简体中文 · English

OneModel: Platform-wide ranking across multiple surfaces

Fidelity: core-mechanism reproduction. Runs multi-surface statistics, surface projections, SAIM-style gating, and global/local hierarchical representations. The production Transformer and private features are not reproduced.

Paper information

Field	Details
Paper	arXiv 2608.18606
Institution	Xiaohongshu
First public release	2026-08-19 (arXiv v1)
Author code	No author-released code found (checked 2026-08-24).
Local adapter or method key	onemodel
Local reproduction code	src/auto_research/reproductions/onemodel/

Original paper summary

Background and main changes

Recommendation, ads, and merchant recommendation typically train separate rankers. This duplicates resources and limits sharing of cross-surface behavior. OneModel projects heterogeneous features to unified event tokens and models mixed long histories with a shared causal decoder. SAIM gates condition the FFN on the surface; the hierarchical representation combines global attentive pooling with recent local state, balancing transfer and specialization.

flowchart LR
  H["Mixed recommendation, ads, and merchant histories"] --> P["Surface-specific feature projection"]
  P --> C["Surface + action + time + position context"]
  C --> D["Shared causal decoder"]
  D --> S["SAIM surface modulation"]
  S --> R["Global pooling + final local state"]
  R --> T["Task heads for three surfaces"]

Key figure from the paper

Figure 1: Overall design and components. From the paper; copyright belongs to the authors. Click for the source.

Core equations

\[ e_i=\phi(W^{(s)}x_i^{(s)}+b^{(s)}),\qquad z_j=e_{i_j}+\operatorname{Embed}(s_j,a_j,\Delta t_j,pos_j). \]

SAIM generates a gate from the surface embedding and modulates the shared FFN:

\[ g_t=\sigma(W_g e_s+b_g),\qquad \operatorname{SAIM}(h_t)=\operatorname{SiLU}(W_1h_t)\odot g_t. \]

Offline and online results

Online Xiaohongshu results: Explore Feed time +0.33% and engagement +1.25%; Feed Ads value +3.43% and CTR +8.18%; Merchant Recommendation DGMV +1.1867% and GPM +2.1585%. This is the evidence for inclusion in the repository's industrial queue.

Local reproduction

Local comparison: A shared global transition ranker is the baseline. The experimental arm adds three surfaces derived from MovieLens genres, surface transitions, sigmoid gating, and global/local fusion. Its NDCG@10 is 0.0870 versus 0.1285, a 32.29% decrease, not an improvement.

The negative result shows that artificially dividing the single MovieLens domain into surfaces loses data density. It cannot be extrapolated to the paper's cross-business gains. See metrics/movielens-1m-seed42.json.

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

Scope

The local run uses public MovieLens-1M, not private Xiaohongshu multistream logs, action tokens, a unified Transformer decoder, early gradient isolation, selective backpropagation, or online serving. The core surface gating path executes, but metrics verify only the mechanism and its failure boundary.