UniScale: Synergistic Entire Space Data and Model Scaling for Search Ranking¶

Core public-data mechanism, separate from online evidence.

Paper information¶

Field	Details
Paper	arXiv 2603.24226
Institution	Taobao & Tmall Group / Alibaba.
First public release	2026-03-25 (arXiv v1)
Author code	No official author code linked by the paper as checked on 2026-08-24.
Local adapter or method key	`uniscale`
Local reproduction code	`src/auto_research/reproductions/uniscale/`

Paper method and evidence¶

ES³ expands training beyond exposed items; HHSFT hierarchically combines heterogeneous examples. Scale data and model capacity jointly with low-cost deployment.

flowchart LR
  D[Entire-space data expansion] --> H[Hierarchical heterogeneous training]
  H --> M[Scale ranking model]
  M --> S[Low-cost serving]

Original paper figure¶

The original figure shows joint data/model scaling. From the original paper; copyright remains with the authors.

Core equation¶

\[ h^{l+1}=h^l+\sum_d g_d(h^l)\,FFN_d(h^l) \]

Reported results¶

Section 4.5 reports GMV +2.04% at 5% traffic for ten days; not comparable with MovieLens.

Local reproduction¶

Shared transition + content scorer vs independent UniScale core on matched MovieLens-100K 220 users/360 items, full-catalog candidates and seed 42. NDCG@10 0.0540→0.0528 (-2.20%); Hit@10 0.1091→0.1045. Validation selects only blend weight; test does not tune.

auto-research reproduce --paper uniscale --dataset-dir data --seed 42

Fixed metrics. Independent model state and scoring execute; no private features, production foundation model or online traffic. The negative result is retained.