GR4AD: generative recommendation for large-scale ads¶
Fidelity: core mechanism. Trains UA-SID, parallel LazyAR, VSL, and list-wise RSPO. Private multimodal ads, eCPM, online continual learning, and dynamic beam serving are not reproduced.
Paper information¶
| Field | Details |
|---|---|
| Paper | arXiv 2602.22732 |
| Organization | Kuaishou |
| First public release | 2026-02-26 (arXiv v1) |
| Authors' code | The paper did not provide official or author code when checked on 2026-08-09. |
| Adapter / method | gr4ad |
| Local reproduction code | src/auto_research/reproductions/gr4ad/ |
Original paper¶
Motivation and method¶
Generative ad recommendation must meet business-semantics, real-time update, and latency requirements. GR4AD uses multigranularity UA-SID to encode ads, LazyAR to relax dependencies between autoregressive levels, VSL for business-value-weighted supervision, and RSPO to optimize ranked lists directly. Dynamic Beam Serving adjusts beam width with load.
flowchart LR
A["multimodal ad + business"] --> B["UA-SID"]
C["user history"] --> D["LazyAR parallel decoder"]
B --> D --> E["VSL"] --> F["RSPO online update"] --> G["dynamic beam serving"]Key figure from the paper¶
Figure 1: Core architecture, modules, and their connections. From the original paper; copyright remains with the authors.
Core equations¶
\(L_{VSL}=-w(v_i)\sum_l\log p(c_{i,l}|h)\). Local RSPO uses the reward distribution \(q_j\propto e^{r_j/\tau}\) to optimize \(L_{RSPO}=-\sum_jq_j\log p_\theta(j|h)\).
Paper results¶
LazyAR nearly doubles QPS; off-peak beam can grow 60%. Ad revenue rises up to +4.2% over DLRM, with deployment to over 400 million users.
Local reproduction¶
Local comparison: Against common DIN, GR4AD NDCG@10 is +69.67%. Adding UA-SID/LazyAR/VSL/RSPO to the internal DLRM baseline yields +49.88%. Both carry concentration risk, with head share 0.505.
60 steps, 180 users, 280 items, seeds 42–44. Common DIN (100 steps) Hit/NDCG is 0.0481/0.02167; GR4AD is 0.0704/0.03676, NDCG +69.67% relative. The internal DLRM ablation gives +49.88%. Head share reaches 0.505, showing concentration from the MovieLens popularity value proxy. See the metrics.