Generative Pseudo-Labeling for Pre-Ranking with LLMs¶
Public-data core mechanism, not private production serving.
Paper information¶
| Field | Details |
|---|---|
| Paper | arXiv 2602.20995 |
| Institution | Alibaba Group / Renmin University. |
| First public release | 2026-02-24 (arXiv v1) |
| Author code | No official author code linked by the paper as checked on 2026-08-24. |
| Local adapter or method key | gpl-prerank |
| Local reproduction code | src/auto_research/reproductions/gpl_prerank/ |
Paper method and evidence¶
An LLM derives interest anchors from user behavior and labels unexposed retrieval candidates. The pre-ranker learns from real and pseudo-supervision without calling the LLM online.
flowchart LR
H[User history] --> A[LLM interest anchors]
A --> P[Pseudo-label unexposed candidates]
P --> R[Train pre-ranker with real labels]
R --> S[Serve without LLM]Original paper figure¶
The original figure shows LLM supervision for unexposed candidates. From the original paper; copyright remains with the authors.
Core equation¶
Reported results¶
Section 4.5/Table 4 reports CTR +3.07% across two weeks with hundreds of millions of DAU. Not comparable to MovieLens.
Local reproduction¶
NDCG@10 0.0540→0.0609 (+12.84%), Hit@10 0.1091→0.1045. Same MovieLens-100K 220 users/360 items, full catalog, split and seed 42, using a shared transition + content baseline. Validation selects only blend weight; test does not tune. The local model has independent state and scoring. No private features, production checkpoint or online serving; retain negative results.