0x00 한

Research / Daily reports

  • Sources read: 2
  • Canon files updated: 0
  • Radar entries: 1
In one lineByteDance reported merging reranking and mix-ranking into a single slate-generating decoder deployed on its e-commerce feed. The tracked services had no new sources.

Other changes reviewed today:

  • Reddit started testing ads on search results pages. It is a search surface rather than a feed, so we did not fold it into the canon.

1ByteDance builds the impression slate with one generative policy — radar

Source OneTrans-GSR: A Generative Slate Ranking Paradigm Unifying Reranking and Mix-Ranking with One Autoregressive Transformer — ByteDance global e-commerce recommendation team, 2026-11-03 v1, paper.

Background Industrial recommendation places reranking and mix-ranking after the ranking that scores each candidate. These stages build the slate (the set of items shown at once) that is actually sent to the screen, and mix-ranking mixes candidates from different channels, such as product cards, short videos, live streams, and ads, on one screen. These later stages rely on handcrafted rules and small models, so people tune the balance among objectives. This paper turns the two stages into a training problem in which one decoder generates the slate item by item, and it reports the result from a production A/B test.

What we learnedA. A generative policy trained with a reward model that sees the whole slate replaced the rule-based later stages

  • The decoder picks T items from the candidate set without duplicates and builds an ordered impression slate verified — arXiv HTML.
    • This decoder extends the OneTrans backbone (the same team's large-scale Transformer ranking model) that the team has already deployed.
    • The previous production approach was a Generator–Evaluator structure, in which several generators produced candidate slates and a separately trained evaluator picked one.
  • The training reward comes from a reward model that evaluates the whole slate; this model is used only in training, so it is not bound by serving cost constraints verified — arXiv HTML.
    • The efficiency objectives are ListClick, ListOrder, and ListGMV, and the long-term value objective predicts repurchase within 14 days.
    • The scores for the multiple objectives are combined into a single reward by a geometric mean (GMR-PO).
  • Business constraints are applied at each decoding step through hard masks and finite logit adjustments verified — arXiv HTML.
    • Per-tag impression frequency control and continuous score adjustment are introduced this way.
  • The paper reports GMV/u +2.63%, advertiser value +2.06%, and repurchase rate +2.03% in an online A/B test verified — arXiv abstract.
    • It writes that end-to-end latency across the two stages, reranking and mix-ranking, fell by 13%.
    • We could not find the experiment duration or the traffic share verified.

"Deployed in a large-scale industrial recommendation system, OneTrans-GSR mixes heterogeneous content types within a unified slate. In online A/B tests, it improves GMV/u by 2.63%, advertiser value by 2.06%, and repurchase rate by 2.03%, while reducing end-to-end latency across the reranking and mix-ranking stages by 13%." — OneTrans-GSR (paper), 2026-11-03

Canon update Not folded into the canon. Added an entry to ## New this month in radar.md.

Why it matters This adds one more piece of production evidence for the shift toward generative models even at the slate stage. It is not evidence about a tracked service, and it concerns an e-commerce surface, so it goes on the radar. Combined with the same team's OneTrans-V2 (ranking cascade unification), ByteDance e-commerce is moving everything from retrieval to the slate onto a single Transformer family inferred. The basis is that both papers use the same OneTrans backbone as their starting point.

2Reddit tests ads on search results pages — Reddit

Source We Are Testing Ad Placements in Reddit Search Pages — Reddit newsroom, 2026-11-03, blog post.

Background The canon's description of the Reddit ads funnel covers the eligibility filter, a lightweight ranker, the heavy prediction model, and the auction order, and it does not cover which surfaces carry ads. Search is a surface where users write down exactly what they want, so purchase intent is high. This announcement says Reddit is testing ad placements on search results pages.

What we learnedA. Search ads appear based on several factors, including contextual relevance to the query, and start with only some advertisers

  • Search ads use the existing ad targeting on search results pages and appear based on several factors, including contextual relevance to the search query verified — Reddit newsroom.
    • As an example, it writes that a search for "espresso station" may show ads for espresso machines, coffee beans, and mugs.
  • It writes that the test starts with only some qualified advertisers and will be refined based on user reactions verified — same post.
    • It did not disclose the ranking approach, the placement positions, or the share of users in the test.

"Ads in search may appear based on a range of factors, including contextual relevance to a user’s search query." — We Are Testing Ad Placements in Reddit Search Pages (blog), 2026-11-03

Canon update Not folded into the canon. This is an ads announcement for a search surface rather than feed recommendation, and it does not disclose the ranking mechanism, so it stays in the daily only as a boundary source.

Why it matters The product fact that Reddit's ad surface is expanding beyond the feed into search is confirmed, but it does not change our understanding of the ads funnel. The basis is that the announcement describes the selection factors only as "a range of factors".

Radar

What to check next