From Local Indices to Global Identifiers: Generative Reranking via Global Action Space¶
Core public-data generation mechanism.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | City University of Hong Kong / Kuaishou / UC San Diego. |
| First public release | 2026-04-28 (arXiv v1) |
| Author code | No official author code linked by the paper as checked on 2026-08-09. |
| Local adapter or method key | glorank |
| Local reproduction code | src/auto_research/reproductions/glorank/ |
Paper method and evidence¶
A global SID action space turns local reranking into full-catalog generation. Listwise SFT precedes group-relative optimization of slate utility and tail coverage.
flowchart LR
A[Map candidates to global SIDs] --> B[Listwise supervised training]
B[Listwise supervised training] --> C[Group-relative slate utility]Original paper figure¶
Figure shows global-action reranking. From the original paper; copyright remains with the authors.
Core equation¶
\[
J(\theta)=\mathbb E_{\pi_\theta(Y|h)}[R_{list}(Y)],\quad R=R_{rel}+\lambda R_{global}.
\]
Reported results¶
Kuaishou 7.8% traffic, 14 days: watch time +0.095%, effective views +0.111%, likes +0.286%.
Local reproduction¶
MovieLens-100K, 260 users/420 items, full catalog, seed 42: shared transition+content versus GloRank, NDCG@10 0.0354→0.0197 (-44.26%). Fixed metrics. No private logs, traffic, checkpoints or production serving; local results do not reproduce paper lift.