OneMall: Unified end-to-end generative recommendation across scenarios¶

Core scenario prompts, three-level SIDs and cross-behavior fusion on a reduced public dataset.

Paper information¶

Field	Details
Paper	arXiv 2601.21770
Institution	Kuaishou.
First public release	2026-01-29 (arXiv v1)
Author code	No official author code linked by the paper as checked on 2026-08-09.
Local adapter or method key	`onemall`
Local reproduction code	`src/auto_research/reproductions/onemall/`

Paper method and evidence¶

Unify previously separate product-card, short-video and livestream channels using common product SIDs, Query-Former compression, cross-attention over behaviors, a sparse MoE generator and RL alignment.

flowchart LR
  H[Multiple behaviors and scenario prompt] --> Q[Query-Former compression]
  Q --> A[Cross-behavior attention]
  A --> S[Generate hierarchical item IDs]
  S --> R[Ranking reward]

Original paper figure¶

The original figure shows the multi-scenario generative system. From the original paper; copyright remains with the authors.

Core equation¶

\[ \mathcal L_{\mathrm{gen}}=-\sum_{\ell=1}^{L}\log p(c_\ell\mid c_{<\ell},h,s), \qquad \mathcal J=\mathbb E_{\pi_\theta}[R_{\mathrm{rank}}]. \]

Reported results¶

Paper online result: product-card GMV +13.01%, short-video orders +15.32%, livestream orders +2.78%. Not local.

Local reproduction¶

Shared transition/content ensemble versus adding scenario prompts, three residual SID levels and behavior fusion on MovieLens-100K: NDCG@10 0.03540→0.03693 (+4.33%), Hit@10 -4.35%. Validation selects blend weights, not test. Metrics.

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

Genre only proxies scenario. No private product features, sparse MoE, production RL or online serving.