EGR: Shared-LM Native Vector Retrieval¶
Core mechanism evidence. Paper and local evaluations are separate.
Paper information¶
| Field | Value |
|---|---|
| Paper | arXiv 2607.23038 |
| Affiliations | Snap Inc. |
| First public release | 2026-07-25 (arXiv v1) |
| Authors' code | No official author code linked as of 2026-09-02. |
| Adapter / method | egr |
| Local code | src/auto_research/reproductions/egr/ |
Original paper¶
Background and contribution¶
One encoder produces indexed item vectors and user query vectors. Train item representation learning and next-item prediction jointly throughout.
flowchart LR
I[Items and histories] --> E[Shared encoder] --> V[Native item and query vectors] --> R[Retrieval]Key figure from the paper¶
Original Figure 2: EGR. From the paper; copyright belongs to the authors. Not a local result.
Core mechanism¶
\[
\mathcal L=\mathcal L_{IRL}+\lambda_t\mathcal L_{NIP},\quad s(u,i)=q_u^\top z_i.
\]
Paper results¶
Snap DPA two-week 10%/10% traffic experiment: impressions +0.15%, CTR +0.23%, CVR +2.91%, then full deployment. Not local effects.
Local reproduction¶
Matched full-catalog MovieLens-1M subset: NDCG@10 0.05401→0.05025 (−6.95%). Three-seed results, SD, and 95% CI.
Reproduction limits¶
Negative public proxy result. No private logs, online bucketing, production scale, or serving stack.