Enhancing Talent Search Ranking with Role-Aware Expert Mixtures and LLM-based Fine-Grained Job Descriptions

Sep 18, 2026

Talent search is a cornerstone of modern recruitment systems, yet existing approaches often miss nuanced job-specific preferences, model recruiter behavior too coarsely, and suffer noise from subjective human judgments. A novel framework improves talent search effectiveness by using LLMs to extract fine-grained recruitment signals from job descriptions and historical hiring data, and by applying a role-aware multi-gate MoE network to model differences across recruiter roles. Multi-task learning jointly optimizes CTR, CVR, and resume matching relevance, achieving relative AUC gains of 1.70% (CTR) and 5.97% (CVR), plus a 17.29% lift in click-through conversion rate.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages10

Reading Time25 min

AccessFree to download

Introduction

Recruitment process and behavioral signals

Related work and remaining challenges

Proposed framework and contributions

Enhancing Talent Search Ranking with Role-Aware Expert Mixtures and LLM-based Fine-Grained Job Descriptions Jihang Li1, * , Bing Xu2,∗ , Zulong Chen2,∗ , Chuanfei Xu5,†, Minping Chen1,†, Suyu Liu2 , Ying Zhou4 , Zeyi Wen1,3 , 1HKUST (GZ), Guangzhou, China 2Alibaba Group, Hangzhou, China 3HKUST, Hong Kong, China 4 Zhijiang Lab, Hangzhou, China 5 Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ), Shenzhen, China Correspondence: [xuchuanfei@gml.ac.cn](xuchuanfei@gml.ac.cn), [mchen779@connect.hkust-gz.edu.cn](mchen779@connect.hkust-gz.edu.cn) Abstract Talent search is a cornerstone of modern recruitment systems, yet existing approaches often struggle to capture nuanced job-specific preferences, model recruiter behavior at a finegrained level, and mitigate noise from subjective human judgments. We present a novel framework that enhances talent search effectiveness and delivers substantial business value through two key innovations: (i) leveraging LLMs to extract fine-grained recruitment signals from job descriptions and historical hiring data, and (ii) employing a role-aware multi-gate MoE network to capture behavioral differences across recruiter roles. To further reduce noise, we introduce a multi-task learning module that jointly optimizes click-through rate (CTR), conversion rate (CVR), and resume matching relevance. Experiments on real-world recruitment data and online A/B testing show relative AUC gains of 1.70%(CTR) and 5.97%(CVR), anda 17.29% lift in click-through conversion rate. These improvements reduce dependence on external sourcing channels, enabling an estimated annual cost saving of millions of CNY. 1 Introduction The rise of online recruitment platforms has revolutionized how employers and jobseekers connect, enabling efficient matching between talents and open positions (Kenthapadi et al., 2017 ; Geyik et al., 2018b) . A core component of these systems is talent search, which allows recruiters to identify qualified talents for specific job postings. As shown in Figure 1, talent search plays a central role in the recruitment process, where recruiters interact with talent profiles by issuing queries and browsing retrieved results. This stage involves three key steps: (i) exposure, where talent profiles are surfaced to the recruiter; (ii) click, where the * Equal Contribution † Corresponding Authors Figure 1: Recruitment process of our online system. recruiter views a specific profile; and (iii) application initiation, where the recruiter initiates a hiring evaluation for the candidate. These interactions generate behavioral signals such as click-through rate (CTR) and conversion rate (CVR), which are central to assessing and optimizing search effectiveness. A detailed workflow of our deployed talent search system is provided in Appendix A. Recent advancements in talent search have improved ranking quality through better text modeling and personalization. Those techniques include extracting skill signals from recruiter queries and resumes (Manad et al., 2018), using deep models for learning-to-rank (Ramanath et al., 2018), and applying BERT-based keyword extraction (Devlinet al., 2019 ; Wang et al., 2021) . Incorporating historical recruiter behavior and talent features has further enhanced performance (Geyik et al., 2018a ; Ozcaglar et al., 2019 ; Yang et al., 2021) . Despite these advancements, several challenges remain. (i) Insufficient modeling of recruitment preferences: Existing methods often fail to accurately capture the nuanced requirements of individual job postings, limiting the relevance of retrieved talents. (ii) Lack of role-aware personalization: Existing methods typically treat all recruiters uniformly, ignoring role-specific behavioral patterns. This omission limits personalization and leads tomismatches between recruiter intent and retrieved talents. As shown in Table 1, sourcing assistants (SA) generate nearly 50% of page views, yet their 1

FAQs

What problems does the paper address in talent search ranking?

It addresses insufficient modeling of job-specific recruitment preferences, lack of role-aware personalization across recruiter types, and noisy behavioral signals caused by subjective judgments and expertise gaps.

How does the framework use LLMs to improve job-talent matching?

It uses LLM-based extraction to derive fine-grained recruitment preferences from job descriptions and historical hiring records, capturing implicit signals beyond surface-level keywords.

What is the role-aware multi-gate MoE component used for?

It models behavioral heterogeneity across different recruiter roles by using a role-aware multi-gate mixture-of-experts network, enabling more personalized talent ranking.

Sep 18, 2026

Uploaded by wps_ap_test_251126_0180