OneTrans-GSR: A Generative Slate Ranking Paradigm Unifying Reranking and Mix-Ranking with One Autoregressive Transformer

OneTrans-GSR. A Generative Slate Ranking Paradigm Unifying Reranking and Mix-Ranking with One Autoregressive Transformer. By Feng Cheng and fifteen others, from ByteDance. Published on the third of November.

Industrial recommender systems spend a great deal of computation scoring candidates one at a time. Then, at the very end, the final exposure slate, the set of items a user actually sees, is assembled by separate reranking and mix-ranking stages that run on hand-crafted rules and specialized, low-capacity models. Mix-ranking is the stage that combines product cards, short videos, livestreams and advertising cards, each supplied by a different upstream business channel.

Those rules and models encode valuable domain knowledge. The trouble is that they are inherently limited in how much additional data and computation they can put to use. They constrain the combinatorial space of slates that can even be considered, and they settle the trade-offs between objectives by manual tuning. The last step before the user, in other words, is the one step that does not get better when you give it more.

The paper reformulates reranking and mix-ranking as a single learning problem, one whose performance can improve with computation. That formulation is instantiated across both stages as OneTrans-GSR: a single decoder-only policy that extends the team's deployed OneTrans backbone so that it generates the exposure slate end-to-end. Instead of rules picking and ordering items, one autoregressive model writes the slate.

Training and decoding each carry part of the load. During training, a reward model evaluates complete slates, not individual items, and that reward guides reinforcement learning toward slate-level and long-term objectives. During decoding, business constraints are enforced explicitly, so the requirements that the old rules existed to protect are still respected while the policy does the composing. In deployment, the same policy mixes heterogeneous content types within one unified slate.

Offline, the scaling behaviour the reformulation was meant to unlock does appear: slate quality consistently improves as reward-model compute and policy compute increase. Online, in A/B tests in a large-scale industrial recommendation system, OneTrans-GSR improves GMV per user by two point six three percent, advertiser value by two point zero six percent, and repurchase rate by two point zero three percent.

Those gains come with a reduction in cost rather than an increase. End-to-end latency across the reranking and mix-ranking stages falls by thirteen percent, even though two stages of rules and specialist models have been replaced by one learned policy.

The authors frame this as a paradigm shift for slate generation, away from human-engineered heuristics and toward a general learned policy. Stated plainly, the argument is about where future improvement comes from. A rule only gets better when someone rewrites it. A policy trained against a slate-level reward gets better as compute is added, and that, by their account, opens a scalable path for further algorithmic optimization.