From Item-Only to Query-Item: Query-Conditioned Generative Search with QGS in Quark¶
Core public-data search mechanism.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | University of Science and Technology of China / Alibaba. |
| First public release | 2026-05-25 (arXiv v1) |
| Author code | No official author code linked by the paper as checked on 2026-08-24. |
| Local adapter or method key | qgs |
| Local reproduction code | src/auto_research/reproductions/qgs/ |
Paper method and evidence¶
Serialize query and item jointly for generative search. Linear HSTU encodes history at linear complexity and HFG-Attention fuses sparse crossed features.
flowchart LR
A[Serialize query and item] --> B[Linear HSTU history]
B[Linear HSTU history] --> C[Fuse sparse crosses]Original paper figure¶
Figure shows query-conditioned generation. From the original paper; copyright remains with the authors.
Core equation¶
\[
p(y|q,h)=\prod_t p(y_t|q,h,y_{<t})
\]
Reported results¶
Production CTR +0.62% over seven days on 2% traffic; not a local metric.
Local reproduction¶
Matched MovieLens-100K 220 users/360 items, full catalog and seed 42: shared transition+content baseline versus QGS. NDCG@10 0.0540→0.0544 (+0.73%), Hit@10 0.1091→0.0955. Only validation chooses blend. Fixed metrics. No private logs, traffic, checkpoints or production serving; local results do not reproduce paper lift.