One learned system selects and orders recommendation slates
OneTrans-GSR: A Generative Slate Ranking Paradigm Unifying Reranking and Mix-Ranking with One Autoregressive Transformer · arXiv:2611.04720
In plain words
A recommendation system often scores items first, then uses separate rules and tools to decide which products, videos, livestreams, and ads appear together and in what order.
OneTrans-GSR is a method that makes those choices in one process, selecting each item after considering the user, available items, and earlier choices.
It learns from scores assigned to whole slates, but uses its generator alone when serving recommendations. In one production test, replacing the existing system improved measured business outcomes and reduced latency; the test covered one system and the complete replacement.
The research
A slate’s value can depend on which items appear together and where they appear, yet many recommendation systems assemble it with hand-built rules. OneTrans-GSR puts selection and ordering into one learned decision.
Its production test changes the evidence readers can expect from that approach: the complete replacement improved measured business outcomes and reduced latency in the tested system.
OneTrans-GSR combines reranking and mix-ranking in one slate-generating policy, trained with rewards for complete slates and constrained during serving.
For offline evaluation, the decoder sampled candidate slates and selected one by its sequence score; reward models then compared it with the logged exposure slate. The selected slates scored higher on click, order, and gross merchandise value measures, but these are model estimates rather than observed outcomes.
A randomized production test compared the complete replacement with the deployed system. It found improved business measures and lower latency across the replaced stages, supporting the method in that production setting.
The online test covered a complete system replacement in one production recommendation system, so it cannot identify which components produced the changes. Offline scores for generated slates are model estimates; the logs contain no observed outcomes for those counterfactual slates.
What the evidence shows
- The slate is the decision
OneTrans-GSR is a recommendation method that combines reranking and mix-ranking: it selects and orders candidates from each request’s pool to create the final slate. The paper addresses a design constraint in existing systems, where hand-built rules and specialized components restrict which slates can be assembled and require manually tuned trade-offs. Its central question is whether one learned system can make the whole slate decision.
- One process selects and orders
The method generates the slate one candidate at a time, conditioning each choice on the user, the available candidates, and the items already selected. A reward model scores complete slates during offline training, while the decoder generates the served slate directly and applies business constraints during generation.
This division lets the training signal consider whole-slate outcomes without requiring the reward model or a separate evaluator at serving time.
- The score responds to order
In controlled tests, the reward model using a final-state readout changed its ListGMV score much more when researchers swapped slate positions than models that summed item scores. This shows that the readout was more sensitive to order in those tests.
Because the altered slates had no outcome labels, the comparison says nothing by itself about whether their actual value rose or fell.
- Offline estimates meet a live test
In offline evaluation, OneTrans-GSR’s selected slates scored above logged exposure slates on the reward models’ click, order, and gross merchandise value measures. These are estimated score differences, not observed outcomes for the generated slates.
A randomized production test supplied separate evidence: against the deployed Generator–Evaluator system, the complete replacement improved measured business outcomes and reduced latency across the replaced stages.
- The conclusion has a boundary
The production comparison supports the complete OneTrans-GSR replacement in the tested e-commerce system. Since the test changed the whole reranking and mix-ranking system at once, it cannot show which component drove the results.
The offline logs also lack observed outcomes for generated alternatives, so their score gains remain model-based estimates.
Deep
- Original paper authors
- Feng Cheng, Yu-Hsuan Huang, Fengyi Wang, Chaofan Zhou, Zhennan Zhu
- Submitted to arXiv
- License
- arXiv non-exclusive