0x00 한

Research / Daily reports

  • Sources read: 3 (all radar candidates)
  • Canon files updated: 0
  • Radar entries: 2 new and 1 open thread update
In one lineThere was no new announcement from the tracked services (X, Meta, Netflix) that would change the canon. Instead, we added two arXiv papers with production claims to the radar.

Other changes reviewed today:

  • ByteDance ads said it deployed in production a ranking model whose sequence length is extended to 20K.
  • Webtoon presented an RL post-training stage that aligns a recommendation foundation model with business metrics.
  • HD-Rec, with Kuaishou researchers participating, has only public benchmark results, so we added only one line to the radar open thread.

1ByteDance ads deployed a ranking model with sequence length 20K in production — Radar

Source TM20K: Teacher Retains Full Tokens, Student Merges Efficiently — E-Commerce Sequence Modeling in Ad Recommendation — arXiv 2608.07055, 2026-08-07, paper (12 ByteDance authors)

Background The claim that "feeding long user sequences directly into the ranking model works better" is currently being tested in different directions at three tracked services.

  • Meta HSTU/GEM (Meta's two sequence-model lineages; HSTU is on the organic side, GEM on the ads side) claim scaling that grows parameters and sequence length together verified — topics/sequence-transformer-ranking.md
  • LLaTTE, the user model of Meta ads, set the opposite condition: "scaling holds only with semantic features" verified — daily/2026-08-09.md
  • Netflix Foundation Model (FM) saw "consistent improvements" in the 2M→1B shadow verified.
    • A shadow (production-shadow) is an evaluation stage that runs alongside live traffic without showing the results to users.

The three statements half fit together and half collide. The ByteDance paper posted to arXiv on August 7 (read today) adds a fourth statement that points in yet another direction. Instead of growing parameters, this paper expanded sequence length to 20K. To achieve this while barely increasing production serving latency, it chose teacher/student KD (knowledge distillation, a method in which a student model learns to follow the signal produced by a teacher model).

What we learnedA. Both teacher and student use full attention, and the student gets lighter through token merge

  • Both teacher and student use full attention rather than target-only attention (a common optimization that folds attention for a single query) verified.

    "Both teacher and student models take the full attention mechanism rather than pure target-sequence attention for effective sequence scaling." — TM20K paper abstract, 2026-08-07

  • The student gets lighter not by folding attention but by compressing the sequence itself with token merge.

  • The teacher is trained heavily to produce the correct signal, and the student follows that signal through KD to raise its performance.

What we learnedB. Deployed in production at sequence length 20K, with serving latency up 5.6%

  • The paper reported ADSS +1.036% and serving latency +5.6% at sequence length 20K, and said training and serving cost is nearly the same as the existing SOTA verified.

    • ADSS is a ByteDance ads metric. The exact definition must be checked in the paper body.

  • The paper states that it deployed this model in ByteDance's e-commerce ad recommendation system verified.

    "successfully deployed in ByteDance's e-commerce advertising recommender system" — Same abstract, 2026-08-07

  • If it is true that latency rose by only 5.6%, the serving burden of the 20K expansion is smaller than the market expected.

Canon update None. It does not belong to a tracked service, so we did not edit any canon file. Instead, we added it as a new entry in the "New this month" section of radar.md.

Why it matters The claim that long sequences are good gets stronger, but companies still give different answers on how to make them servable. TM20K is a data point from ByteDance following Meta and Netflix, which the scaling law thread covers, but unlike HSTU, LLaTTE, and Netflix FM, which talk about parameter scaling, it handles sequence-length scaling with KD. If statements that use KD for very long sequences accumulate at three or more places, we will decide whether to promote it to a topics/ note.

2Webtoon added RL post-training to a recommendation foundation model — Radar

Source Progressive Alignment of Recommender Foundation Model through Multi-Phase Post-Training — arXiv 2608.06792, 2026-08-07, paper (Webtoon, RecSys '26 Industry Track)

Background A recommender foundation model (RFM) is an approach in which one pretrains once and reuses the model across multiple serving surfaces such as the home feed, For You, and search. Netflix Foundation Model and Meta GEM are representative examples of this family (services/netflix.md, services/facebook.md). Per-surface adaptation is usually handled with SFT (Supervised Fine-Tuning, a method that fine-tunes on a downstream task with supervised learning). However, when SFT targets immediate signals such as click or like, it tends to diverge from the business metrics one actually wants to optimize; those are sparse, delayed signals such as retention and revenue.

The Webtoon paper posted to arXiv on August 7 presents an approach that resolves this mismatch with three-stage post-training. It carries the SFT → RLHF sequence familiar from LLMs over to recommendation as is.

What we learnedA. Post-training is split into three stages: LP, FFT, and RFT

  • The pipeline has three stages, LP → FFT → RFT verified.
    • LP (Linear Probing) freezes the pretrained representations and first stabilizes only the randomly initialized downstream head. It is a warm-up stage that keeps the head from swinging widely. The downstream head is the layer attached on top of the pretrained representations to produce the output of the downstream task.
    • FFT (Full Fine-Tuning) specializes the whole model for the target task.
    • RFT (Reinforcement Fine-Tuning) uses a learned reward model to align the policy with business metrics. Here the policy is the recommendation model actually served.

What we learnedB. Business metrics are used only to train the reward model, and the policy is trained on dense signals

  • The core design is to train the policy on dense implicit feedback (implicit behavior signals that accumulate often, such as clicks) and to use sparse business targets only for the reward model verified.

    "Rather than directly optimizing the serving policy on sparse business targets, we train the policy on dense implicit feedback and use business-metric supervision only for reward modeling." — Progressive Alignment abstract, 2026-08-07

  • The reason is clear. Optimizing directly on a sparse target (for example, payment conversion) leaves almost no gradient flowing.

  • The reward model acts as an intermediary between dense and sparse signals, and the policy is stable because it is trained on dense signals.

What we learnedC. It states online A/B results but gives no numbers

  • The abstract states online A/B results but gives no numbers verified. On the A/B it writes only this: "Large-scale online A/B tests further show that the proposed framework improves production recommendation quality over a conventional non-foundation baseline".
  • The specific metrics and figures must be checked in the body.

Canon update None. We added it as a new entry in the "New this month" section of radar.md.

Why it matters The Webtoon paper goes one step further at the point of attaching a downstream head to a foundation model, adding one more RL alignment stage. Our canon (services/netflix.md, Foundation Model section) is still thin on how Netflix FM attaches downstream heads. If this approach is reproduced at two or three more places, post-training, the second half of the standard recsys FM pipeline, will need to be redrawn; for now it is one company's first statement, so we only keep watching.

Radar

Of the 30 new arXiv cs.IR papers screened today, 3 passed the radar acceptance criteria (we look for platform-company authors or production claims, novelty of mechanism, and large-scale results). We created new entries for only two of them.

  • TM20K (ByteDance) is the paper in section 1 above. We created a new entry in radar.md New this month.
  • Progressive Alignment (Webtoon) is the paper in section 2 above. We created a new entry in the same place.
  • HD-Rec (arXiv 2608.06997, 2026-08-07) has only public benchmark results and no production claim yet. So we did not create a new entry and only appended one line, "2026-08-11 additional evidence", to the generative retrieval / semantic ID open thread.
    • Kuaishou's senior line (Ruiming Tang, Guorui Zhou, Han Li) took part.
    • It extends Semantic ID (SID, a representation of items as hierarchical codes) to cross-domain. The means are a hierarchical codebook and domain sparse MoE (mixture of experts, a structure that selects and mixes several expert sub-networks through a gate).
    • We record it only as a signal that the Kuaishou line has joined the SID family.
  • The remaining 27 were rejected as purely academic increments: RAG optimization, academic retrieval methodology, geo/legal/edu domains, and recsys limited to the music domain.
  • The industry blog sweep is done on Mondays, so we did not do it today. The next sweep is on 2026-08-17 (Monday).

What to check next