RADAR: Recall Augmentation through Deferred Asynchronous Retrieval¶
Core mechanism on public data, not production serving.
Paper information¶
| Field | Details |
|---|---|
| Paper | arXiv 2506.07261 |
| Institution | Meta. |
| First public release | 2025-06-08 (arXiv v1) |
| Author code | No official author code linked by the paper as checked on 2026-08-09. |
| Local adapter or method key | radar |
| Local reproduction code | src/auto_research/reproductions/radar/ |
Paper method and evidence¶
Run a full ranker asynchronously after the current request. Cache its high-value items to supplement real-time retrieval on the next request.
flowchart LR
R[Current request] --> A[Run ranker asynchronously]
A --> C[Cache candidates]
C --> U[Merge on next request]
N[Next request real-time retrieval] --> UOriginal paper figure¶
The original figure shows deferred candidate retrieval. From the original paper; copyright remains with the authors.
Core equation¶
\[
C_{t+1}=R_{live}(h_{t+1})\cup\operatorname{TopK}(f_{rank}(h_t,\mathcal I)).
\]
Reported results¶
The paper reports approximately doubled online Recall@200 and engagement +0.8%.
Local reproduction¶
MovieLens-100K, 260 users/420 items, seed 42, full-catalog ranking. Shared transition + content versus adding RADAR only: NDCG@10 0.0354→0.0344 (-2.78%). No private features, traffic, serving or kernels. The local decline is distinct from the online claim.