ByteDance's OneTrans-GSR Unifies Reranking with One Transformer, Lifts GMV/u 2.63% in A/B Tests
ByteDance · hf · 2026-11-09
ByteDance has published OneTrans-GSR, a generative slate ranking paradigm that unifies reranking and mix-ranking in industrial recommender systems as a single learning problem.
Method
- A single decoder-only autoregressive policy built on the deployed OneTrans backbone, generating the exposure slate end-to-end (mixing product cards, short videos, livestreams, and ads).
- A reward model evaluates complete slates during training, guiding RL toward slate-level and long-term objectives; business constraints are explicitly enforced at decoding.
Results (large-scale online A/B tests)
- GMV/u +2.63%, advertiser value +2.06%, repurchase rate +2.03%
- End-to-end latency across reranking and mix-ranking cut by 13%
The authors frame it as a paradigm shift from hand-crafted heuristics toward a general learned policy that scales with compute.
More from Infra
- Short seller alleges over 20% of Nvidia's FY2026 compute revenue still flows from China despite 'zero' claim — FuSheng_0306 · 2026-11-09
- Qwen-Image 2.1 Turbo (20B) runs fully on a phone NPU: 40s per image on Snapdragon 8 Gen 1 — shirosoro · 2026-11-09
- Leaker: Huawei Ascend DT production in 2026 to far exceed the rumored 200K chips — zephyr_z9 · 2026-11-09
- AI writes code faster than Uber can merge it: how SubmitQueue was rebuilt for the AI era — AI Engineer · 2026-11-09
- Lam Research may owe Samsung around ₩1 trillion over alleged leak at Korean unit — zephyr_z9 · 2026-11-09
- Leak: CXMT to start HBM mass production within a month, Ascend 960PR seen doing well in China — zephyr_z9 · 2026-11-09