De-conflating Preference and Qualification: Constrained Dual-Perspective Reasoning for Job Recommendation with Large Language Models

TL;DR

This paper tackles the dual-perspective nature of job matching, where candidate preference and employer qualification must be reconciled. It introduces JobRec, a framework that uses the Unified Semantic Alignment Schema (USAS) to symmetrically encode candidate and job attributes across four layers, and a Two-Stage Cooperative Training with a Lagrangian-based policy alignment to decouple and optimize preference and qualification under explicit eligibility constraints. A synthetic, expert-refined dataset is constructed to provide reliable dual-perspective supervision, enabling robust evaluation. Experimental results on a CS-domain benchmark show JobRec outperforms both LLM-based baselines and general deep recommenders, with strong gains in qualification ranking and controllable trade-offs via the eligibility threshold. The work demonstrates that explicit de-conflation and policy-aware ranking can significantly improve both the relevance and feasibility of job recommendations, offering practical benefits for policy control and fairer, more effective recruitment systems.

Abstract

Professional job recommendation involves a complex bipartite matching process that must reconcile a candidate's subjective preference with an employer's objective qualification. While Large Language Models (LLMs) are well-suited for modeling the rich semantics of resumes and job descriptions, existing paradigms often collapse these two decision dimensions into a single interaction signal, yielding confounded supervision under recruitment-funnel censoring and limiting policy controllability. To address these challenges, We propose JobRec, a generative job recommendation framework for de-conflating preference and qualification via constrained dual-perspective reasoning. JobRec introduces a Unified Semantic Alignment Schema that aligns candidate and job attributes into structured semantic layers, and a Two-Stage Cooperative Training Strategy that learns decoupled experts to separately infer preference and qualification. Building on these experts, a Lagrangian-based Policy Alignment module optimizes recommendations under explicit eligibility requirements, enabling controllable trade-offs. To mitigate data scarcity, we construct a synthetic dataset refined by experts. Experiments show that JobRec consistently outperforms strong baselines and provides improved controllability for strategy-aware professional matching.

Figures (6)

• Figure 1: Job recommendation is inherently dual-perspective: candidates focus on preference (interest or willingness to apply) while employers enforce qualification (eligibility to be shortlisted). Effective matching should prioritize (B) "high-interest-high-eligibility" region and particularly avoid both (A) “high interest but low eligibility” and (D) “high eligibility but low interest” recommendations.
• Figure 2: The architecture of our framework, consisting of (A) Unified Semantic Alignment Schema (USAS) for symmetric candidate--job feature alignment, (B) Problem Decomposition: Preference versus Qualification to obtain disentangled scores $s_{\text{pref}}$ and $s_{\text{qual}}$, (C) Lagrangian-based Joint Optimization to derive a constraint-aligned ranking policy, (D) Expert-in-the-Loop Data Synthesis to construct dual-perspective supervision, and (E) Two-Stage Cooperative Training to learn disentangled experts and perform policy alignment.
• Figure 3: Controllability analysis by varying the eligibility target $\epsilon$ in our constraint-aligned policy. Larger $\epsilon$ enforces a stricter qualification requirement and yields improved qualification-aware ranking while maintaining stable preference performance.
• Figure 4: Top-$K$ agreement on preference ranking. Using softmax-normalized $s_{\text{pref}}$, we measure per-user ranking consistency between Pref-only and Ours via Top-$K$ Jaccard overlap and bidirectional Top-1 containment (whether one method's Top-1 appears in the other's Top-$K$).
• Figure 5: Job dataset overview. Summary statistics of the synthesized CS-domain job postings. The figure reports the distribution of job industries (top-left), the distribution of minimum GPA requirements (top-right), the top required skills across postings (bottom-left), and the minimum academic level requirements (bottom-right).