Title: Reimagining Sales Lead Scoring via LLM-Driven Hierarchical Preference Ranking

Abstract

The process of converting sales leads in high-stakes sectors, such as the automotive and real estate industries, is fundamentally distinct from e-commerce recommendation systems. This divergence stems from extended decision-making cycles and complex, multi-stage sales funnels. Conventional lead scoring techniques—ranging from rule-based scorecards and machine learning algorithms to pointwise Click-Through Rate (CTR) models—struggle with significant hurdles. These include sparse supervision signals, a semantic disconnect when processing unstructured Customer Relationship Management (CRM) logs, and a general inability to accurately capture the relative priority of leads.

Although Large Language Models (LLMs) provide superior semantic comprehension of customer interactions, standard, general-purpose LLMs are poorly equipped for lead ranking. They are designed to generate text rather than produce comparable scores and fail to align with the hierarchical priorities inherent in sales funnels. To address this, we present an LLM-based discriminative framework for sales lead scoring that facilitates the joint modeling of structured CRM data and unstructured customer interactions.

Building upon this foundation, we introduce HPRO (Hierarchical Preference Ranking Optimization). This approach enhances sales lead scoring by incorporating a hierarchical preference ranking objective. HPRO utilizes a margin-aware Bradley-Terry formulation to convert sparse binary labels into dense, funnel-aware preference pairs. This mechanism allows lead scoring to benefit from both pointwise and pairwise supervision.

Our experiments, conducted on large-scale data from a prominent New Energy Vehicle (NEV) brand, reveal state-of-the-art results in both classification (achieving an AUC of 0.8161) and ranking performance, which saw a 39.7% increase in precision among the top-ranked leads. Furthermore, a 132-day online A/B test confirmed a 9.5% uplift in sales volume, validating the model’s real-world commercial impact.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC