Debt Collection Negotiations with Large Language Models (MADeN)

This repository accompanies the paper:

Debt Collection Negotiations with Large Language Models: An Evaluation System and Optimizing Decision Making with Multi-Agent
Xiaofeng Wang, Zhixin Zhang, Jinguang Zheng, Yiming Ai, Rui Wang
arXiv: 2502.18228

Overview

This project explores the automation of debt collection negotiations (DCN) using large language models (LLMs). Traditional DCN methods are labor-intensive and lack dynamic negotiation capabilities. Our work introduces:

• An Evaluation Framework: A novel system with 13 metrics across 4 aspects to assess DCN performance.
• MADeN Framework: A Multi-Agent Debt Negotiation system incorporating planning and judging modules to enhance decision-making rationality.
• Optimization Techniques: Application of post-training methods, including Direct Preference Optimization (DPO) with rejection sampling, to improve negotiation outcomes.

Our experiments reveal that while LLMs have potential in automating DCN, they tend to over-concede compared to human negotiators. The proposed MADeN framework addresses this by enhancing decision rationality and efficiency.

Features

• Dynamic Negotiation Simulation: Leveraging LLMs to simulate realistic debt collection scenarios.
• Multi-Agent Collaboration: Implementing planning and judging modules to mimic human negotiation strategies.
• Comprehensive Evaluation: Utilizing a 13-metric framework to assess negotiation effectiveness.
• Post-Training Optimization: Enhancing LLM performance through DPO and rejection sampling techniques.

Getting Started

Note: The codebase is under development. The following instructions will be updated upon release.

Citation

If you find this work useful, please cite our paper:

@article{wang2025debt,
  title={Debt Collection Negotiations with Large Language Models: An Evaluation System and Optimizing Decision Making with Multi-Agent},
  author={Wang, Xiaofeng and Zhang, Zhixin and Zheng, Jinguang and Ai, Yiming and Wang, Rui},
  journal={arXiv preprint arXiv:2502.18228},
  year={2025}
}