SimRPD: Optimizing Recruitment Proactive Dialogue Agents through Simulator-Based Data Evaluation and Selection

TL;DR

SimRPD tackles data scarcity in goal oriented recruitment dialogue by closing the loop between a high fidelity user simulator and a dual level evaluation framework grounded in Chain-of-Intention. The approach first generates large-scale synthetic data with SFT+RL, then filters it with global distribution and instance level metrics, and finally trains a recruitment proactive dialogue agent with supervised fine tuning and PPO optimization. Empirical results show distributional alignment gains, reduced hallucinations, and a real-world uplift of 15.8% in WeChat acquisition in production, validating the method’s industrial relevance. The framework emphasizes data quality over sheer volume and offers a transferable workflow for business-oriented dialogue scenarios, while acknowledging domain specificity and the need for reference data for global evaluation.

Abstract

Task-oriented proactive dialogue agents play a pivotal role in recruitment, particularly for steering conversations towards specific business outcomes, such as acquiring social-media contacts for private-channel conversion. Although supervised fine-tuning and reinforcement learning have proven effective for training such agents, their performance is heavily constrained by the scarcity of high-quality, goal-oriented domain-specific training data. To address this challenge, we propose SimRPD, a three-stage framework for training recruitment proactive dialogue agents. First, we develop a high-fidelity user simulator to synthesize large-scale conversational data through multi-turn online dialogue. Then we introduce a multi-dimensional evaluation framework based on Chain-of-Intention (CoI) to comprehensively assess the simulator and effectively select high-quality data, incorporating both global-level and instance-level metrics. Finally, we train the recruitment proactive dialogue agent on the selected dataset. Experiments in a real-world recruitment scenario demonstrate that SimRPD outperforms existing simulator-based data selection strategies, highlighting its practical value for industrial deployment and its potential applicability to other business-oriented dialogue scenarios.

Figures (3)

• Figure 1: Background of this work. High-quality data in training proactive dialogue agents in real-world applications is sparse, therefore we train a user simulator to synthesize data and propose a dual-level evaluation protocol to select premium data.
• Figure 2: Overview of the SimRPD framework, illustrating the pipeline from simulator training to multi-granularity data selection and final proactive agent optimization.
• Figure 3: Real-world intent transition heatmap.