CGR: Constraint-aware generative reranking¶
Core public-data constraint mechanism.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | Bilibili. |
| First public release | 2026-03-04 (arXiv v1) |
| Author code | No public author code found as of 2026-09-05. |
| Local adapter or method key | cgr |
| Local reproduction code | src/auto_research/reproductions/cgr/ |
Paper method and evidence¶
One network generates the sequence and estimates reward. Encode slot and ad-load rules in autoregressive decoding, and prune infeasible branches early rather than repairing output afterward.
flowchart LR
A[Generate and estimate reward] --> B[Check constraints during decode]
B[Check constraints during decode] --> C[Prune infeasible branches]Original paper figure¶
Figure shows constraint-aware reward pruning. From the original paper; copyright remains with the authors.
Core equation¶
\[
y_t=\arg\max_{i\in\mathcal F(y_{<t})}\left[s_\theta(i|y_{<t})+\lambda^\top r(i)\right].
\]
Reported results¶
The paper reports revenue and engagement gains with business constraints met and over 85% lower inference latency on millions of daily ad requests.
Local reproduction¶
MovieLens-1M fixed catalog: NDCG@10 0.05401→0.05715 (+5.82%). Three-seed metrics, mean, SD and 95% CI. No private features, production traffic or serving. These local public-data results do not reproduce online lift.