UniFormer: unified model-centric scaling for industrial recommendation¶

简体中文

Fidelity: core mechanism. Runs the defining mechanism verifiable on public data. Private data, production models, and serving remain outside scope.

Paper information¶

Field	Details
Paper	arXiv 2606.27058
Organization	Kuaishou
First public release	2026-06-25 (arXiv v1)
Authors' code	The paper did not provide official or author code when checked on 2026-09-02.
Adapter / method	`uniformer`
Local reproduction code	`src/auto_research/reproductions/uniformer/`

Original paper¶

Motivation and method¶

Separate item-independent and item-dependent features and tasks into semantic tokens. A unified interaction module scales sequences, features, and tasks together.

flowchart LR
  A["Public behavior and candidates"] --> B["UniFormer core mechanism"]
  B --> C["Matched-budget method output"]
  A --> D["Paper-defined direct baseline"]
  C --> E["Common metrics and structural diagnostics"]
  D --> E

Key figure from the paper¶

Figure 2: Core architecture, modules, and connections. From the original paper; copyright remains with the authors.

Core equations¶

\[ H'=\\operatorname{Attn}([T_{ind};T_{dep};T_{task}]),\\quad \\hat y_k=g_k(H'_{task,k}). \]

Paper results¶

Over seven days at 5% traffic on the Kuaishou main and Lite apps, Lite dwell time rose +0.260% and watch time +1.113%. User–item decoupling raised QPS +48%. This is paper evidence, not a local public-data outcome.

Local reproduction¶

Local comparison: Under the same full-catalog protocol, baseline NDCG@10 is 0.05401 and UniFormer NDCG@10 is 0.04131, -23.50% relative, a negative proxy result. The local proxy does not extrapolate to online lift.

See the three-seed artifact for complete results, means, standard deviations, and 95% confidence intervals.

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

Reproduction boundaries¶

The local run uses a public MovieLens-1M subset and auditable proxies to validate the core computation. It does not reproduce private logs, production models, online traffic assignment, or serving. No claim is made to reproduce absolute paper metrics or online gains.