Policy-Constrained Dynamic Facet Recommendation for Job Search¶
Core public-data facet mechanism.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | LinkedIn. |
| First public release | 2026-05-15 (arXiv v1) |
| Author code | No public author code found as of 2026-09-02. |
| Local adapter or method key | policy-facet |
| Local reproduction code | src/auto_research/reproductions/policy_facet/ |
Paper method and evidence¶
Humans and LLMs maintain a policy-compliant facet taxonomy. An embedding SLM retrieves facets, and a generative SLM scores them pointwise with one token. Prefix caching and batching serve the latency target.
flowchart LR
A[Policy-compliant taxonomy] --> B[Retrieve facet candidates]
B[Retrieve facet candidates] --> C[One-token generative score]Original paper figure¶
Figure shows taxonomy, retrieval and scoring. From the original paper; copyright remains with the authors.
Core equation¶
\[
\mathcal C_K=\operatorname{TopK}_{c\in\mathcal T}\cos(e_{u,q},e_c),\qquad s_c=\log p_\theta(\mathrm{Okay}\mid u,q,c).
\]
Reported results¶
LinkedIn Job Search A/B: facet CTR +34.8%, applications per impression +2.6%, successful search sessions +1.6%.
Local reproduction¶
MovieLens-1M full catalog: NDCG@10 0.05401→0.04713 (-12.73%). Three-seed metrics, mean, SD, and 95% CI. No private logs, production models, bucketing or serving. Public proxies cannot establish online lift.