Generate to Explore, Select to Exploit: Aligning LLM-based Headline Generation with Personalized Recommendation¶
Public-data core mechanism. No private logs, production checkpoint, or serving stack.
Paper information¶
| Field | Value |
|---|---|
| Paper | arXiv v1 |
| Organization | Baidu |
| First public release | 2026-09-14 (arXiv v1) |
| Authors' code | No public author repository found as of 2026-09-16. |
| Adapter | gese |
| Local code | gese |
Original paper¶
Background and contribution¶
Explore sets of diverse, faithful headlines, then select a personalized headline using current context. The paper's production A/B reports +2.57% CTR and +0.87% dwell time, not local results.
flowchart LR
H[Public histories] --> M[GESE core] --> R[Common candidate scoring] --> E[Same-split offline evaluation]Key figure from the paper¶
Original architecture from the paper; copyright belongs to the authors.
Scoring and evidence¶
The local deterministic candidate scorer combines relevance, set diversity, faithfulness, and contextual selection. The source also describes LazFormer separately: transferable pretrained mapping and a ranking residual adapter, with mixing selected only on validation. Offline matched-public-set results are not production A/B.
Local reproduction¶
MovieLens-100K results use seeds 42/43/44. On the first seed and the same candidate set and split, NDCG@10 is 0.05401 baseline vs. 0.04421 experiment, −18.15% relative.
Reproduction limits¶
Public-data mechanism only, not private training, a large headline generator, production feedback, or serving. No registry-only Evolve mapping.