
ByteDance's Learned Shopping Feed Still Boosts Ads by a Hand-Set Number
ByteDance's OneTrans-GSR hands the assembly of an e-commerce feed to one model trained against a slate-scoring reward and reports gains in GMV per user, repurchase and advertiser value. Advertiser value is not in that reward but is steered by a constant ad boost chosen by people, which the paper's own sweep shows costs shopper conversion and whose production setting it never discloses.
Neritus Vale
ByteDance’s e-commerce engineers have handed the assembly of a shopping feed to one learned model and call it a move “beyond human-engineered heuristics”. The model now decides where product cards, short videos, livestreams and ads sit. How hard it pushes the ads is chiefly one constant chosen by people, and the paper does not say where it was set.
OneTrans-GSR was submitted to arXiv on 3 November by ByteDance’s Global E-Commerce Recommendation Foundation Team. It replaces reranking, which picks and orders items already scored upstream, and mix-ranking, which interleaves content from separate business channels. One decoder now writes the slate, trained by reinforcement learning against a reward model that scores whole slates. The setting is named only as “a large-scale industrial recommendation system for e-commerce”.
The live test pitted the new policy against the old cascade. Accounts were split at random for a week, with “the business rule set” held fixed. The paper does not say whether the ad boost, or the old system’s equivalent, sits inside that set, though the introduction lists “commercial boosting” among the old system’s hand-built rules. Gross merchandise value (GMV) per user rose 2.63% and the repurchase rate rose 2.03%; both are statistically significant. The paper never defines the GMV denominator or reports a conversion rate.
The GSR abstract promises to generate the exposure slate “end-to-end”, and the introduction is more careful. Kuaishou’s OneRec paper said in 2025 that it “replaces the cascaded learning framework with a unified generative model”. GSR’s introduction keeps the rules in view: “What changes is the role of the rules rather than their presence.”
The rules that survived are the ones we called merchandising policy in July. During decoding, masks and score adjustments govern spacing between content types and categories, product deduplication, exposure quotas and “advertising exposure”. A separate training reward, weighted at one-tenth of the main one, favours less fatigue, more discovery and cold-start exposure. The authors concede that reward weighting “cannot enforce position-specific requirements”.
Against the old system, those rules do work the learning has not absorbed. In offline tests without the decoding adjustments, GSR’s slates showed 3.41% more fatigue than the old system’s: more repeat exposure to unclicked categories. Their share of novel items fell too. With the adjustments on, fatigue dropped 7.65% below the old system’s, though discovery still trailed it. Against score-sorted baselines, the learned policy alone cut fatigue by 19.41% to 40.91%.
Advertiser value is not among the reward model’s objectives; the introduction says it is “controlled during decoding”. The reward blends predicted clicks, orders, GMV and repurchase, with weights “tuned as hyperparameters” and left unpublished. Yet the paper faults the hand-built systems it replaces for “manually tuned fusion weights”. Ads get something blunter: a constant boost to every advertising candidate’s score; raising it “trades user engagement for advertiser value”.
The paper’s sweep shows what that dial is worth. Advertiser value rose at every step tested: 0.71% at a boost of 0.05, 1.83% at 0.1 and 4.61% at 0.3, each measured against no boost, as the chart plots it. Every nonzero setting also lowered the user-level conversion rate, by about 0.5% at the two lower settings and about 0.74% at the top. The authors conclude: “Selecting a production operating point requires considering both outcomes alongside rule compliance.”
In production, advertiser value rose 2.06% against the old cascade, and the paper does not say which boost was served. The sweep’s top setting added 4.61% on its own, but against no boost rather than against the old system, and the paper does not say whether the sweep ran online or offline. The metric is never defined; the chart labels its axis “ROI2”. No reader can separate what the policy learned from where the dial was set.
The production table is fuller than Kuaishou’s and still has no conversion rate. The Kuaishou ad generator covered here on 2 November was scored online on revenue alone; this table adds advertiser value, repurchase and latency to GMV per user. The paper calls its slate checks “offline proxies rather than direct measures of user response”. Neither those checks nor the table count the slots that went to ads.
A live test contradicted the reward model’s own verdict. An arm trained against a single reward model scored higher offline but lifted GMV per user by only 1.41%, Table 9 shows. The authors call that “consistent with errors in the proxy reward”, a hazard OpenAI researchers measured in 2023, but say the test cannot identify the cause. The ensemble arm’s 2.63% matches the headline gain, which suggests that is the scorer ByteDance shipped.
Stage by stage, ByteDance’s OneTrans papers move the cascade into models and leave the ad trade-off outside as a dial. In September, OneTrans-V2 folded retrieval and ranking into one transformer, with ads among the goals steered by a “business offset” whose “strength is exposed as an online parameter”. It also printed a sponsored share: 33.7% of its retriever’s exposures were sponsored items. GSR, by the same corresponding author, prints none for the final slate.
The strongest objection is that a dial is the right design. How much of a feed to sell is a business call, and an explicit setting is easier to audit than a buried weight. GMV per user and repeat purchases still rose under the same rules. For the thesis to fail, the dial would have to matter little: the policy would reach similar outcomes wherever it was set.
The paper’s own evidence says the dial matters. One constant moves advertiser value by amounts that bracket the production gain, at a conversion cost the sweep records. The paper defines no GMV denominator and reports no production conversion rate, so that cost has no counterpart to compare against. An explicit setting is auditable only once it is printed, and this one is not.
The paper never mentions apparel or sellers, so what follows for clothing sellers is extrapolation. If a feed like this carried a seller’s promoted placement, the ad boost would decide how far a paid card is lifted over organic items. A similarity penalty also scores any item resembling those placed above it lower, so a second black dress like the first scores below it. Neither setting’s production value is published.
ByteDance automated the layout and kept the lean toward advertisers as a setting.
If platforms keep reporting advertiser gains without the boost behind them, sellers cannot tell a better model from a higher setting. The remedy takes three numbers: the served boost, the conversion change and the share of slots given to ads. No amount of compute decides how much of a feed to sell, and ByteDance’s paper shows it is still a choice.