OneRanker: Unified Generation and Ranking with One Model in Industrial Advertising Recommendation¶
Core public-data ranking mechanism.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | Tencent. |
| First public release | 2026-03-03 (arXiv v1) |
| Author code | No official author code linked by the paper as checked on 2026-08-09. |
| Local adapter or method key | oneranker |
| Local reproduction code | src/auto_research/reproductions/oneranker/ |
Paper method and evidence¶
Use a fake item token to unify generation, click-value estimation and ad ranking. Distribution consistency aligns generated probability with the value head.
flowchart LR
A[Fake item token] --> B[Generate and estimate value]
B[Generate and estimate value] --> C[Align preferences for ranking]Original paper figure¶
Figure shows one model for generation and ranking. From the original paper; copyright remains with the authors.
Core equation¶
\[
\mathcal L=\mathcal L_{gen}+\lambda_v\mathcal L_{value}+\lambda_cD_{KL}(p_{gen}\Vert p_{rank}).
\]
Reported results¶
Full WeChat Channels ad rollout reports GMV +1.34%.
Local reproduction¶
MovieLens-100K full catalog, 260 users/420 items, seed 42: shared transition+content scorer versus adding OneRanker only, NDCG@10 0.0354→0.0327 (-7.65%). Fixed metrics. No private logs, traffic, checkpoints or production serving; local results do not reproduce paper lift.