Abstract
Modern online recruitment platforms face significant challenges in matching the right candidates to specific job requirements due to the nuance gap in job descriptions, the behavioral heterogeneity of different recruiter roles, and the inherent noise in user interaction data. This paper presents a novel talent search framework that integrates Large Language Models (LLMs) for fine-grained job description encoding with a role-aware Multi-gate Mixture-of-Experts (MMoE) architecture. By leveraging Chain-of-Thought (CoT) prompting to extract implicit requirements and tailoring ranking logic to specific recruiter personas, the proposed system achieved a 17.29% lift in online conversion rates, resulting in substantial annual cost savings.
Introduction
In the competitive landscape of online recruitment, the efficiency of talent search engines is a primary driver of success. Traditional systems often rely on keyword matching or basic semantic similarity, which frequently fail to capture the subtle complexities of professional requirements. Furthermore, these systems typically treat all recruiters as a monolithic group, ignoring the fact that a Sourcing Assistant and a Team Leader have vastly different search behaviors and conversion patterns. Finally, behavioral data—such as clicks—is often noisy and does not always correlate with a high-quality match.
To address these hurdles, we introduce a framework that combines the reasoning capabilities of LLMs with the architectural flexibility of Mixture-of-Experts (MoE) models. Our approach focuses on three core innovations: deep semantic extraction from job descriptions, personalized ranking based on recruiter roles, and multi-task learning to de-noise behavioral signals.
Methodology
1. Fine-Grained JD Encoding with LLMs
Standard Job Descriptions (JDs) often contain "hidden" requirements. For example, a requirement for a "Backend Engineer" at a high-growth startup might implicitly demand strong computer science fundamentals and experience with scalable architecture, even if those terms aren't explicitly listed. We utilize an LLM with Chain-of-Thought (CoT) prompting to analyze JDs and historical hiring data. This process generates an "augmented summary" that captures these implicit preferences. These summaries are then transformed into high-dimensional vectors to serve as features for the ranking model, bringing LLM-level reasoning into a high-speed production environment without the latency of real-time LLM inference.
2. Role-Aware Personalized MMoE
Recruiter behavior is highly heterogeneous. Our analysis indicates that Sourcing Assistants (SAs) often perform broad searches with high page-view volume but lower conversion, whereas Team Leaders (TLs) are more selective. To account for this, we implemented a role-aware gating network within a Multi-gate Mixture-of-Experts (MMoE) framework. This network dynamically routes input features through different "expert" sub-networks based on the recruiter's specific role (SA, Sourcing Generalist, or TL), allowing the model to learn specialized ranking strategies for different user personas.
3. Multi-Task Learning and De-noising
To mitigate the impact of noisy behavioral data, the system is trained using a Multi-Task Learning (MTL) module. The model jointly optimizes for:
- Click-Through Rate (CTR): Predicting the likelihood of a recruiter clicking a profile.
- Conversion Rate (CVR): Predicting the likelihood of a recruiter initiating contact or moving a candidate to the next stage.
- Resume Matching Relevance: An auxiliary task that uses resume content as a ground-truth signal of candidate quality. This task helps the model distinguish between an "accidental" click and a genuine match, effectively de-noising the training signal.
Experimental Results
Dataset and Baselines
The model was trained and evaluated using a proprietary dataset from a major recruitment platform, comprising approximately 495,000 training samples and 257,000 test samples. The data spanned 1,000 recruiters and 200,000 talent profiles. We compared our framework against several industry-standard models, including GDCN, ESSM, MMoE, PLE, and STEM.
Performance Metrics
In offline evaluations, the proposed framework demonstrated significant improvements over the best-performing baselines, achieving relative AUC gains of 1.70% for CTR and 5.97% for CVR.
More importantly, a live A/B test revealed a 17.29% lift in Click-Through Conversion Rate (CTCVR). This improvement in ranking accuracy directly translated to business value by reducing the reliance on expensive external headhunting services, resulting in estimated annual savings of millions of CNY.
Discussion and Limitations
The success of this framework highlights two key insights: first, that LLMs are highly effective as offline feature extractors for complex text; and second, that modeling recruiter roles is essential for true personalization in B2B platforms.
However, the system has limitations. It currently struggles with cold-start scenarios, where new job postings or recruiters lack the historical interaction data necessary for role-aware modeling. Additionally, the model does not yet account for temporal dynamics, such as shifting market trends or evolving hiring needs over time.
Conclusion
By combining LLM-based semantic enrichment with role-aware expert mixtures, we have developed a talent search ranking system that significantly outperforms traditional models. Future research will focus on incorporating temporal modeling and developing strategies to improve performance for cold-start entities, further refining the precision of professional talent matching.
Summary updated