%, launched

• What: Pinterest's standard ads lightweight ranking model in its ads stack. It keeps the query and Pin towers of the existing two-tower model and adds a cross tower and an inter module, so one model outputs a fast two-tower score for all candidates and an accurate three-tower score for the top subset at the same time.
  • Details: The cross tower takes three kinds of user sequences and four kinds of candidate features as input. A query token built from the candidate features performs cross-attention over the user sequences.
    • User sequences: organic engagement, ads engagement, search history
    • Candidate features: advertiser ID, campaign ID, GraphSAGE embedding, Pin PinnerSAGE embedding
  • Details: The inter module combines the query, Pin, and cross embeddings with a shared-bottom MLP and generates the logits of three tasks.
    • pCTR: click probability
    • pGCTR30: probability of a good click that lasts 30 seconds or more
    • pOCTR: probability of an outbound click to the advertiser's site
  • Details: Serving is split into two stages.
    • Stage 1 scores all candidates with the co-trained two-tower head.
    • Stage 2 rescores only the top 40K by utility with the three-tower, then blends the two scores to pick the final top-K. This top range is at roughly the 70th percentile on Home Feed and Related Pins and the 90th percentile on Search.
  • Details: The loss gives a strong weight to the main term (three-tower) and an auxiliary weight to the co-train term (two-tower).
    • The formula is L = (100·L_main_ctr + 5·L_main_gctr30 + 1·L_main_octr) + (10·L_cotrain_ctr + 5·L_cotrain_gctr30 + 1·L_cotrain_octr).
• Why notable: In the standard ads production A/B, CTR, gCTR30, and oCTR each rose ~1% and CPC fell ~1%, and the model launched with a modest increase in GPU spend verified.
  • Details: Offline loss fell ~30% on the three-tower prediction. The baseline is the existing production two-tower model.
  • Details: The serving optimization results are quantified:
    • BF16 adoption: P99 103ms → 56ms at 50K candidates, 45ms → 26ms at 20K candidates
    • Late expansion of user features: -15ms at 100K candidates
    • pre-layer normalization: -3ms
    • Moving Pin preprocessing into the Pin tower: -2ms
  • Details: Combined, these cut P99 model inference from over 200ms to 30ms, and the latency increase over production was only 1-2ms.
  • Details: In the repository corpus, Pinterest CLR covered the candidate generation axis of organic Home Feed, and PinDCO covered the ads creative selection axis. 3-Tower Co-Train is a statement on the ads lightweight ranking axis, so its role differs.
  • Details: In particular, the pattern of "putting fast + slow two heads in one model by co-train and splitting serving into two stages" is the first production claim in the repository's lightweight ranking role.
  • Details: It reads as a sequence: the Nexus serving stack (Part 1) separated candidate generation and scoring, and this Part 2 expanded scoring on top of it.
• Evidence: verified Beyond Two Towers: Launching the 3-Tower Engagement Co-Train Model (Part 2), 2026-09-17, Pinterest Engineering. There are 8 authors, and the last author is Manager II, Ads Lightweight Ranking.
  • Details: The authors are Longyu Zhao, Gwendolyn Zhao, Peng Yan, Yuanlu Bai, Yuan Wang, Yao Cheng, Ang Xu, and Zhaohong Han.
• Status: Watching. Pinterest is not a tracked service, so no services canon is created. Part 1 (Nexus) has not yet been covered in the repository. A detailed analysis is in daily/2026-09-28.md.
  • Details: The next checkpoint to watch is Nexus's architecture that separates candidate generation and scoring. We also watch whether a second company states a co-train fast and slow pattern in the ads lightweight ranking role. We also look for any follow-up statement from the company on the direction after 3-Tower (longer sequences, expansion to objectives beyond engagement).

AgentX-Model (Kuaishou, inferred) — dual-agent recommender research loop; 5 production A/Bs: acquisition +10-15%, ad spend +15-20%, watch time +0.3-0.8%, FLOPs −10%

• What: AgentX-Model is a framework that automates the research cycle of a production recommender team (proposal → experiment → next question) with two agents. The Research Agent produces independently reviewed proposals from papers and earlier experiments, and the Model Agent turns each proposal into code, measurements, and unresolved questions.
  • Details: The Research Agent receives the results again, selects the starting implementation, and drafts the next question, so follow-up experiments build on earlier results.
  • Details: There are 4 research actions:
    • Reproduce: reproduction
    • Follow-up: extension
    • Composition: combination
    • Diagnose: collects evidence for diagnosis and repair candidates for problems such as issues found in online evaluation or prediction bias measured by PCOC.
• Why notable: In production A/B, it recorded acquisition efficiency +10-15%, target-segment advertising spend +15-20%, and watch time +0.3-0.8% (mean over 5 tests). The watch-time model cut FLOPs and parameters by about 10%. And 560 of 636 completed model-changing experiments (88%) beat the business baseline on offline AUC verified.
  • Details: In 2026-09, this is the second case where a research-automation angle was explicitly named and published. The first is Meta VDD MBR·EvoPilot's "37-day human-gated autoresearch campaign caught an evaluator output-depth 3000/600 defect and recovered offline +3.20 pp".
  • Details: If Meta's angle automated search retrieval improvement, AgentX-Model's angle automates the entire recommendation model research loop.
  • Details: It also released a dependency-aware historical-replay benchmark. The initial result states that "when the agent has already analyzed and selected concrete candidates, more complex scheduling gives no consistent efficiency gain".
• Evidence: arXiv 2609.30001, 2026-09-24 v1, 37-page technical report. The affiliations of the 23 authors are not rendered in the v1 HTML version. The authors include Peng Jiang (Kuaishou VP), Wenwu Ou, and Zhaojie Liu, and the earlier AgentX v1 also includes Kun Gai and Han Li, so we infer a Kuaishou line inferred.
• Status: Watching (deep dive needed, awaiting affiliation confirmation). Kuaishou is not a tracked service, so no services canon is created. When v2, a conference acceptance, or a company blog post appears, promote [inferred] to [verified]. A detailed analysis is in daily/2026-09-27.md.
  • Details: The novelty of this source lies on the organization axis, not the ranking or retrieval mechanism axis, so it is far from the feed pipeline itself, which is the repository's main interest. Therefore, a deep dive starts when a person directs it.

OneTrans-V2 — unifies the industrial recommendation retrieval, pre-rank, and fine-rank cascade in one Transformer; GMV +9.74%, 3.2× throughput

• What: OneTrans-V2 is an approach that unifies the 3-stage cascade of industrial recommender systems (retrieval → pre-rank → fine-rank) into one Transformer. The user behavior sequence is encoded only once as shared context, while the candidate features and computation of each stage are kept.
  • Details: Joint training makes the three stages reinforce each other, and in-model knowledge distillation flows from fine-rank to pre-rank.
  • Details: The backbone scales with sparse MoE, and μP-style parameterization stabilizes scaling.
  • Details: Instead of keeping a separate retrieval channel per objective, it unifies them with DCGR (Decision-Conditioned Generative Retrieval). DCGR first predicts a decision prefix, then generates items conditioned on that prefix.
  • Details: SNT (Sequence-Native Training) amortizes sequence encoding across impressions.
• Why notable: In production A/B, it recorded GMV +9.74% and 3.2× throughput at the same hardware budget verified.
  • Details: In the repository corpus, recent production statements on cross-stage unification were UNIQUE (Baidu, RecSys 2026) and UniRec (Kuaishou). OneTrans-V2 adds a third angle: binding all 3 stages in one Transformer.
  • Details: Its A/B figure is also much larger than those of the two papers. UniRec reports app usage duration +0.616%, and UNIQUE did not publish an A/B figure in its abstract.
• Evidence: verified arXiv 2609.28589, 2026-09-23 v1. There are 21 authors (Hannan Cao, Jun Guo, Haolei Pei, et al.), and Appendix A of the HTML version splits them into a 12-person Foundation Team and a 9-person Architecture Team. There is no affiliation listing, and the corresponding author's email domain is bytedance.com, so we infer it is a ByteDance team source inferred (checked 2026-10-19).
• Status: Watching (awaiting affiliation confirmation). The HTML version lists no affiliation, so we include it with the affiliation inferred only from the corresponding author's domain. A detailed analysis is in daily/2026-09-26.md.
  • Details: If a conference acceptance notice or v2 confirms the affiliation, we revisit the canon update. If a second company states that it put full 3-stage unification into production, it becomes a new topic promotion candidate.

Comcast Technology AI — LLM-based user profiling for streaming recommendation (offline production comparison)

• What: A paper that compares, factorially, two paradigms for building user embeddings on a production content-based recommendation dataset from Comcast Peacock/Xfinity. The LLM approach is far more expensive, so the core question is "when is that cost justified." There is no online A/B. The paper also does not state which Peacock/Xfinity product dataset it is.
  • Details: The two paradigms compared are as follows:
    • aggregate: numerically aggregates the user's recent vs historical item embeddings.
    • LLM-generated: feeds the user history to an LLM to get a natural-language summary, then vectorizes it with a text encoder.
  • Details: It compares 4 combinations: the two approaches × temporal disentanglement (recent vs historical) on/off.
  • Details: It states that results differ by user behavior type, and that different trade-offs are observed on the accuracy and beyond-accuracy (diversity, etc.) axes.
• Why notable: Comcast (Peacock streaming, Xfinity) was not a tracked service of this repository, and this is the first public signal of its production stack.
  • Details: The approach of generating user profiles with an LLM differs in angle from LLM applications in search and recommendation such as Spotify SPDD and GLIDE. It gets the user representation itself as a natural-language summary, not through decoding or retrieval.
  • Details: The question the authors answer with the factorial comparison, "how much and when is an LLM profile better than aggregate," is in a form that other streaming services (Disney+, Paramount+, Hulu, etc.) that want to put LLM profiles into production can refer to.
• Evidence: verified arXiv 2609.27183, 2026-09-23 v1. We confirmed the author affiliations in the HTML version. 3 of the 4 authors are at Comcast Technology AI (Sunnyvale, Washington DC), and first author Milad Sabouri has dual affiliations with DePaul University and Comcast Technology AI.
  • Details: The three Comcast Technology AI authors are Neeraj Sharma, Sardar Hamidian, and Shaghayegh Agah.
• Status: Watching. With no online A/B, the strength of production verification is low. A detailed analysis is in daily/2026-09-25.md.
  • Details: The checkpoints to watch are whether Comcast later publishes production deployment and A/B results, and whether the LLM profile direction spreads to other streaming services.

Spotify Robust Fusion — QSS + dual-sample dropout to curb behavioral over-reliance in an LLM reranker

• What: Robust Fusion is a training procedure that suppresses the shortcut learning that arises when the LLM cross-encoder reranker in Spotify personalized search fuses semantic and behavioral signals. The authors propose a deterministic dual-sample feature-dropout.
  • Details: QSS (Query Slice Stats) is an interaction-derived feature that summarizes the past success rate of a query-candidate pair. Simply putting this feature in the prompt raises ranking on slices that have QSS, but performance drops on rare/unseen queries that have no QSS.
  • Details: dual-sample feature-dropout feeds each training example in twice: once with QSS and once with QSS removed.
  • Details: Offline, ranking quality rose +13.3% when QSS was present, and on the QSS-removed slice it improved +4.0% over naive QSS. In the online A/B, both QSS-aware variants recorded live search success ≈+2%.
  • Details: It states that the aggregate test cannot distinguish dual-sample from features-only training, but that the cold-start comparison agrees in direction with the offline results.
• Why notable: A statement that a cross-encoder LLM reranker was validated by A/B in production search is itself uncommon. Many papers stop at offline metrics.
  • Details: Spotify deployed an LLM candidate generator in GLIDE and inference-time decoding in SPDD. This time even the countermeasure for LLM reranker shortcut learning moves to a production A/B.
  • Details: The dual-sample dropout lever is a production methodology that secures robustness on slices where behavioral features are absent or stale while keeping the gain when they are present. It is in a form directly applicable to other companies introducing an LLM reranker.
• Evidence: verified arXiv 2609.25825, 2026-09-22 v1, USRW Workshop @ RecSys 2026 accepted. The email domain of all 7 authors is @spotify.com (confirmed in the HTML version).
• Status: Watching. Spotify is not a tracked service, so no services canon is created. A detailed analysis is in daily/2026-09-24.md.
  • Details: The checkpoint to watch is whether dropout of behavioral features is migrated to LLM rerankers for organic feed recommendation. We also watch whether similar dual-sample training is published by other search engineering organizations (Amazon, eBay, Booking, etc.).

UNIQUE (Baidu + Beihang + CityU) — unified retrieval + ranking with single-layer flat quantization in an industrial mobile feed, RecSys 2026 accepted

• What: UNIQUE targets two problems of the retrieval-ranking pipeline in industrial mobile feeds with improvements on two axes. It unifies generative code-based retrieval and target-aware ranking into an early-fusion single architecture trained end-to-end on a shared representation, while still keeping candidate generation efficient.
  • Details: The two problems are as follows:
    • instability of hierarchical (RQ-VAE family) quantization in candidate retrieval
    • information loss between separated retrieval and ranking stages
  • Details: The authors write that this approach targets long-tail and cold-start problems while keeping candidate generation efficient, and they report P99 89ms and online inference MFU 44.23% as serving measurements.
• Why notable: Baidu is not a tracked service but is a platform company that operates a large-scale industrial mobile feed (Baidu App). This source takes the angle of replacing the hierarchical quantization instability of the RQ-VAE family with single-layer flat quantization.
  • Details: In the repository's generative retrieval / semantic ID thread, most companies that adopted SID (Google PLUM, Kuaishou OneRec, Spotify GLIDE, Meta AMBER, Yandex Gryphon, etc.) use hierarchical codebooks. UNIQUE goes in a different direction.
  • Details: The angle of binding retrieval and ranking by early fusion also differs from Kuaishou UniRec's cross-stage joint learning. UniRec softly aligns the two stages, and UNIQUE unifies them in one architecture. This difference adds one more data point.
• Evidence: verified arXiv 2609.23718, 2026-09-20 v1, RecSys 2026 accepted. We confirmed the author affiliations in the HTML version: Baidu Inc., Beihang University, and Hong Kong Institute of AI for Science (CityU).
• Status: Watching (deep dive needed). The abstract states neither a deployment claim nor A/B figures, so the main text needs a close read. A detailed analysis is in daily/2026-09-23.md.
  • Details: The next checkpoint to watch is whether the RecSys 2026 presentation names the Baidu App feed surface and gives production figures, and whether other companies adopt single-layer flat quantization.

MuSeR (Kuaishou) — multi-interest sequence retrieval on MGS that compresses 10^4–10^5-scale user histories with hierarchical temporal compression

• What: MuSeR (Multi-interest Sequence Representation) is a retrieval framework built on the MGS system deployed at Kuaishou, and it combines three axes.
  • Details: The three axes are as follows:
    • hierarchical temporal compression: recent actions are kept at full resolution, and older segments are compressed by progressive pooling. This fits 10⁴–10⁵ per-user history within a fixed serving budget.
    • multi-interest representation: models a user's multiple intents across heterogeneous modalities such as news, Q&A, and short video as a separate axis, because sparse ID embeddings alone are not enough.
    • integration with the MGS deployment system
• Why notable: The angle of pre-compressing ultra-long behavior sequences is adjacent to ChronicleRec (Tencent PCG), but the role where compression sits differs. ChronicleRec focuses on the input cache of the ranking backbone (Chronicle Token, target-independent), and MuSeR places compression in the retrieval layer.
  • Details: Kuaishou already has the SWIM (evaluator), UniRec (cross-stage), DrEM (pxtr noise), PushDualGen, HD-Rec, and OneRec-Thinking lines. MuSeR adds one more data point on the lifelong user modeling axis.
• Evidence: verified (abstract-level statement) arXiv 2609.23677, 2026-09-20 v1. The author domain was confirmed as Kuaishou in the HTML version.
• Status: Watching. Kuaishou is not a tracked service, so no services canon is created. A detailed analysis is in daily/2026-09-23.md.
  • Details: The next points to watch are whether the main text has A/B figures for the MGS deployment system and the actual size of the fixed serving budget (the compute budget cap). We also look at where exactly it sits relative to ChronicleRec's target-independent style, in contrast with a target-aware style.
  • Details: If the repository's lifelong user modeling angle spreads across three companies (Tencent PCG, Kuaishou, and the Meta HSTU lineage), it becomes a candidate for a new section in topics/sequence-transformer-ranking.md.

Meta Search Hybrid GPU-CPU Retrieval — co-serves a trillion-doc source corpus through two paths, GPU billion and CPU tens-of-billions, deployed in production

• What: The first statement to reveal Meta's personalized search retrieval layer at the production architecture level. Two hardware paths, GPU and CPU, jointly serve document snapshots that each path selected separately.
  • Documents of the multi-trillion-scale source corpus pass policy, quality, language, freshness, and dedup. Then hardware-specific selectors build the GPU snapshot and the CPU snapshot independently. The GPU snapshot is billion-scale, and the CPU snapshot is tens of billions, roughly 20× larger.
  • The GPU pathway fuses two-tower retrieval and interaction pre-ranking, putting interaction pre-ranking into the serving path. It then passes 224 candidates to the shared ranker.
  • The CPU pathway combines a dense embedding index, eager term-at-a-time scoring, bulk scanning, and finer clustering, and uses only dot product.
  • Each request can turn on only one of the two paths, turn on both, or disable them. The results of the two paths are deduplicated at the aggregator.
  • It states that in a full-system A/B against the legacy CPU-only configuration, model-scored relevance and substantive engagement rise together.
  • It claims a 4× capacity improvement for the CPU pathway over the dense retrieval baseline. This figure is limited to vector ops.
• Why notable: Meta's search infrastructure papers had a long gap after the 2020 EBR and the 2021 Que2Search. This is the first angle in this repository's corpus that covers depth–breadth heterogeneous co-serving with production A/B figures.
  • The existing topics concentrate on the ranking and retrieval model layers and have not covered the serving hardware division-of-labor layer.
  • This paper appears to share a substantial part of its infrastructure (Xue et al., 2026 "existing GPU platform") with the VDD/EvoPilot paper, which uses the same metric, Good Search Result Rate, or GSRR.
  • The axis of this observation is that retrieval stack layer information for Meta's search and discovery surfaces is opening across two papers together.
• Evidence: verified arXiv 2609.21281, 2026-09-18 v1 submitted, KDD 2027 accepted. All 21 authors are at Meta Platforms (Menlo Park), confirmed in the HTML version.
• Status: Watching (deep dive needed). The surface differs from Facebook Feed, which this repository tracks, but the infrastructure may be shared. A detailed analysis is in daily/2026-09-22.md.
  • Confirm which paper the Xue et al., 2026 reference is, and attach it to the layer information in the repository topics.
  • We also watch whether the Google/Bing/Yandex families publish depth–breadth heterogeneous co-serving with production A/B figures. Whether a later statement says the P99 inflation of the GPU pathway was reduced is also the next checkpoint to watch.

Meta VDD Model-Based Retrieval + EvoPilot — seed video + pivot query + user context MBR, online catch-up 37-day campaign, GSRR +0.66% relative

• What: VDD (Video Deep Dive) is Meta's online follow-on discovery surface. When a user opens a seed video, it follows up with videos worth watching next through search and ranking. This paper covers MBR, the retrieval model of VDD, and EvoPilot, an LLM experiment protocol that runs in the campaign layer above it.
  • The paper states VDD's serving scale as: "hundreds of millions of users".
  • A request consists of three elements (s, q, u): seed video s, machine-generated pivot query q, and user/context u.
  • On top of that sits MBR (model-based retrieval), a two-tower dense retrieval. The request tower Q takes the pivot query, user context, and request-side seed features, and the candidate tower D takes only candidate video features.
  • The primary loss is an out-of-session in-batch InfoNCE. Examples from the same session are removed from the negatives.
  • The candidate-conditioned interaction head G(h_r, h_d, a(d)) outputs a weighted BCE over click and engagement auxiliary labels.
  • The index is republished every hour at the scale of hundreds of millions of videos, but the weights are not retrained every hour.
  • The baseline is P=K=3000 with the interaction head disabled.
  • EvoPilot is a human-gated LLM autoresearch protocol. In the campaign layer above, it runs 7 direction campaigns over 37 days.
  • The earlier primitive autoresearch reached the wrong conclusion that the interaction head yields hit-rate −22 pp. EvoPilot traced this conclusion to an output depth defect in the evaluation code (3,000 vs 600), and after fixing the defect it gained +3.20 pp offline.
  • A 7-day online randomized experiment confirmed a lift of GSRR relative +0.66% in the US·Canada VDD-pivot slice. The experiment covers tens of millions of accounts.
  • After the experiment review, the interaction head was adopted as the default serving configuration of VDD MBR, and it is on as of the manuscript cutoff.
• Why notable: It is a new axis in two ways. One is the in-house retrieval stack of a Meta discovery surface, and the other is the verification layer that automated LLM experiments require.
  • It is the first production claim that a surface mixing "search + recommendation" inside Meta has its own retrieval stack. It is a discovery axis different from the friend-graph axis of Facebook Feed and the watch-sequence axis of Instagram Reels.
  • It is the first case study showing that industrial LLM automated experiments on a production pipeline require a human-gated verification layer that even catches evaluator defects.
  • Its role differs from the "LLM inside the model" angle covered by the spread of LLM-native ranking thread. This one is "LLM on top of experiment automation".
• Evidence: verified arXiv 2609.21257, submitted 2026-09-18 v1, KDD 2027 accepted. All 5 authors are affiliated with Meta Platforms (Menlo Park), which we confirmed in the HTML version. 4 of the 5 authors overlap with the author list of the Hybrid GPU-CPU paper above.
• Status: Watching (deep dive needed). The original text does not state whether the VDD surface appears on Facebook Watch, Reels, or Instagram Explore UI. The detailed analysis is in daily/2026-09-22.md.
  • We watch whether the official blog, IR, or Transparency Center cards reveal the surface mapping.
  • The first checkpoint to watch is to check the Xue et al., 2026 "existing GPU platform" reference and link the infrastructure layer to the Hybrid GPU-CPU Retrieval paper.
  • The second axis is whether other teams adopt the EvoPilot protocol and whether alternative protocols (Google DeepMind, Anthropic, etc.) are published at production scale.

LIGE-GR (Meta) — smooth leap from ranking to generative recommendation on Instagram Reels and Facebook Video, time-spent +1.14% / +0.72% relative

• What: A Meta paper claiming a "smooth leap" from ranking to generative recommendation. It reports the results of deploying to two surfaces, Instagram Reels and Facebook Video, at the abstract level.
• Why notable: It is direct evidence for two of the services this repository tracks (Instagram and Facebook), and the deployment claim came with production A/B figures. The point to watch is whether it becomes the third axis of Meta LLM/GR ranking after Meta AMBER and ConnectionMind. AMBER concerns input tokenization, and ConnectionMind concerns a graph path reasoner.
• Evidence: verified arXiv 2609.18148, v1 submitted 2026-09-16.
• Status: Watching (deep dive pending). In the next batch we read the HTML version to understand the mechanism, baselines, and deployment scale, then decide whether to fold it directly into this repository's service canon.
  • A deep dive is needed. We will read the original HTML and review it as a candidate for the canon update (services/instagram.md, services/facebook.md).

SARA (Kuaishou) — articulated rationale scaling in MLLM-based recommendation, deployed in production to 240M users for 30+ days

• What: An approach that scales rationale in an articulated way in a Multi-modal LLM (MLLM)-based recommendation system. It was deployed in Kuaishou production for more than 30 days, and the paper claims an engagement gain in an online A/B test on 240M users.
• Why notable: Kuaishou is not one of this repository's tracked services. However, it is a large-scale production MLLM recommendation deployment, so it is a new company data point for the spread of LLM-native ranking thread we are watching. Kuaishou already has SID lines such as OneRec and PushDualGen, but MLLM rationale is a new angle.
• Evidence: verified arXiv 2609.17639, 2026-09-15 v1.
• Status: Watching. It is not a promotion candidate. We add a one-line summary to the LLM-native ranking thread.

ANGLE (Tencent) — one-step retrieval framework for real-time sponsored search ads, GMV +2.16%

• What: An end-to-end LLM ranking approach that replaces multi-stage cascades. It claims consumption +1.81% and GMV +2.16% in a real-world search deployment.
• Why notable: One-step retrieval is a different angle from the SID and GenRec families, replacing multiple stages with a single model. If Tencent PCG's TGR is a statement of deploying organic content generative ranking to 5 surfaces, ANGLE ties sponsored search ads retrieval into a single LLM. This suggests Tencent is pushing LLM-native angles on both the organic and ads axes.
• Evidence: verified arXiv 2609.18296, v1 2026-09-16. All 8 authors are affiliated with Tencent Inc., and the email domain @tencent.com was confirmed in the HTML version.
• Status: Watching (Tencent confirmed). We watch whether follow-up sources reveal the ranking improvement details and the surface (WeChat search or a WeChat channel).

PCap (Meta Facebook Marketplace) — personalizes retrieval-stage diversity capping with per-user Shannon entropy buckets, launched on Marketplace Feed, VPV +0.30% + PDP +0.22% + Sessions +0.17%

• What: PCap is a framework that enforces per-user diversity constraints as per-category caps at the candidate generation stage of Meta Facebook Marketplace.
  1. It computes the Shannon entropy $S(u) = -\sum p(c|u)\log p(c|u)$ from user u's recent clicked-category distribution and uses it as a proxy for diversity preference.
  2. It sorts users into 6 buckets, from bucket-1 to bucket-6. bucket-1 is the users who click most narrowly, and bucket-6 is the users who click most broadly.
  3. It applies per-category caps at each multi-source indexer shard, capping jointly in the two stages of scanning and aggregation. The cap is retrieval source fetch count × tuned multiplier fk × user bucket diversity multiplier mu.
  • The key that groups categories is the hundreds of categories of the Facebook Product Taxonomy (FPT).
  • Parameter search uses a new method, Parameter Tuning Sequence (PTS). PTS is a sequential range-narrowing A/B procedure that keeps the control arm fixed and exposes candidate arms to a small amount of traffic for several days each, replacing online grid search.
• Why notable: In the Marketplace Feed production A/B, it keeps a significant lift while the carrier metric, Marketplace Sessions, also rises verified. The new angles are that serving latency inflation stays at 0.8ms and that capping attaches to retrieval scanning rather than the ranking stage. Existing MMR and DPP sit in the ranking and re-ranking layers.
  • Phase 1 compared a uniform cap with no cap, and the results are VPV +0.3088% (p<0.01) and Latency +8.6 ms.
  • Phase 2 compared PCap with a uniform cap, and the results are VPV +0.2243%, PDP clicks +0.2250%, Marketplace Sessions +0.1708%, and Latency +0.8 ms.
  • The diversity score changes that split by bucket (bucket-1 −1.05%, bucket-6 +0.25%) are evidence that the cap was personalized as "narrow for users who click narrowly, broad for users who click broadly".
  • Facebook Marketplace is not a service canon of this repository. However, in light of feed-diversity.md and the candidate allocation discussion the repository already tracks, this is the first production claim that personalizes a retrieval-stage cap along a diversity axis.
• Evidence: verified arXiv 2609.16452, submitted 2026-09-15. All 11 authors are affiliated with Meta (Menlo Park·Boston), and the @meta.com domain was confirmed in the HTML version.
• Status: Watching. Marketplace is not a canon service of this repository, so this is not a promotion target, but with further observations it becomes a promotion candidate for topics/feed-diversity.md. The detailed analysis is in daily/2026-09-17.md.
  • We watch whether a statement appears that the same stack expanded to other Meta surfaces (Facebook Feed, Instagram, Reels).
  • The next points to watch are whether a second company adopts diversity personalization of retrieval-stage caps and whether PTS extends to other tuning scenarios (such as ranking hyperparameters).

PinDCO (Pinterest) — production DCO that selects ad creatives in real time with whole-page impact, launched on Pinterest Ads, Ad CTR +3.09% relative

• What: PinDCO is a dynamic creative optimization (DCO) system attached to Pinterest's ads pipeline. It runs in parallel with ads ranking, so latency is bounded within the ranking upper bound.
  • The Creative Component Fusion Network (CCFN) has a tower for each of the image, title, and layout components. After matching convergence speed with per-component hyperparameters (e.g., dropout rate), a component fusion layer produces a delta score to add to the logit of the ads ranking model.
  • The Pixel-aware Adjustment Module (PAM) reflects the cannibalization in which a tall creative in the waterfall grid pushes out the exposure space of neighboring slots. The approach is the aspect ratio penalty $P(ar_{rel}) = \mathrm{clip}(1 - \tanh(k(ar_{rel} - 1)), 0, 1)$.
  • In serving, creative-id-based sharded caching and dynamic batching absorb the QPS. Removing dynamic batching makes P99 +87% and P90 +114%.
• Why notable: In the Pinterest Ads production A/B, it launched on the platform with ad CTR +3.09% relative and Successful Session +0.04% relative, keeping the whole-page metric positive verified. It is the first production claim in this repository's corpus to treat DCO as a separate ranking system, and the first to address whole-page-aware creative ranking on a waterfall grid surface.
  • The baseline Lightweight-Only is +1.49% CTR and −0.24% Session, and Peri-CR is +1.70% CTR and −0.12% Session.
  • Both baselines cut the whole-page metric while raising ad CTR. PinDCO, in contrast, keeps both axes positive at the same time.
  • Its role differs from Pinterest CLR (Home Feed organic candidate generation conditional embedding). CLR is organic candidate generation, and PinDCO is ads creative selection.
• Evidence: verified arXiv 2609.11943, submitted 2026-07-21, RecSys 2026 accepted. All 8 authors are affiliated with Pinterest Inc., and the email domain @pinterest.com was confirmed in the HTML version.
• Status: Watching. Pinterest is not one of this repository's tracked services, so we do not create a services canon. The detailed analysis is in daily/2026-09-15.md.
  • We watch whether the RecSys 2026 talk includes statements on component expansion (description, CTA, landing) and on detailed instrumentation of the whole-page metric.
  • The next point to watch is also whether a second company issues a statement treating waterfall-surface ads creative optimization as whole-page-aware.

ChronicleRec (Tencent PCG) — pre-compresses ultra-long user behavior sequences into target-independent Chronicle Tokens for Weixin Moments Ads pCVR, GMV +1.61% (95% CI [0.678%, 2.547%])

• What: ChronicleRec is a pre-train-and-transfer framework for lifelong user modeling. It compresses an ultra-long behavior sequence into target-independent Chronicle Tokens once per user and supplies them to the ranker through a cache.
  • Recency-aware multi-granularity merge keeps recent events fine-grained and merges older events coarsely.
  • Interleaved causal encoding distributes learnable query tokens across the merged sequence and uses causal attention. Each query summarizes only the history before its own temporal anchor.
  • Multi-horizon parallel branches combine masking of different recent-history windows to learn complementary long-range interests.
  • Mask-and-predict pre-training and Chronicle Alignment align historical signals with near-present intent.
• Why notable: On the Weixin Moments Ads pCVR (conversion rate prediction) surface, a 7-day A/B gave GMV +1.61%, a statistically significant result whose 95% CI [0.678%, 2.547%] excludes 0 verified. It belongs to the target-independent family that leaves the ranker backbone unchanged and feeds in only the lifelong representation through a cache, so it is a new data point that the adoption cost is low.
  • On this surface, the Chronicle Token was incorporated into both paths, NS-Token and S-Token, of the MixFormer ranker backbone.
  • In ablations on the Tencent AdLive dataset, the interleaved-causal configuration is highest at 0.8034 GAUC. Removing Chronicle Alignment causes the largest drop, −0.0086 GAUC.
  • On KuaiRand, GAUC holds at 0.5499 even when the most recent 1,000 events are fully masked. This is well above the recent-window Short-Attn (0.5307).
  • The Meta HSTU/Wukong lineage and the Netflix Foundation Model contrast with ChronicleRec in that they were approaches that scale the sequence encoder.
  • If the TGR (Tencent PCG) line was a 4-in-1 framework unifying ranking, retrieval, slate, and reason, ChronicleRec addresses the lifelong user representation that feeds into ranking itself.
• Evidence: verified arXiv 2609.12375, submitted 2026-09-11. Based on the Weixin Moments Ads pCVR surface deployment claim and the recurring names, we judged the 12 authors to be the Tencent PCG ads line inferred. The email domain is not exposed in the HTML version.
  • The recurring names include Junwei Pan, Chao Zhou, Chengguo Yin, and Jie Jiang.
• Status: Watching. Tencent PCG is not one of this repository's tracked services, so we do not create a services canon. The detailed analysis is in daily/2026-09-15.md.
  • We wait for a follow-up statement that it was ported to Tencent PCG surfaces outside Weixin Moments Ads (news, video, music, web novels).
  • We also wait for statements that other companies (Meta, Alibaba, ByteDance) adopted target-independent compression in production.
  • If such statements appear, it is a candidate for a new section in topics/sequence-transformer-ranking.md.

UniRec (Kuaishou) — aligns cascade pre-ranking + ranking fusion with two-stage joint learning, app usage duration +0.616%

• What: UniRec (Unified Cross-stage Recommendation Fusion) aligns the fusion agents in two stages of an industrial recommender cascade through joint learning. The two stages are pre-ranking and ranking. The two agents partially share input embeddings and are trained together in a single computation graph so that gradients flow between the stages.
  • Dual-axis preference alignment works on two axes. On the vertical axis, it transfers downstream pairwise preference to the upstream fusion score.
  • On the horizontal axis, it recasts dozens of pairwise objectives over heterogeneous prior signals as bidirectional preference evidence.
  • Attribute group-relative regularization suppresses the side effect in which unconstrained end-to-end fusion over-concentrates on high-reward attribute groups.
• Why notable: Offline, it consistently beat single-stage fusion and cross-stage coordination baselines verified. In the online A/B, it gained +0.616% relative in app usage time and was fully deployed on the Kuaishou platform verified. This statement takes a new approach to an old problem (cross-stage inconsistency) by training the fusion agents of the upper and lower cascade stages jointly.
  • In August and September alone, Kuaishou issued its fifth or sixth production claim, following the HD-Rec, PushDualGen, DrEM, SWIM and OneRec-Thinking lines, outside the Shopify HSTU lineage thread as well.
  • The earlier per-line axes are: HD-Rec is SID cross-domain, PushDualGen is SID·CoT push, DrEM is pxtr noise robustness, and SWIM is a session-survival evaluator.
  • The statements so far differ from this one in that they were new axes of algorithm lines (SID, evaluator, noise robustness, etc.).
  • It does not connect directly to the Meta HSTU lineage or the LLM-native recsys thread. However, it adds a data point to the observation "how long have industrial cascades lived without touching the misalignment between stages".
• Evidence: verified arXiv 2609.11052, v1 submitted 2026-09-10. The 8 authors appear to be on the Kuaishou line, but their affiliation is not yet confirmed (the affiliation must be checked in the HTML version) inferred.
• Status: Watching. Kuaishou is not one of this repository's tracked services, so we do not create a services canon. If the author affiliation is confirmed in the HTML version, we will revisit the canon update.
  • If a similar cross-stage joint learning statement from a second company appears, it becomes a promotion candidate for topics/multi-task-ranking.md or a new cascade-fusion topic.

Shopify Foundational Generative Recommender — HSTU-based sequence predictor deployed to the Shop app and merchant storefronts, first statement in the e-commerce domain

• What: Shopify's Foundational Generative Recommender System is an autoregressive sequence predictor newly added to product recommendations in the Shop app (the buyer app) and merchant storefronts. It serves as a candidate generator and as a ranking candidate source.
  • It consumes buyer event sequences (search, lookup, add to cart, purchase) in raw form and predicts the next interaction autoregressively.
  • It takes Meta HSTU as the starting point and adapted details such as time encoding, negative sampling, and recall optimization to e-commerce.
  • RoPE-based rotary encoding and relative attention bias encode absolute time and relative recency together.
  • Sharing negatives across batches and positive-aware hard negative selection raise the quality of contrastive training.
  • Boosting that treats the predictions of other ensemble models as hard negatives puts candidate diversity into the training objective.
• Why notable: In the production A/B, four metrics rose, including Shop orders +0.94% relative verified. This source is the sixth industrial company statement in the HSTU lineage and the first statement in a separate domain, e-commerce product recommendation.
  • The other three metrics are high-quality CTR +5% relative, conversion rate +0.71% relative, and final product recall@4 +4.8% relative.
  • CUDA kernel optimization gave a 7.3x speedup in the training pipeline.
  • It is served at the scale at which the whole platform handled 2.2 trillion edge requests during BFCM 2025.
  • The earlier HSTU lineage statements are Meta HSTU/Wukong, Netflix Foundation Model, Meta ads LLaTTE, Alibaba TransRetrieval, ByteDance ReST, and Tencent PCG CCFormer. These statements were concentrated in ads, streaming video, and social content.
  • Semantic ID has not yet been deployed. Shopify explicitly named it as the next step, on the grounds that "large embedding tables are the bottleneck".
• Evidence: verified Shopify Engineering blog, 2026-02-25, Ali Khanafer·Yang Liu(Senior Staff ML Engineers).
• Status: Watching. Shopify is not one of this repository's tracked services, so we do not create a services canon. The detailed analysis is in daily/2026-09-09.md.
  • Because this is a 2026-02 source, it was found later than its publication date. The retrospective point for the batch is that the source itself had not been caught by the industry blog sweep.
  • The identity and parameter counts of the baseline models (the previous two-tower, DLRM, GBDT, etc.) are not public, so the strength it adds to the scaling law thread is weak.
  • The next points to watch are whether follow-up statements appear at the RecSys 2026 industry track or KDD 2027 and whether the results of adopting semantic ID are published.

PTDG (Huawei AppGallery) — personalized task dependency graph replacing MMoE, CVR +1.2% / eCPM +1.9% in a production A/B with 100M+ DAU

• What: Personalized Task Dependency Graphs addresses the multi-task learning architecture of industrial recommenders. It rewires the dependencies between tasks in the conversion funnel for each item.
  • Existing MMoE-family models assume that dependency strength is the same across all traffic. This work starts from the observation that task correlation differs with item characteristics.
  • Low-rank approximation makes the strength of dependency pathways differ per item, while causal constraints such as Click→Pay are enforced with hard masking.
  • Adaptive Progressive Masking (APM) is added to train the parameters of sparse tasks and dense tasks separately and to suppress optimization instability.
• Why notable: It was deployed as the ranking model of a surface presumed to be Huawei AppGallery. In a 2-week A/B it gained CVR +1.2% and eCPM +1.9% over the deployed MMoE baseline verified. Its distinguishing point is that it redefines the dependency structure of the static conversion funnel itself as a learnable personalized graph.
  • The paper describes this surface as: "a mainstream app distribution platform, serving over 100 million daily active users".
  • The A/B ran as a 10%×10% split, and the remaining 80% kept the original production model.
  • Offline, the AUC of sparse conversion tasks on KuaiRand1K and the in-house industrial dataset rose by up to +1.45% verified.
  • The direction differs from parameter- and sequence-axis scaling such as Meta HSTU/Wukong, Netflix Foundation Model, and Alibaba TransRetrieval.
  • Like Xiaohongshu GateDiffInt, which appeared alongside it, a redesign of the CVR head itself is newly added to what we are watching.
• Evidence: verified arXiv 2609.04862, 2026-09-04 v1, CIKM 2026 accepted. All 11 authors are affiliated with Huawei Technologies Shanghai / Nanjing / Dongguan / Beijing, which we confirmed from the affiliation listing in the HTML version and the @huawei.com domain.
• Status: Watching. Huawei is not a tracked service in this repository, so we do not create a services canon file. A detailed analysis is in daily/2026-09-08.md.
  • If a second company issues a similar personalized task dependency statement (GCN, causal masking, APM), it becomes a candidate for a new section in topics/multi-task-ranking.md.
  • We will check whether the CIKM 2026 presentation slides disclose the scale of the industrial dataset and the absolute-figure gap between KuaiRand1K and the in-house data. Whether it expands to other Huawei surfaces such as HiSuite/Petal Ads is another checkpoint to watch.

Pinterest CLR — expands Home Feed conditional embedding retrieval into a unified architecture, p90 latency down 85%

• What: Conditional Learned Retrieval (CLR) conditions the user tower of Pinterest's Home Feed two-tower candidate generator on explicit retrieval context, producing multiple condition-aware embeddings. The contexts are Interest, Pin, and Board. This source summarizes how CLR, which started from an initial interest condition, evolved along five axes.
  • Condition type expansion: it widened from Interest to Pin to Board, and an LLM generates new interest signals.
  • Model expansion: it moved from the Conditioned User Sequence Transformer to CLR based on the PinFM Foundation Model, and introduced a contrastive alignment loss.
  • Condition representation unification: Unified CLR and Slot Architecture provide three slots: condition_os, condition_id, and condition_type_id.
  • Training efficiency: Request-level Training and M-Falcon raised throughput 2x.
  • Serving redesign: after unifying single requests, it was redesigned with GPU serving and NVEmbed.
• Why notable: It saved a 7-figure cost, cut p90 model latency from 80ms → 12ms (85% reduction), and enabled follow-up initiatives such as CLR FM and Search CLR verified. In particular, it states explicitly that after moving to GPU serving it kept reading experiment feedback with Claude Code to automatically tune the per-request condition budget, which is the first statement of attaching an AI agent to a recsys production tuning loop inferred.
  • The semantic ID has a structure of 5 hierarchical layers × 2048 codes each, and sits next to the image signature embedding in the feature cross of the Pin tower.
  • The blog post says this automatic tuning was done separately for each condition type: Interest, Pin, and Board.
  • A contrast emerges: Netflix Foundation Model unified above the scoring function, while Pinterest CLR unified along the conditioning axis of retrieval.
• Evidence: verified Scaling Conditional Learned Retrieval for Pinterest Home Feed, 2026-08-26, Pinterest Engineering, 6 authors. The prior papers are RecSys'24 Bootstrapping Conditional Retrieval and KDD'25 Synergizing Implicit and Explicit User Interests.
• Status: Watching. Pinterest is not a tracked service in this repository, so we did not create a services canon file. A detailed analysis is in daily/2026-09-07.md.
  • Together with LinkedIn Follows and Spotify GLIDE, it forms three branches of the pattern of "attaching an LLM/foundation-based retrieval channel to the candidate generation layer". LinkedIn Follows is on the profile verbalization side, and Spotify GLIDE is on the Semantic ID generation side.
  • The detailed PinFM architecture and parameter count, and the objective function and feedback cadence of the agentic tuning pipeline, are the next checkpoints to watch.

LinkedIn Follows Recommendations — LLM bi-encoder EBR deployed as the candidate generator on two surfaces, MyNetwork and Home Feed

• What: LinkedIn rewrote candidate generation for Follows Recommendations as embedding-based retrieval with a fine-tuned bi-encoder LLM. The encoder is a mid-sized instruction-tuned open-source model.
  • Profiles are given to the encoder through narrative prompting, which restructures profile fields into natural-language sentences.
  • Creator embeddings are precomputed in nightly batches on a Ray distributed GPU cluster and loaded into a FAISS IndexFlatIP exact K-NN index (~10M creators).
  • Viewer embeddings are inferred in real time on Proxima and queried against hosted vector search.
  • In L2 ranking, a dense NN consumes behavioral, categorical, and collaborative-filtering features plus task-aware projected embeddings features. The projected embeddings reduce 1024–4096 dimensions to 64–128 dimensions.
• Why notable: It was deployed on two surfaces, the MyNetwork tab and Home Feed. In the production A/B it produced a statistically significant lift in follow rate per segment, and the statement is that the lift is largest in the new-member segment. It did not disclose specific uplift percentages or latency figures.
  • Following Netflix GenRec, Yandex, Alibaba, 2 from Meta, and Spotify GLIDE, it joins the spread of LLM-native ranking thread as the sixth company.
  • Its role is the same "LLM embedding candidate generator channel" as Spotify GLIDE, and it is complementary to Pinterest CLR (conditioning).
• Evidence: verified Rebuilding LinkedIn's Follows Recommendations with LLM-Based Semantic Retrieval and Ranking, 2026-08-24, author Haohua Wan.
• Status: Watching. The lack of specific figures is a weakness. A detailed analysis is in daily/2026-09-07.md.
  • We will watch whether follow-up papers such as those at RecSys/CIKM disclose absolute figures per segment.
  • The improvement in the new-member segment is a meaningful data point in the cold start debate.

Spotify GLIDE — deploys a 1B LLM + Semantic ID candidate generator on Home podcasts

• What: GLIDE is an additional candidate generator in which a 1B-parameter LLM generates Semantic IDs to pick candidates. A Semantic ID represents a catalog item as a short token sequence. It was deployed on the Spotify Home podcast discovery surface and supplies candidates to the downstream ranking pipeline.
• Why notable: In a 21-day A/B at a scale of ~20M impressions, it gained non-habitual podcast listening +5.4% and new-show discovery +14.3%, while keeping the latency and engagement requirements verified. Catalog hallucination is avoided thanks to Semantic IDs.
  • The NEO framework mentioned in the same blog post (a model unifying search, recommendation, and reasoning) was tested at a scale of 10M+ items × 15M users, but it is at a research stage with no production deployment.
  • It is added to both the LLM-native ranking thread and the SID/generative retrieval thread.
• Evidence: verified From Models to Products: LLMs for Recommendation at Spotify Scale, 2026-08-03, Spotify Research blog post.
• Status: Watching. Whether a production statement on NEO appears, and whether GLIDE expands beyond Home podcasts to other surfaces (such as music discovery), are the checkpoints to watch. A detailed analysis is in daily/2026-09-07.md.

Spotify SPDD — inference-time multi-objective decoding on top of SASRec

• What: Stochastic Primal-Dual Decoding, SPDD for short, is a lightweight controller that intervenes only at the decoding (inference) time of a pretrained generative recommender. It dynamically adjusts the weights of auxiliary objectives (e.g., relevance and diversity) as decoding progresses, and does not require retraining.
  • Spotify applied it to a production A/B of SASRec-based playlist ranking and to offline evaluation (TransAct, etc.).
• Why notable: In a production A/B at a scale of ~1 million users × 600,000 playlists, auxiliary stream-share was +5.44% over control, about 1.8 pp apart from the +3.60% of fixed-weight. It achieved this result without a decrease in user consumption, and the increase in skip rate was also lower.
  • Whereas Kuaishou PushDualGen avoided serving cost with SID and a skippable copy structure, SPDD is an alternative angle that leaves the serving model as is and changes only the decoding policy.
  • It is added to the SID/generative retrieval thread.
• Evidence: verified Balancing Multiple Objectives in Generative Recommendations with Adaptive Decoding, 2026-08-07, Spotify Research blog post.
• Status: Watching. If a second company issues a similar inference-time multi-objective controller statement, it becomes a candidate for a new section in topics/multi-task-ranking.md. A detailed analysis is in daily/2026-09-07.md.

TGR (Tencent PCG) — binds generative-paradigm ranking, retrieval, slate generation, and reason tokens into one frame, deployed on 5 surfaces

• What: This is a four-component framework from the 13-person TGR Team at Tencent Platform and Content Group (PCG). The paper states deployment across five surfaces (anonymous Scenario 1–5): news feeds, short videos, long videos, music, and web novels.
  • CCFormer: Transformer ranking
  • BARGE: hierarchical semantic ID generative retrieval
  • HiGR: whole-slate generation
  • TGR-Reason: offline generation of semantic-ID reason tokens and their injection into the serving prompt
• Why notable: It is the first case to replace both HSTU and OneRec as production baselines. The SID line (→ generative retrieval / semantic ID thread) gains the second-largest statement by company scale.
  • The paper states its relationship to OneRec explicitly: "aligned with the one-model direction of OneRec at Kuaishou".
  • CCFormer was fully launched in Scenarios 1 and 2, carries all production traffic, and replaced HSTU and DLRM. CTR rose +3.57% and ad revenue 1.71%, GFLOPs are half those of HSTU, and training is 2.21x.
  • BARGE was fully rolled out as a generative retrieval channel, with Hit@5 +10.2–16.9% over OneRec.
  • HiGR recorded slate quality +15.9–21.3% over OneRec and 5x inference speed (A/B in three scenarios).
  • TGR-Reason recorded +477.8% in cold start new-user Hit@1.
• Evidence: verified arXiv 2609.00986, 2026-09-01 v1. We confirmed the affiliation from the HTML.
• Status: Watching. It is added to the SID thread as strong evidence. If a third company statement unifying HSTU/OneRec in one frame appears, it becomes a promotion candidate to topics/llm-in-recsys.md or a new topic (one-model recommendation).
  • The items needing a deep dive are the BARGE OSQ-VAE codebook structure, the substantive difference between HiGR PCRQ-VAE and OneRec SID, the Think model backbone and reason token corpus scale, and the Scenario→Tencent product mapping.

ReST (ByteDance / TikTok Shop Ads) — production deployment of behavior-sequence Transformer scaling in ads ranking

• What: This is a rec-native Transformer ranking framework deployed on TikTok Shop Ads. It combines components along a signal quality axis and a compute asymmetry axis.
  • Signal quality axis: Dual-gated attention, rotary positional and temporal embedding, stabilized residual normalization, training-only auxiliary objectives
  • Compute asymmetry axis: heavy reusable encoder and lightweight cross decoder with projection-free KV attention
• Why notable: In a 1-week A/B on TikTok Shop Ads it gained AUC +1.31% and Advv +11.93% (p<0.01), while holding 50 ms P99 latency. The paper states explicitly that it is fully deployed in production. Following Meta HSTU and Wukong, LLaTTE, and Alibaba TransRetrieval, a ByteDance claim joins the scaling law thread.
  • The authors claim that ReST scales consistently at the sequence length, depth, and width where LLM-style Transformer blocks saturate.
  • The sizes are 5.69M parameters for ReST Base and 22.64M for Large. Both values count only the sequence modeling part.
  • It is added as evidence to the recommendation model scaling law thread.
• Evidence: verified arXiv 2609.01240, 2026-09-01 v1, HTML. All 12 authors are affiliated with ByteDance.
• Status: Watching. This paper does not explicitly stand on a conditional confirmation of the debate between LLaTTE and HSTU over the "necessity of semantic features", but it aligns with the direction that a rec-native design is needed instead of LLM blocks. Because TikTok Shop Ads is an ads domain, whether it expands to organic feeds is the checkpoint to watch.

CAMIE (Snap DPA) — shared encoder + co-engagement signal for I2I retrieval

• What: It addresses two problems of item-to-item search in Snap Dynamic Product Ads (DPA, ads). The abstract states production DPA deployment.
  • Separately trained visual, text, and multimodal encoders fragment the embedding space, which lowers retrieval quality between different item types.
  • If co-engagement signals (the relation of the same user viewing two items together) are handled only with symmetric contrast, they do not reflect the interaction context of the items.
• Why notable: Recent cases of Snap issuing a production statement on I2I retrieval in the ads domain are rare. Whether Snap's next statement continues this line is the next checkpoint to watch.
  • The same group issued SetMIR (multi-interest retrieval as set prediction) at a similar time. However, the SetMIR abstract does not state a Snap production claim, so we put it on hold.
• Evidence: verified arXiv 2608.30255, submitted 2026-08-31. There are 12 authors (Xiaodong Liu, Siman Wang, Congfei Zhang, Hsiang-wei Chao, Xiao Bai, and others), and the HTML version lists the affiliation of all of them as Snap Inc.
• Status: Watching. The I2I retrieval thread is not yet in the canon. If Snap issues a follow-up statement, or a second company (Pinterest, Meta ads, etc.) issues a similar statement using a shared encoder and co-engagement together, it becomes a topic promotion candidate.

VK Friend Recommendation GNN — a case of scaling a message-passing GNN on a production social graph of 194M users and 28B edges

• What: It tackles head-on the two kinds of problems, systems and modeling, that arise when attaching message-passing GNN training and serving to a production-scale social graph of hundreds of millions of users and tens of billions of edges verified.
  • For friend recommendation, multi-hop social context is more decisive than user attributes, so a message-passing GNN is a natural fit.
  • Multi-hash user embeddings combine embedding lookups from several hash functions for each user ID. This cuts the ID embedding table size by more than 98% while maintaining ranking quality.
  • Temporal neighbor sampling reduces neighbor sampling from O(deg(v) + k) to O(log(deg(v)) + k) with timestamp-sorted CSR storage and binary search. The neighbor traversal cost of supernodes (accounts with an extremely large number of follows) is held to logarithmic time verified.
• Why notable: On a social graph of 194M users · 28B edges, it gained friend additions from recommendations +16% and unique friend adders +11.5% over a strong production baseline.
  • A double-digit uplift over a strong production baseline is rare in GNN papers.
  • If Meta ConnectionMind is an LLM that reasons over paths on a graph, this is a case of pushing a pure GNN to the scale limit of a social graph.
  • It reinforces the observation that the domain where GNNs ultimately survive in practice is a fully explicit graph such as a friend graph or a follow graph.
  • It is also a source that can be contrasted with X's SimClusters (SVD-based) and Meta's GNN·LLM dual stack.
  • The abstract does not state explicitly that it was applied without serving latency. However, that is exactly the purpose of the two optimizations, table shrinking and logarithmic neighbor traversal.
• Evidence: verified arXiv 2608.27413, 2026-08-27 v1, CIKM 2026 accepted. There are 3 authors (Maksim Utushkin·Andrei Ovsiannikov·Alexander D'yakonov).
  • We presume the affiliation is VK from the author names and the scale mentioned in the text inferred. The HTML abstract has no affiliation listing, and a code release is mentioned.
• Status: Watching. VK is not a service tracked by this repository's canon, so promotion is on hold. After the conference slides (CIKM 2026, October) appear, affiliation and infrastructure details may be disclosed additionally.
  • If a similar production GNN statement comes from a second company (other than Meta), it becomes a candidate for topics/embedding-retrieval.md or a new social-graph-gnn topic. A GNN update to LinkedIn friend recommendation or Facebook People You May Know is an example.

SWIM (Kuaishou main feed) — redesigns the Evaluator of Generator-Evaluator re-ranking as a session survival process

• What: It redesigns the Evaluator that scores lists in the G-E (Generator-Evaluator) architecture of industrial re-ranking. SWIM models user behavior as a finite-horizon prefix session-level survival process verified.
  • Previous evaluators miss the session perspective because they look only at single-request metrics (e.g., expected watch time).
  • How much the current list contributes to the session goal is factorized with a recursive survival distribution and reached-position conditional rewards verified.
• Why notable: It was deployed in production for main feed re-ranking of the Kuaishou short-video app (>400M DAU). An A/B with 5% of traffic for 7 days gained App stay time +0.351% and 7-day retention +0.048% verified.
  • The confidence intervals are 95% CI [0.27%, 0.42%] for App stay time and CI [0.01%, 0.09%] for retention verified.
  • It replaced the previous evaluator, CAVE, and kept the upstream retrieval and generator.
  • After DrEM, HD-Rec, and PushDualGen, this is Kuaishou's fourth production statement in August alone. It is a new branch in that it redesigns the evaluator itself.
• Evidence: verified arXiv 2608.25104, 2026-08-25, the authors are 6 from Kuaishou Technology, 2 from USTC, and the unaffiliated Kun Gai.
• Status: Watching. If another company's statement redesigning the evaluator itself as a session survival process appears, it becomes a candidate for topics/feed-diversity.md or a re-ranking topic.
  • The fact that Kuaishou issued production statements in four papers on different algorithm lines in the same month is a company-level checkpoint to watch. The four lines are SID (semantic ID, which represents items as sequences of discrete token IDs) cross-domain, SID+CoT push, ensemble ranking noise, and G-E evaluator survival.

MetaStrategy (Alibaba/Taobao) — The LLM generates "executable ranking-strategy JSON" instead of scoring

• What: An LLM policy takes the request context and generates a typed JSON bundle; a deterministic validator and compiler then instantiate it as an isolated Generator verified. This Generator competes atomically with existing strategies under the list-level Evaluator of the GE (Generator-Evaluator) architecture verified.
  • The JSON bundle consists of objective weights, content/category preferences, experience constraints, and position policies.
  • Training builds a 0.8B student from a 4B×2 teacher. It combines production-path replay, 3 kinds of reward, a self-competitive curriculum, and Evaluator-routed on-policy distillation.
  • Production-path replay re-runs logged requests without user impressions. The three rewards are selection, relative-rank, and baseline-lift.
• Why notable: It was deployed to Taobao Homepage Guess You Like with a diff-triggered nearline approach. In a 7-day user-randomized A/B test, it recorded click PV +2.11%, IPV +3.12%, and transaction value +2.83% verified.
  • Because LLM inference sits outside synchronous ranking, no RT impact was observed.
  • In the same A/B test, the treatment group won 27.93% of GE calls verified.
  • Netflix GenRec is an LLM that scores online and Yandex Gryphon-v2 is an LLM that generates online; MetaStrategy is an offline policy-compiler angle in contrast to them. So it counts as the third mechanism of the LLM-native thread.
• Evidence: verified arXiv 2608.09440, 2026-08-10.
• Status: Watching. If a third company states a similar nearline LLM policy compiler, it becomes a candidate for introducing a section in topics/llm-in-recsys.md.
  • It is symmetric with DREAM, which covers the same surface (Taobao home feed). MetaStrategy is a policy compiler, and DREAM is real-time parameter adjustment.

DREAM (Alibaba/Taobao) — Adds an agentic meta-control layer on top of the existing pipeline

• What: An architecture that leaves the retrieval/rank/re-rank pipeline as is and adds a policy layer on top verified.
  • A 3-layer Intent Engine (L0/L1/L2) fuses on-device signals into structured intents. An edge-cloud trigger chain compresses reporting volume to ~8.7%.
  • The MetaModel then reasons in the order M1, M2, M3 to adjust pipeline parameters, and deploys them after they pass safety guardrails.
  • M1 summarizes the intent, M2 plans a strategy by consulting Strategy Memory, and M3 translates it into parameters.
  • A Reward Dual Loop (offline simulation and online feedback) performs dual optimization verified.
• Why notable: The gain comes from parameter adjustment alone, without replacing models. In the Taobao home feed A/B test, applying it only to re-ranking control gave IPV +2.06%, Core IPV +2.39%, and GMV +0.88% verified.
  • When extended to fine ranking, the figures were +2.71%, +3.06%, and +1.31%, respectively, and PV rose consistently by more than +1% verified.
  • The authors stated that it neither replaces pipeline models nor harms serving stability: "neither replacement of pipeline models nor compromise of serving stability".
  • Many authors overlap with MetaStrategy, so the two may be experiment lines of the same organization. Overlapping authors include Chengyu Lai, Jiuning Lin, Bin Zhang, and Han Zhu.
• Evidence: verified arXiv 2608.09408, 2026-08-10.
• Status: Watching. Together with MetaStrategy, it forms a trend of "LLM/agentic control layers on top of industrial recommendation pipelines". Whether the two approaches converge or are deployed in parallel is the next point to watch.

TM20K (ByteDance ads) — Deploys 20K-sequence ranking to production with teacher/student KD

• What: The teacher produces ground truth through heavy training, and a student that compresses the sequence with token merge is raised through KD (knowledge distillation) verified.
  • Both teacher and student keep full attention. It is not target-only attention.
• Why notable: This is ByteDance's production claim on the proposition "longer sequences = better ranking" asserted by Meta HSTU/GEM (ads) and the LLaTTE family. It actually ran up to 20K in ByteDance ads.
  • Deployment results are sequence length 20K, ADSS +1.036%, and serving latency +5.6%.
  • It states that training and serving cost are nearly the same as the existing SOTA: "nearly the same".
  • It is a new data point for the scaling law thread. However, it extracts sequence scale through KD rather than pure parameter scale, so the direction differs.
• Evidence: verified arXiv 2608.07055, 2026-08-07.
• Status: Watching. If ultra-long-sequence claims accumulate from a third company or more, we will add a section to topics/sequence-transformer-ranking.md.
  • The 20K-sequence KD line joins the unresolved debate between Meta LLaTTE and HSTU as a third approach.
  • LLaTTE, one side of the debate, holds that "scaling does not work without semantic features", and HSTU, the other side, holds that "scaling alone is enough".

Progressive Alignment (Webtoon) — Introduces RL post-training to a recsys foundation model

• What: When SFT-ing a recommendation foundation model (FM) per serving surface, the click/like objective diverges from business metrics; this is solved in 3 stages: LP, FFT, and RFT.
  • First, LP (Linear Probing) freezes the pretrained representation and stabilizes only the randomly initialized downstream head.
  • Next, FFT (Full Fine-Tuning) specializes the whole model to the target task.
  • Finally, RFT (Reinforcement Fine-Tuning) aligns to business metrics with a learned reward model.
  • The key trick is to train the policy on dense implicit feedback and use the sparse business target only for the reward model verified.
• Why notable: This is a production claim that transplants the LLM-style RLHF pipeline directly to downstream alignment of a recommendation foundation model. It claims improvement over non-FM baselines in a large-scale online A/B test, but the abstract has no concrete numbers.
  • It can be contrasted with how the Netflix Foundation Model attaches downstream heads. Netflix attaches a head per serving surface by training, and Webtoon adds one more RL alignment stage on top.
• Evidence: verified arXiv 2608.06792, 2026-08-06 (RecSys '26 Industry Track).
• Status: Watching. It is a source for cross-checking the downstream adaptation sections of Netflix FM and Meta GEM. If more production cases appear, it becomes a candidate for topics/multi-task-ranking.md or a new topic (recsys post-training).

Gryphon-v2 (Yandex Music) — A single generate-and-rank model replaces the cascade

• What: It handles retrieval, rerank, and rank with a single model verified.
  • It encodes the user history once and generates Semantic-ID candidates with an autoregressive decoder.
  • It maps the generated candidates to catalog items and passes them to an item-level ranker that reuses the shared encoder state.
• Why notable: In a production A/B test it replaced the existing cascade, active users rose +1.41%, and serving latency is comparable to the cascade.
  • The authors described the replaced cascade this way: "replaces the production cascade of 15+ candidate generators, pre-ranking and final ranking".
  • It is a production claim that directly answers the Netflix GenRec lineage. GenRec is an LLM in the ranking path, whereas Gryphon-v2 unifies retrieval and ranking.
• Evidence: verified arXiv 2608.06213, 2026-08-06.
• Status: Watching. Although it is the music domain, the mechanism is domain-independent, so it is a promotion candidate to canon topics (llm-in-recsys, embedding-retrieval). If a claim that a single generate-and-rank model replaced a cascade appears on another platform again, we will decide on topic promotion.
• 2026-08-13 additional evidence: Yandex Music Sona (arXiv 2608.11015, 2026-08-11), a follow-up from the same company, unified candidate generation and ranking into a single architecture on the "My Vibe on smart speakers" surface. It stated that it replaced the entire cascade in a production A/B test, and the abstract reports active users +4.53% and total listening time +6.30% verified.
  • The abstract's wording on cascade replacement is: "replaced the entire production cascade".
  • It appears to deploy a different model on a different surface, smart speakers, from Gryphon-v2 (entire cascade, +1.41% AU). However, we could not confirm the relationship between the two.
  • The fact that Yandex Music published two production papers in a short period is itself a signal that company-level resources are committed to the ranker family that unifies SID and LLM.

DrEM (Kuaishou short video) — Ensemble ranking with dual correction of pxtr noise

• What: In the ensemble ranking stage, it corrects noise in the per-action predicted probabilities (pxtr) emitted by upstream multi-task models on both the supervision side and the feature side.
  • Examples of pxtr are pctr/pvtr/pcmtr/pftr.
  • Prediction noise causes two problems. It can flip proxy preference labels, and it can flow in as feature-side input and shake the ranking score.
  • On the supervision side, it uses a robust loss that corrects the empirical risk with the pairwise flip probability.
  • On the feature side, it uses a preference-preserving ranking consistency regularizer over perturbations sampled from the prediction noise distribution.
  • Theoretically, it proved that the robust loss is superior to the basic pairwise loss even when the flip probability estimate has error verified.
• Why notable: It ran a 7-day user-randomized A/B test on Kuaishou's industrial-scale short-video platform on top of the EMER and EASQ production backbones. It allocated 5.1% of main traffic, and all metrics had p<0.005.
  • On the EMER backbone, treatment-group metrics are Follow +1.197%, Comment +1.388%, Video View +0.691%, and Long View +0.401%. The remaining metrics are Like +0.625%, Forward +0.683%, App Stay Time +0.116%, and LT7 +0.017%.
  • The authors themselves observed that sparse behaviors such as follow and comment gain more than dense behaviors such as view and long view.
  • This observation matches the theoretical prediction. The prediction is that sparse tasks have large prediction variance and thus a high flip probability, so there is more room for robust correction.
• Evidence: verified arXiv 2608.12778, 2026-08-13, 8 authors (a collaboration between Kuaishou Technology Beijing and Shenzhen University).
  • We confirmed the affiliations in the HTML. Kaiqiao Zhan / Xiaoxiao Xu / Tiantian He and others are the Kuaishou line.
• Status: Watching. We will decide on topic promotion when a similar claim from a second company appears.
  • With a production claim in the pxtr ensemble line following the HD-Rec/PushDualGen SID line, Kuaishou recorded its third in August.
  • The three production claims are SID cross-domain, SID+CoT push, and ensemble ranking noise robustness; all belong to different algorithm families.
  • This is the first case where Kuaishou confirmed the theory of label noise and feature noise in ensemble ranking in large-scale production. So it is a valid reference for companies that suffer pxtr noise from combining multi-objective ranking. X, Meta, and Netflix all count.

SDF (Google Discover) — 2-year production staleness filter, dual learned models for supersession and decay

• What: It decomposes "staleness" in a content feed into two different mechanisms and puts a learned filter on each verified.
  • Supersession is when a new item invalidates an earlier item. An example is an earlier negotiation article after a transfer is confirmed.
  • The supersession filter uses a pairwise classification framework. It builds synthetic training data with an LLM pipeline and then serves a student model through knowledge distillation.
  • Relevance Decay is when an item decays naturally over time even without a follow-up item. An example is a "meteor shower tonight" notice.
  • The decay filter takes the multimodal content of items in text, image, and video as input. It is a multi-horizon PTR (Predicted Traffic Ratio) regression model trained on lifetime visit traffic.
  • The two filters are combined by disjunction (OR) and placed in the upstream candidate pruning stage before ranking verified.
• Why notable: It was deployed in production on Google Discover for 2 years, and user-filed staleness reports fell 54.9% against the pre-deployment baseline. DAU is in the hundreds of millions and MAU in the billions.
  • The supersession filter alone gives 64.0%, and the decay filter alone 34.4%.
  • It was tuned under an engagement-neutral constraint, and artificially aggressive pruning was excluded because metrics improve but engagement drops verified section 4.2.
  • It is a fully learning-based approach in contrast to X's IS_STALE_POST14D bool feature. That X feature is a 14-day time rule plus a count zero stamp.
  • Because it removes staleness before ranking altogether, it differs fundamentally from X's position of "giving the ranker an explicit label".
• Evidence: verified arXiv 2608.15780, submitted 2026-08-16, CIKM Applied Research Track 2026 accepted (DOI 10.1145/3799682.3840082).
  • All 6 authors are affiliated with Google LLC. We confirmed the affiliation notation and the @google.com domain in the HTML version.
• Status: Watching. Google Discover is not yet a tracked service in the canon because it has not released the rest of the stack such as retrieval, ranking, and diversity. If a second component is released, we will reconsider canon registration.
  • If a similar claim from a second company appears, it becomes a candidate for introducing a staleness section in topics/.
  • We could not confirm how Netflix Foundation Model and Meta ConnectionMind handle stale items.

Tlow (WeChat/Tencent) — A flow-based item tokenizer as an RQ-VAE alternative, CTR +10.32% in production

• What: In the item tokenizer of the Semantic ID (SID) line, it approaches the three known problems of the standard method RQ-VAE with a normalizing flow. It claims to beat RQ-VAE on two axes of tokenization, codebook utilization and semantic locality verified.
  • The item tokenizer is a core component of the SID line that converts each item's semantic embedding into a discrete token ID sequence. RQ-VAE is Residual-Quantized VAE.
  • The three problems are codebook collapse, residual accumulation of quantization error, and discontinuity where similar items map to completely different token sequences.
  • Codebook collapse is the phenomenon where most code slots die and only a few are used.
  • Flow transforms the embedding space into a learnable distribution with an invertible neural network and inverts it back verified.
• Why notable: The abstract explicitly states that a production A/B test was run on WeChat (Tencent), China's largest social media platform. The token-ID-based retrieval model recorded user CTR +10.32% globally and +11.64% for new items.
  • The larger gain on new items suggests it is especially strong at cold start. In effect, it confirms the core benefit of the semantic tokenization line in production metrics.
  • So far the SID line has mostly been a combination of "RQ-VAE + downstream model". This is the first claim that replacing the tokenizer layer with a flow-based one also improves production metrics.
  • It is also circumstantial evidence that RQ-VAE's codebook collapse was a problem in real serving inferred. We interpret this as Tencent having hit this problem and offering an alternative.
• Evidence: verified arXiv 2608.24176, 2026-08-25 v1. The author affiliation is the Tencent WeChat rec group.
  • We confirmed the affiliation from the author list, including Chonggang Song and Nian Li, and the WeChat production claim.
• Status: Watching. A new axis has appeared in the SID/generative retrieval thread. The tokenizer architecture itself has begun to diverge in production.
  • There are two points to watch. One is the follow-up release of code and the detailed flow structure, and the other is whether a similar tokenizer-alternative claim appears from a second company (Kuaishou/ByteDance/Xiaohongshu).
  • Tencent WeChat rec has been quiet recently. It is worth watching whether claims from other surfaces such as Weixin Video Channels follow this paper.

TAGR — Anonymous large e-commerce live-stream ads platform, revenue +16.1%

• What: It proposes Temporally Adaptive Generative Recommendation tailored to fast-changing live-streaming ads. The abstract does not disclose the concrete temporal adaptation mechanism (whether a tokenizer retraining cycle or an online update) verified.
  • Live-streaming ads demand strong freshness from the recommender because content, products, and user feedback change by the minute.
  • Yet until now the generative recommender (SID-based) line has mostly assumed a static catalog.
  • The abstract describes the model as built for rapidly changing live content, promoted products, and user feedback: "designed for rapidly changing live content, promoted products, and user feedback".
• Why notable: The abstract states a production deployment to a large-scale e-commerce live-streaming ads platform. Against the production baseline it gained live-room entry rate +8.5%, shopping-cart click rate +7.4%, and revenue +16.1%.
  • The abstract wrote the deployment target as "a large-scale e-commerce live-stream advertising platform".
  • After Twitch (Amazon) Multi-Objective live and Netflix's live cold start, it is the third company to join the live-streaming ranking thread.
  • Because it is an ads surface and a Chinese e-commerce live angle, generative recommenders are accumulating production claims in ads, live, and fast-changing domains too.
  • However, the abstract does not disclose the author affiliation or the platform name. The leading candidates for Chinese e-commerce live-stream are Kuaishou, Alibaba Taobao Live, Douyin, and Xiaohongshu, but the author names do not directly match any of these companies inferred.
• Evidence: verified arXiv 2608.24034, 2026-08-25 v1.
• Status: Watching. Not being able to confirm the platform is the biggest weakness, and checking the author email domains is the next step. If author affiliations are released after v2, we will reconsider canon promotion.
  • If a similar claim covering live and generative together appears from a second company, it becomes a promotion candidate to topics/multi-task-ranking.md or a new live-streaming topic.

HEGM (VK Russia) — Short-video watch-time distribution prediction, detailed production A/B figures released

• What: It proposes HEGM (Hierarchical Exponential-Gaussian Mixture), which predicts the conditional distribution of watch time, the core regression target of short-video ranking. It repeatedly reproduced the three failure modes of the prior SOTA EGMN and put countermeasures in place verified.
  • The conditional distribution of watch time is near-zero-inflated, long-tailed, and multimodal.
  • EGMN (Exponential-Gaussian Mixture Network) models the whole density with a Gaussian mixture. So it supports prediction of CDF threshold events rather than a single point estimate.
  • The first failure mode, variance collapse, is when σ_k(x)→0 creates a likelihood spike on particular samples.
  • The second failure mode, component redundancy, is when several Gaussians converge to μ_i≈μ_j·σ_i≈σ_j.
  • The third failure mode, inactive components, is when most mixture weights collapse to 0.
  • In HEGM, a sigmoid gate predicts the skip probability p_skip(x) = σ(W_skip h + b_skip). If skipped, the density is combined with the exponential component; otherwise, with the Gaussian mixture.
  • This hierarchical skip-watch decomposition separates "quick skip = indifference" from "engaged viewing = actual watching" at the verdict layer.
  • In addition, it adds KL-based variance regularization, applies structured initialization, and removes the forced Gaussian shift and the entropy regularizer verified.
• Why notable: It ran a production A/B test on the large-scale short-video surface of VK (a large Russian social platform). Session depth was +9.26%, shares +14.59%, and dislikes −8.34%.
  • Results by watch length are deep watch (≥10s) +5.75%, non-short watch (≥3s) +4.59%, and skips (<3s) −6.23%.
  • In a 1.5-month test (1 month forward and 2 weeks reverse swap), 5% of traffic was allocated to each of the treatment and control groups, and each group processed ~11M requests per day verified HTML version.
  • After Kuaishou DrEM, this is a claim from another company (VK) in a short-video production ranking paper. In particular, it confirms that watch-time conditional distribution prediction (the EGMN line) is evolving as a separate lineage.
  • All authors hold joint positions at AI VK and Lomonosov Moscow State University. Russian research sources, a different company from the Yandex Music Gryphon-v2/Sona line, are accumulating production claims in succession.
• Evidence: verified arXiv 2608.23356, submitted 2026-08-24. Code and models are released at github.com/rw404/HEGM.
• Status: Watching. If an EGMN-family production paper appears from a second company besides VK, it becomes a candidate for introducing a section in topics/multi-task-ranking.md.
  • It is a source to compare with the Netflix Foundation Model's watch-time head and the 5-quantile head of X's ranking model Phoenix (Gen4+4).
  • Because HEGM treats the regression itself as distribution prediction, it is a different angle from Netflix and X, which combine a point estimate and quantiles.

User Behavioral Densing Law (Ant/Alipay) — scaling law confirmed by a third company

• What: Derives a densing law between the minimum sufficient capacity of tokenization and data scale, through theory and experiments, and on that basis proposes ALGN, a variable-length approach verified.
  • It addresses the problem that when user behavior sequence data grows to billion scale, scaling raw text alone runs into diminishing returns and tokenization becomes necessary.
  • The relation between the minimum sufficient capacity C* and data scale s is ln C*(s) = β + α ln(s/s_0) in a single dimension and ln C*(s) = β + Σ_i α_i ln(s_i / s_{i,0}) in multiple dimensions.
  • ALGN adaptively adjusts tokenization capacity according to data scale verified.
• Why notable: Following the scaling confirmations from the Meta HSTU·Wukong·GEM family and the Netflix Foundation Model, it formalized a scaling law using Ant Group (Alipay) data. However, it has only offline benchmarks, no production A/B claim, and it did not disclose the deployment status.
  • The new angle this time is a law for tokenization capacity scale rather than parameter scale.
  • It carries the naming and formula structure of the Densing Law in the LLM domain (arXiv 2412.04315) over unchanged to tokenization in the recommendation domain.
  • The offline benchmark consists of a 0.5M user held-out set and 3 downstream tasks (classification, text retrieval, user-to-user retrieval).
• Evidence: verified arXiv 2608.23392, submitted 2026-08-24. A collaboration between the DeepFind Team @ Ant Group and Zhejiang University.
• Status: Watching. It adds a third angle, tokenization capacity scaling, to the scaling law thread. If another Ant Group production claim appears (for example, a follow-up saying this densing law was folded into the Alipay serving stack), it becomes a promotion candidate for topics/sequence-transformer-ranking.md.

Descriptive reasoning traces reduce recommendation effectiveness (Spotify Research) — first rebuttal claim for the SID+CoT line

• What: Uses a controlled experiment to check whether the recent trend of attaching chain-of-thought reasoning traces to Semantic ID (SID)-based generative recommenders (SIDReasoner and others) actually improves traditional offline metrics. Introducing explicit reasoning traces decreases traditional offline effectiveness in both SFT and RL verified.
  • The design is a 2×2 factorial: {SID vs Title item representation} × {No Reasoning vs Reasoning}, compared on a common Qwen3-1.7B backbone across 3 Amazon domains.
  • For example, for Office Products with Titles, R@10 dropped from .1587 to .1416 (−11%). Across all domains, Titles fell by −19% to 0% and SIDs by −9% to 0%.
  • Title produces more grounded and interpretable traces than SID. Even so, extensive SID alignment only raises trace quality and does not restore recommendation effectiveness.
  • A richer reward signal partially narrows this gap verified.
• Why notable: Kuaishou PushDualGen claimed improved production metrics by combining SID with skippable copy (CoT-lite). In contrast, the Spotify Research controlled experiment reports the opposite observation: CoT lowers recommendation effectiveness.
  • The metrics PushDualGen claimed are effective play rate +8.50% and dissatisfaction rate −37.70%.
  • However, the two sources differ in conditions. Kuaishou ran a production A/B on the push surface with a real catalog, while Spotify used Amazon reviews and a small (1.7B) backbone offline. The goal is to work out under which conditions the negative result holds.
  • Spotify produced a reproducible rebuttal point: "attaching traces lowers the metrics, and adding more alignment does not help." So when reading the next production claim on the SID+CoT line, the conditions of backbone size, the offline-versus-online gap, and reward design are what matter.
• Evidence: verified arXiv 2608.23154, submitted 2026-08-24. All 6 authors are affiliated with Spotify Research.
  • We confirmed the affiliation in the HTML version from the email domain @spotify.com.
• Status: Watching. llm-in-recsys.md links it as a counterpoint to the branch that combines SID and reasoning. If a similar rebuttal or conditional confirmation claim appears from a second company, it becomes a promotion candidate for a separate section in the topic (SID and reasoning traces: production vs benchmark debate).

GateDiffInt (Xiaohongshu) — a CVR ranker applying diffusion denoising to user behavior sequences plus LLM multi-intent distillation

• What: Addresses the noise-intent coupling problem in user behavior sequences with two components, diffusion-based denoising and LLM multi-intent distillation, and predicts CVR.
  • Noise-intent coupling is the problem that noise dilutes the true intent, and without a structured intent prior there is no denoising target either.
  • Gate-Mediated Controllable Diffusion enhances and denoises the behavior sequence through a controllable forward diffusion process with dual gating.
  • Multi-Intent LLM Distillation has an LLM teacher distill four kinds of structured intents into a lightweight student. The four kinds are long-term / short-term / latent / conversion.
  • The distilled intents are fused by attention to predict CVR verified.
• Why notable: It is deployed on Xiaohongshu's production primary traffic, serves hundreds of millions of DAU, and claims "substantial GMV improvements". The abstract does not disclose concrete uplift numbers.
  • Where OneModel is the angle of unifying rankers across organic, ads, and shops, this one is independent, at the angle of the same company's behavior sequence encoder and CVR head.
  • It uses the LLM only as a teacher rather than placing it in serving, so it is a cousin of MetaStrategy, an offline policy compiler.
• Evidence: verified arXiv 2608.18764, 2026-08-19 v1 / 2026-08-20 v2. The author email domain is xiaohongshu.com (confirmed from the PDF metadata).
• Status: Watching. If a second company publishes a production ranker claim that uses diffusion for sequence denoising, it becomes a promotion candidate for a new section in topics/sequence-transformer-ranking.md. We plan to check whether concrete uplift numbers are released after v3.
  • Xiaohongshu issued two production claims (OneModel and GateDiffInt) in a short period. The two address different layers: ranker unification and the sequence encoder.

OneModel (Xiaohongshu) — unifying 3 surfaces (organic, ads, shops) into one ranker

• What: Serves organic recommendation in Explore Feed, ads in Feed Advertising, and shop recommendation in Merchant Recommendation with a single ranker verified.
  • It maps heterogeneous behaviors into a shared event sequence and builds a long-context user representation with an action-oriented backbone.
  • Scenario-aware Information Modulation balances cross-stream transfer and stream-specific specialization.
  • The production optimizations include stratified user representation, multi-objective training, and feature decomposition. User feature prefetching, shared user-tower computation, and graph-level inference optimization are also part of them verified.
• Why notable: It is the third production claim of one ranker unifying and serving multiple business streams. The Xiaohongshu production A/B reported metrics for all three surfaces together.
  • Explore Feed recorded Time Spent +0.33% and Engagement +1.25%.
  • Feed Advertising recorded value +3.43% and CTR +8.18%.
  • Merchant Recommendation recorded DGMV +1.1867% and GPM +2.1585%.
  • Viewing the three claims side by side, Netflix FM links service diversity from the foundation to the head, and Meta GEM handles surface diversity within the ads domain. Xiaohongshu OneModel unifies ads, organic, and shops as equals.
  • The new angle is that Meta GEM is centered on the ads domain, whereas OneModel unifies organic, ads, and shops as equals.
• Evidence: verified arXiv 2608.18606, 2026-08-19 v2.
• Status: Watching. Xiaohongshu is not yet a tracked service in the canon, and we have not created a service canon (services/xiaohongshu.md). If a similar claim appears from a second company (a platform outside China), it becomes a promotion candidate for topics/multi-task-ranking.md or a new unified-multi-scenario topic.
  • The points to watch are whether details of the "action-oriented backbone" architecture and the cross-stream transfer gating mechanism appear in follow-up papers or talks.

Multi-Objective Ranking for Live-Streaming (Twitch) — delayed feedback window

• What: Trains fresh signals and delayed signals separately with delayed window collection and a multi-model architecture, then combines them verified.
  • Live streaming has the problem that immediate and delayed signals have different time scales. Immediate signals are click and initial watch, and delayed signals are session length and retention.
• Why notable: Twitch (Amazon) claims that an online A/B test on its mobile live feed significantly improved DAV and ARPU.
  • It is adjacent to Netflix's live cold start (RecSys 2025), but this one is multi-objective ranking for a general live feed and adds segment-aware targeting.
• Evidence: verified arXiv 2608.04455, 2026-08-05.
• Status: Watching. It is useful as a source for cross-checking Netflix's live cold start. Folding it into the canon is on hold until more services track live surfaces.

Graduated / closed

(None yet)