COBRA: Cascaded sparse–dense generative retrieval¶
Local sparse S-ID, conditional dense vector, joint training, and coarse-to-fine BeamFusion execute; no production ANN service.
Paper information¶
| Field | Details |
|---|---|
| Paper | arXiv 2503.02453 |
| Institution | Baidu. |
| First public release | 2025-03-04 (arXiv v1) |
| Author code | No official author code linked by the paper as checked on 2026-08-09. |
| Local adapter or method key | cobra |
| Local reproduction code | src/auto_research/reproductions/cobra/ |
Paper method and evidence¶
Quantizing to semantic IDs loses fine detail. Generate a sparse code, then condition an end-to-end trainable dense vector on it. Use beam search to find codes and refine retrieval by dense neighbors.
flowchart LR
H[User history] --> S[Generate sparse code]
S --> D[Code-conditioned dense vector]
S --> B[Beam search]
B --> R[Dense refinement]
D --> ROriginal paper figure¶
The original figure shows sparse-to-dense cascade retrieval. From the original paper; copyright remains with the authors.
Core equation¶
\(L=L_{sparse}+\lambda L_{dense}\), with \(L_{sparse}=-\sum_l\log p(c_l|h,c_{<l})\). Dense loss is full-catalog softmax between the conditional vector and target item. At inference \(s(i)=s_{beam}(c_i)+\alpha q(c_i,h)^Tv_i\).
Reported results¶
The paper beats purely generative and dense retrieval on public benchmarks. Baidu reports online conversion +3.60% and ARPU +4.15%, covering over 200 million daily active users.
Local reproduction¶
180 users, 280 items, seeds 42–44. Shared DIN baseline (100 steps) Hit/NDCG 0.0481/0.02167, COBRA 0.0593/0.02724, or +25.75% relative NDCG. The distinct sparse-only → dense-cascade ablation is +45.83%; these comparisons are not interchangeable. Head share rises 0.0598→0.2313, indicating higher popularity concentration. Metrics.