SORT: A Systematically Optimized Ranking Transformer for Industrial-scale Recommenders¶

Core public ranking mechanism.

Paper information¶

Field	Details
Paper	Paper or original source
Institution	Alibaba International Digital Commerce.
First public release	2026-03-04 (arXiv v1)
Author code	No official author code linked by the paper as checked on 2026-08-24.
Local adapter or method key	`sort-ranking`
Local reproduction code	`src/auto_research/reproductions/sort_ranking/`

Paper method and evidence¶

Redesign feature tokenization, attention and FFN, with hardware-oriented long-sequence compression so one Transformer replaces a traditional DLRM.

flowchart LR
  A[Feature tokenization] --> B[Compress long sequence]
  B[Compress long sequence] --> C[Unified ranking Transformer]

Original paper figure¶

Figure shows hardware-aware ranking architecture. From the original paper; copyright remains with the authors.

Core equation¶

\[ H'=H+\operatorname{MHA}(H),\quad H''=H'+\operatorname{FFN}(H') \]

Reported results¶

Section 4.7/Table 5 reports orders +6.35% across three production scenarios.

Local reproduction¶

MovieLens-100K, 220 users/360 items, seed 42, same split/candidates: shared transition+content baseline vs independent SORT state and scorer. NDCG@10 0.0540→0.0469 (-13.11%), Hit@10 0.1091→0.0909. Validation selects blend; test never tunes.

auto-research reproduce --paper sort-ranking --dataset-dir data --seed 42

Fixed metrics. No private features, production traffic or serving. These local public-data results do not reproduce online lift.