Dual-Rerank: Fusing Sequential Dependencies and Utility for Generative Reranking¶
Core public-data reranking mechanism.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | Kuaishou Technology. |
| First public release | 2026-04-08 (arXiv v1) |
| Author code | No official author code linked by the paper as checked on 2026-08-09. |
| Local adapter or method key | dual-rerank |
| Local reproduction code | src/auto_research/reproductions/dual_rerank/ |
Paper method and evidence¶
An autoregressive teacher learns item-order dependencies; a non-autoregressive student predicts ranks in parallel. Utility distillation optimizes quality and latency.
flowchart LR
A[Autoregressive teacher learns dependencies] --> B[Distill utility]
B[Distill utility] --> C[Parallel rank student]Original paper figure¶
Figure shows teacher and student reranking. From the original paper; copyright remains with the authors.
Core equation¶
\[
\mathcal L=\mathcal L_{rank}+\lambda\,D_{KL}(p_{AR}\Vert p_{NAR})+\gamma\mathcal L_{utility}.
\]
Reported results¶
Kuaishou search on 5% traffic for a month: Long View +1.107% with lower mean and P99 latency.
Local reproduction¶
MovieLens-100K, 260 users/420 items, full catalog, seed 42: shared transition+content baseline versus adding Dual-Rerank, NDCG@10 0.0354→0.0392 (+10.66%). Fixed metrics. No private logs, traffic, checkpoints or production serving; local results do not reproduce paper lift.