2026-08-13

Other changes reviewed today:

• The academic paper on Netflix GenRec published quantitative figures the blog post had not disclosed (+0.006% core metric, the loss formula, and more).
• The landscape of slot 3, where an LLM serves as the ranker, was overturned in two directions within a single day.
• Yandex Music published a paper on Sona, a unified ranker for the smart speaker surface (radar).

1Meta put an 8B LLM in the online path of the organic feed (⚠ Claim changed) — Facebook

That document divides the places where an LLM can be used into six: content annotation, safety verdicts, the ranker itself, user-interest summarization, LLM-as-judge, and natural-language instruction parsing. Slot 3 is the ranker itself. The document asserted that slot 3 was empty for two reasons.

• Meta's blog posts on its own systems GEM, LLaTTE and MARM use the phrases "LLM-scale" and "LLM-inspired". These phrases are only analogies to scaling curves and parameter counts, and the actual rankers were not LLMs.
• The only place the name Llama appeared in Meta's blog posts was the offline automated judge for Groups Search. Even there, applying it to ranking remained in the future tense, "we plan to".

Unlike GenRec, where the scoring function itself is an LLM, the LLM in ConnectionMind is a path reasoner that outputs paths over a graph as JSON. It also takes a hybrid structure in the serving layer.

• For requests from heavy users in the top 5–10%, Llama 3.1-8B runs inference directly on the online hot path.
• The remaining 90–95% is served by a Student GNN (a lightweight graph neural network built by distillation). This GNN is trained on meta-paths (path patterns over node and edge types) that the LLM discovered offline.

Thanks to this arrangement, the statement "an 8B LLM actually enters the online hot path" and the statement "most traffic is still served by the GNN" both hold at the same time.

• The model backbone is Llama 3.1-8B-Instruct, and deployment is confirmed verified — §6.1

  • The experiments covered both 3.2-3B and 3.1-8B, but the model put into production is the 8B.
  • The only Meta system the paper cites is Fan et al. 2019 (GraphRec). The other Meta systems, HSTU, GEM, LLaTTE and MARM, do not appear in the references.

• In the hybrid serving setup, the LLM handles the 5–10% cohort and the Student GNN handles the remaining 90–95% verified — §6.1

  "The top 5-10% of most active users exhibit complex social graphs and high engagement velocity. For this cohort, requests are routed to the full ConnectionMind LLM. For the remaining majority of traffic, we employ an offline distillation pipeline. The LLM processes historical contexts to discover effective meta-paths. These paths are extracted to train a lightweight Graph Neural Network (Student GNN)." — ConnectionMind, §6.1 Hybrid Inference

• The LLM outputs paths over the graph as JSON verified — §4.1
  • The state s_d is fed to the LLM as a serialized prompt of "current paths, candidate next-hop edges, and compact textual node profiles".
  • The LLM produces "a structured action with two components: path expansions and surfaced item–path pairs".
  • It uses neither beam search nor constrained decoding, and verifies the output afterward with a format reward (§4.3).
  • The stopping condition is "when 𝒫_{d+1} = ∅ or when the maximum depth D_max is reached".
• The graph includes Groups and Pages as first-class edge types verified — §3.1
  • The relations that make up the schema are as follows.
    • User–User(Friend, Similar)
    • User–Page(Follow)
    • User–Group
    • Group–Item(Share)
    • Page–Item(Post)
    • Item–Item(Co-watch, Semantic Similarity)
  • Instagram and Threads have no schema equivalent to Groups and Pages.
• Training is a two-stage setup that applies GRPO after SFT verified — §4.2-§4.3
  • The SFT stage trains on BFS shortest paths with teacher forcing. In the paper's words, this corresponds to "For each positive engagement (u,i⁺), we enumerate valid shortest paths from u to i⁺ in the sampled subgraph".
  • The RL stage uses GRPO (of the DeepSeek-R1 family). The reward is a sum of format (−1/+1), F1-based recommendation, and step-wise shaping.
  • The reward weights are α_rec=0.5, α_step=0.3 and α_fmt=0.2.

• The A/B results were measured over several weeks on tens of millions of users verified — §6.2 Table 2

  • The relative lifts over production are watch time +0.43% ± 0.14%, exposure +0.33% ± 0.08%, and video sessions +0.22% ± 0.13%.

  "extensive A/B test on a massive user population (tens of millions of users) over a continuous multi-week period" — ConnectionMind, §6.2

• The offline benchmarks use the public Delicious and Foursquare datasets, not Meta data verified — §5.1

  • The baselines are limited to TrustMF/GraphRec/MHCN/SEPT/DiffNet/DiffNet++/RecDIFF/SGIL/BIGRec. Strong sequential/generative-rec baselines such as LightGCN, PinSage, SASRec and HSTU are missing.
  • The "+88%" lift in Recall@10 reported in Table 2 is an offline result on an academic benchmark, not an online A/B result.
  • The table caption mixes online and offline results, so care is needed when quoting it.

1. The surface is not disclosed. The paper says only "large-scale personalized video recommendation in a social platform setting" (§3) and "short-form video platform" (§6).
   • The surface is most likely Facebook Reels or the unified Facebook Video surface inferred. The three points below all point in the same direction.
     • Groups and Pages enter the graph schema as first-class relations.
     • The main metrics are watch time and video sessions.
     • Since 2024 Q2, Meta has merged Reels, long-form video and Live into a single video recommendation service.
   • However, the paper does not confirm this, so the service name must not be quoted, and the source must be cited as "Meta short-form video platform".
2. There are zero citations of Meta's own systems. HSTU, GEM, LLaTTE and MARM are absent from the references, and the paper does not establish how this work relates to the company's other ranking lines.
   • It may be an experiment run in parallel by a separate team. The five authors are listed only as affiliated with Meta Menlo Park, so even the footnotes do not identify which team (Core Ranking / GenAI / Reality Labs) it is.
3. Serving cost is not disclosed at all. Nowhere does the paper say what SLO the 8B LLM actually meets on the heavy-user hot path. QPS, p99 latency, GPU generation and the fallback policy for cohort routing are all missing.
   • The phrase "millisecond-level latency" does appear, but that statement applies to the Student GNN.
   • This contrasts sharply with the Netflix GenRec serving-layer blog post, which described even vLLM, MSS, FSx pre-materialization and constrained-decoding batch scaling.
4. The offline experiments use only academic benchmarks, and there is no ablation on Meta data. The 60/20/20 split of Delicious and Foursquare lacks strong sequential/generative-rec baselines, which lowers the reliability of the academic comparison. The only statements grounded in Meta data are the three numbers from the production A/B test.

• Added a new subsection to the ranking section of services/facebook.md: "ConnectionMind — a ranking layer where an LLM reasons over graph paths (2026-08 paper, production A/B)".
  • The subsection records the caveat about the surface, the hybrid serving structure, the format the LLM outputs, and the three A/B figures. Items that are not public were moved to the open questions.
  • Added the paper to the sources section. Added three new open questions, covering the surface, serving cost and the relationship to HSTU.
• Changed the following in topics/llm-in-recsys.md.
  • Updated the Meta cell in the slot 3 table.
  • Rewrote the overview sentence "Netflix broke the slot 3 gap" as "Netflix broke it first and Meta is second".
  • Added a new Meta ConnectionMind case section and placed it before the Netflix case.
  • Rewrote the concluding sentence of the framing verification section from "there is no first-party statement from either X or Meta" to "Netflix GenRec was first and Meta ConnectionMind is second".
  • Also updated the Facebook Feed row in the "Who uses this" section.

• Until now, this repository had summarized the position as "Meta uses LLMs as (ads) feature encoders and offline judges. There is no first-party statement from either X or Meta that an LLM enters the organic feed ranking path". ConnectionMind overturned the second of those sentences.
• However, the way it was overturned is subtle. The ranker itself is still not an LLM (it is not a form where the backbone is swapped for an LLM, as in GenRec), and what is entrusted to the LLM is multi-hop reasoning over the graph. Moreover, most traffic (90–95%) is still served by the GNN.
• It also bears on how to read the absence of sources. This repository had long attached the caveat "Meta publishes almost nothing about its organic feed stack — absence does not mean non-use", and ConnectionMind confirmed that the caveat was right. It means Meta did have an organic feed system in operation; it simply was not public.
• The fact that ConnectionMind cites none of Meta's own systems (HSTU, GEM, LLaTTE, MARM) may be a signal that ranking lines inside the same company are still proceeding in several parallel directions. We cannot yet confirm whether this is fragmentation arising from the organizational structure or parallel experiments optimized for different surfaces and services inferred.

↑ Contents

2The GenRec paper published quantitative figures missing from the blog post — Netflix

This paper does. No passage explicitly contradicts the blog post, and instead of overturning the blog's qualitative description, the paper adds quantitative detail, the loss function formula and the team structure.

• The measured value of the online core metric is published for the first time verified — §5.1

  • The blog post wrote only "statistically significant win", whereas the paper gives an absolute value.

  "For example, we observe a +0.006% relative improvement on our core online metric, which is statistically meaningful at Netflix scale." — GenRec arXiv, §5.1

  It matters for quoting that the paper itself characterizes this value as meaningful. The absolute value alone looks small, but the paper attaches the qualifier "at Netflix scale".

• The paper separates the contributions of Phase 1, which builds the foundational LLM, and Phase 2, the post-training on top of it, and the time axis determines their shares verified — §5.3

  "Using the Phase-1 foundational LLM as the base model improves offline ranking metrics by the order of 10-20% compared with using an off-the-shelf LLM as the backbone." "Phase-2 post-training provides a further roughly 35-50% gain in offline ranking metrics when evaluated on the Phase-1 training cutoff date (i.e. when Phase-1 model is the freshest)." "the improvement rises to roughly 80% after two weeks" — GenRec arXiv, §5.3

  The paper states the reason the relative gain from Phase 2 grows over time as "increasing staleness of the less frequently updated Phase-1 backbone". That is, the staler the Phase 1 backbone becomes, the higher the relative value of Phase 2 post-training. The time axis is a new relationship added to the blog post's statement that both stages are needed.

• A concrete example of context compression appeared, and serving cost also falls to 1/3 verified — §5.4

  • The blog post disclosed only "roughly one-third of the original budget".

  "the context length can be reduced to roughly a third of the original token budget (for example, from about 5,000 tokens down to about 1,700 tokens) with only negligible degradation in offline ranking metrics." "Since GenRec is largely compute-bound with serving cost approximately proportional to the context length, we observed the serving cost is reduced to roughly a third of original cost as well." — GenRec arXiv, §5.4

  An example of reducing 5,000 tokens to 1,700 tokens was added. There is also a new quantitative claim that context compression lowers serving cost to 1/3 while causing almost no quality loss. The blog post's qualitative description alone could not have established this combination.

• The loss function formula is published for the first time verified — §4.4-§4.5

  "ℒ = α·ℒ_ranking + β·ℒ_language + γ·ℒ_miscellaneous, s.t. α+β+γ = 1, α,β,γ ≥ 0" "These labels are used in a cross entropy loss over the item catalog (or candidate set), teaching the model to produce a high-quality ordering of items given the verbalized context." "When the catalog is too large to score exhaustively, we can combine the scoring with sampled-softmax instead." — GenRec arXiv, §4.4-§4.5

  • The loss is a convex combination of three terms.
  • The ranking loss is a softmax cross-entropy over the whole catalog (or candidate set), replaced by sampled softmax when the catalog is too large.
  • The specific values of α/β/γ are not disclosed.
  • It is also newly stated that only the ranking head is used at inference time, and the LM head serves to regularize training.
  • The paper leaves "future text generation use cases (e.g., natural-language explanations)" open as a future direction.

• The training data schema is a user/assistant conversation format, and raw engagement signals go into the assistant side verified — §4.2

  "we transform a member's historical interaction data into a single-turn or multi-turn 'conversation' between a user and a recommender. Each turn pairs a user message (combining context, history, and a task) with an assistant message representing the member's actual engagement for that turn." — GenRec arXiv, §4.2

  • The blog post listed only the user message side ("verbalized context, profile, history, item metadata, and task").
  • The paper disclosed the schema in which the assistant message holds the actual engagement (play, play duration, abandon, thumb, etc.).
  • It also stated the definition of a positive label as "high-quality engaged events (e.g., high-duration plays, strong explicit feedback), with denoising logic and thresholds that vary by content types".
  • Thresholds and denoising logic exist per content type. The specific thresholds are not disclosed.

• GRPO did outperform SFT in the preliminary experiments verified — §4.6

  • The blog post went only as far as saying "future work because of higher cost".

  "In preliminary experiments, RL-style methods (e.g. GRPO) showed additional gains over supervised finetuning, but their high training overhead makes them better suited for future work. Therefore, currently we adopt the reward-weighted loss approach as the main alignment mechanism for GenRec, due to its simplicity, stability, and cost considerations." — GenRec arXiv, §4.6

  Netflix tried it and got better results but did not deploy it. This contrasts with Meta ConnectionMind actually deploying GRPO; in effect, the two companies made opposite decisions about the same algorithm.

• A data scaling sweep was run verified — §5.2
  • The experiment increased Phase-2 data from 1x to 20x on each of two backbones, ~1B and ~10B.
  • Both showed monotonic improvement and a similar scaling pattern. The absolute curves and slopes are not disclosed.

• The paper states explicitly that GenRec is not in the semantic ID (SID, a representation of items as hierarchical codes) family verified — §2.2

  "Most LLM-based generative retrieval systems rely on autoregressive decoding with beam search at inference time, which introduces latency overhead that can be prohibitive at scale, especially for large candidate sets. Our work avoids these limitations by augmenting the base LLM architecture with a catalog-aware ranking head, so that it can rank a large candidate set in a single forward pass" — GenRec arXiv, §2.2

  The paper itself sorts out how GenRec differs from SID and generative families such as Google PLUM, Kuaishou OneRec-Think, Spotify GLIDE and Yandex Gryphon. GenRec is not in the semantic ID family. This is because its scoring head is catalog-aware and scores a large candidate set in a single forward pass.

• The team structure is revealed for the first time verified — §8

  • The blog post had no team breakdown. The paper states that three teams collaborated.
  • AI for members (22 people): includes the main author group and Justin Basilico
  • AI platform and serving (13 people): serving infrastructure
  • Product (4 people)
  • Justin Basilico is known as the head of personalization at Netflix. This is the first time his name has been explicitly confirmed in a document adjacent to GenRec.

• The paper does not confirm a full rollout either verified — §7

  "GenRec represents an initial step toward a more LLM-centric recommendation stack at Netflix. Our results suggest that this is a viable and promising path for large-scale personalization in at least some scenarios" — GenRec arXiv, §7 Conclusions

  The phrases "initial step" and "at least some scenarios" show that stance. The open questions the canon has kept, namely whether there is a full rollout, whether it runs on a real-time surface, and who generates the candidate set, cannot be answered from the paper either, so they remain.

• The +0.006% core metric
• The Phase 1/2 contribution breakdown and the time axis
• The 5000→1700 token example and serving cost down to 1/3
• ℒ = αℒ_rank + βℒ_lang + γℒ_misc and the catalog softmax
• The user/assistant conversation schema
• The preliminary GRPO result
• The mention of a distilled backbone
• The positioning relative to Semantic ID
• The three-team structure
• Items that are still not public in the paper

The open questions were kept as they were, because the paper does not answer whether there is a full rollout, whether it runs on a real-time surface, or where the candidate set comes from.

1. The +0.006% core metric can now be quoted. With only the phrase "statistically significant", the scale could not be gauged. Used together with the paper's own characterization ("meaningful at Netflix scale"), the weight of the claim becomes clear.
2. A new combination is established: context compression not only preserves quality but also lowers serving cost to 1/3. This is the key evidence for the argument that GenRec's batch-compute serving is economically viable.
3. The statement that GRPO produced better results but was not deployed contrasts with Meta ConnectionMind actually deploying the same GRPO. This is the first moment the industry-level trade-off, "how much training cost a company can bear decides the deployment form", appears symmetrically in two cases.

The relationship between the blog post and the paper is itself suggestive. Netflix completed the sequence of releasing the blog post first and then reinforcing it with quantitative figures in a paper in a little over ten days (blog 2026-07-30 → paper 2026-08-10). If this sequence repeats, this repository can settle on a scan strategy of folding a blog post into the canon immediately and adding the quantitative figures when the paper appears.

↑ Contents

3Yandex Music released Sona, a unified ranker for the smart speaker surface — Radar

• Sona is a unified generative recommendation model for music streaming verified — abstract
  • It processes the user engagement sequence with a shared encoder, then produces recommendations with an autoregressive decoder and a ranking module.
  • It does not use hand-engineered features.
• The deployment surface is "My Vibe on smart speakers", a surface dedicated to smart speakers. It differs from Gryphon-v2, which covers the broad music recommendation cascade.
• In the production A/B it stated "replaced the entire production cascade", and active users rose +4.53% and total listening time rose +6.30% verified — abstract

No canon update Because it is not a tracked service (X, Meta, Netflix), only a 3–4 line pointer was added to the Gryphon-v2 thread in radar.md as 2026-08-13 additional evidence.

↑ Contents

Corrections

⚠ The claim in role 3 was about half overturned. The old concluding sentence in topics/llm-in-recsys.md was rewritten.

• Old conclusion: "There is no first-party statement from either X or Meta that an LLM enters the organic feed ranking path — across the industry, Netflix GenRec was the first to make that statement"
• New conclusion: "Netflix GenRec (2026-07, LLM = scoring function) was the first, and Meta ConnectionMind (2026-08, LLM = graph path reasoner) was the second"

That the role the LLM plays differs between the two companies is newly established ground. The corpus used for verification is the original text of the 2 arXiv papers read this cycle, which are quoted in sections 1 and 2.

There is also one thing that was not overturned. The statement "Netflix GenRec's scoring structure is not text generation — the LLM is the encoder that produces the pooled hidden state h, and final scoring is an embedding dot product / small MLP" stands as written. The catalog softmax and scoring head φ description in §4.5 of the arXiv paper matches this summary. The newly published loss formula strengthens this statement and does not weaken it.

↑ Contents

Radar

Today the radar updated two items.

• Added Sona (Yandex Music, arXiv 2608.11015, 2026-08-11) to the Gryphon-v2 thread as a 3–4 line pointer. Sona is a unified ranker for the smart speaker surface, and it recorded active users +4.53% and total listening time +6.30%.
• Updated the LLM-native ranking spread open thread in radar.md. It reflects that Meta ConnectionMind entered as the second production claim in role 3, and "ConnectionMind surface confirmation and serving-cost follow-up post" was added to the watch list.

In radar triage, of the 30 new arXiv cs.IR papers scanned today, no item other than the 3 above passed the acceptance criteria.

• The 3 above: 2608.10257 Netflix GenRec, 2608.10187 Meta ConnectionMind, 2608.11015 Yandex Sona
• Acceptance criteria: platform-company authors, production/A/B statements, mechanism novelty, scale results
• Representative examples that did not pass fall into the three families below, and none of them has a statement about production deployment.
  • LLM re-ranker family: 2608.10528·2608.10406
  • Multimodal sequential rec: 2608.10700·2608.10240·2608.10983
  • Academic benchmark CF: 2608.10297·2608.10247

↑ Contents

What to check next

↑ Contents