LEADRE: LLM-based generative ad retrieval with multifaceted knowledge¶
Fidelity: full core path. Runs semantic IDs, long/short-term and commercial intent, semantic and value auxiliary tasks, SFT followed by DPO, and cached and real-time paths. Private ads and production serving are not reproduced.
Paper information¶
| Field | Details |
|---|---|
| Paper | arXiv 2411.13789 |
| Organization | Tencent / WeChat |
| First public release | 2024-11-21 (arXiv v1) |
| Authors' code | The paper did not provide official or author code when checked on 2026-08-09. |
| Adapter / method | leadre |
| Local reproduction code | src/auto_research/reproductions/leadre/ |
Original paper¶
Motivation and method¶
ID retrieval struggles to use ad content. LEADRE builds intent-aware prompts from cross-domain behavior, long-term summaries, and real-time behavior. It quantizes ads to S-IDs, trains ad semantics and commercial-value auxiliary tasks, then aligns with DPO. Online serving combines a long-term generation cache with real-time serving.
flowchart LR
A["long/short/cross-domain behavior"] --> B["intent-aware LLM"]
C["ad content"] --> D["RQ-VAE S-ID"]
B --> E["S-ID generation + auxiliary tasks"]
D --> E --> F["DPO alignment"] --> G["cached + real-time serving"]Key figure from the paper¶
Figure 2: Core architecture, modules, and connections. From the original paper; copyright remains with the authors.
Core equations¶
SFT uses \(L=L_{SID}+\lambda_sL_{semantic}+\lambda_vL_{value}\). DPO uses \(-\log\sigma(\beta[(\log\pi_\theta(y_w)-\log\pi_\theta(y_l))-(\log\pi_{ref}(y_w)-\log\pi_{ref}(y_l))])\).
Paper results¶
HR@1 is 0.0764 and NDCG@8 0.1360. WeChat Channels and Moments GMV increase +1.57%/+1.17%, respectively, and deployment serves tens of billions of requests daily.
Local reproduction¶
Local comparison: LEADRE is +12.94% relative to the shared DIN baseline on NDCG@10. Against the internal S-ID multitask SFT baseline, adding DPO changes NDCG by -4.53%. Report system and alignment ablations separately.
MovieLens-100K, 180 users, 280 items, seeds 42–44: DIN at 100 steps has Hit/NDCG 0.0481/0.02167, and LEADRE 0.0574/0.02447, or +12.94% relative NDCG. Internally, DPO changes SFT's 0.02563 to 0.02447 (-4.53%). The local gain comes from the full S-ID/SFT system, not DPO. See the metrics.