ReST: A recommendation-native sequence Transformer for industrial ranking

简体中文

Public-data dual-gated temporal encoding and reusable user/candidate path.

Paper information

Field	Details
Paper	arXiv 2609.01240
Institution	ByteDance.
First public release	2026-09-01 (arXiv v1)
Author code	No public author implementation found as of 2026-09-05.
Local adapter or method key	rest
Local reproduction code	src/auto_research/reproductions/rest/

Paper method and evidence

Dual-gated attention suppresses behavior noise. RoPE and relative time represent both order and elapsed time. A heavy user encoder runs once per request, then a light cross decoder scores each candidate.

flowchart LR
  H[Behavior history] --> U[Heavy dual-gated user encoder]
  U --> D[Reuse user state]
  C[Candidate] --> R[Light cross decoder]
  D --> R

Original paper figure

The original figure shows the compute-once, decode-many architecture. From the original paper; copyright remains with the authors.

Core equation

\(h_u=\mathrm{DualGate}(H)\) and \(s(u,i)=\mathrm{CrossDecode}(h_u,e_i)\); reuse \(h_u\) across the request.

Reported results

The one-week online A/B reports online AUC +1.31%, core revenue +11.93%, P99 50 ms, followed by full rollout.

Local reproduction

MovieLens-100K, fixed 220 users/360 items: temporal decay, signal gates and candidate crossing execute on CPU. NDCG@10 0.05401→0.05372 (-0.53%). Three-seed metrics. No private logs, production checkpoint, shared-prefix serving kernel or CUDA claim. The negative public-data result does not contradict or reproduce online lift.