ByteDance's OneTrans-V2 unifies retrieval, pre-rank and fine-rank, lifting GMV 9.74%
_reachsumit · x · 2026-09-25
ByteDance's OneTrans-V2 paper unifies the retrieval, pre-rank, and fine-rank cascade of industrial recommenders into a single jointly trained Transformer.
- User behavior sequences are encoded once as shared context while preserving stage-specific candidate features, eliminating repeated encoding and isolated optimization
- Joint training enables mutual reinforcement across stages plus in-model distillation from fine-rank to pre-rank
- Sparse MoE scales the shared backbone with bounded activated compute, stabilized via μP-style parameterization
- Decision-Conditioned Generative Retrieval (DCGR) generates a decision prefix describing the upcoming interaction, then generates items conditioned on it, steering one generative process with business objectives
Online deployment lifted GMV by 9.74%.
More from Research
- PosteriorBench accepted to NeurIPS: point reconstruction accuracy misleads on inverse solvers — AnimaAnandkumar · 2026-09-25
- Dietterich: pretraining's next-token task already forces models to abstract and generalize — tdietterich · 2026-09-25
- Neural Spectral Capacity: a training-free architecture score that beats #Params and prunes LLaMA-7B 5900x faster — CityU-HongKong · 2026-09-25
- HRI 2027 Adds Archival Industry White Paper Track for Real-World Robot Deployments — petitegeek · 2026-09-25
- Study: Reasoning hurts 15.7% of multimodal embeddings; training-free SURE router fixes it — _reachsumit · 2026-09-25
- Google scales learned cross-task relationships in YouTube's production recommender — _reachsumit · 2026-09-25