# MARE: Multi-Agents Collaboration Framework for Requirements Engineering

**Authors:** Dongming Jin, Zhi Jin, Xiaohong Chen, Chunhui Wang  
**arXiv:** [2405.03256](https://arxiv.org/abs/2405.03256)  
**Date:** 2024-05-06  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2405.03256v1/

## Key points

- Requirements Engineering (RE) is crucial for translating stakeholders' needs into detailed requirements specifications
- Recent advancements in deep learning techniques show promise in enhancing various RE tasks
- Effective collaboration across different tasks and roles is necessary for high-quality requirements specifications
- MARE (Multi-Agents Collaboration Framework for Requirements Engineering) leverages large language models to facilitate collaboration in the RE process
- MARE divides the RE process into four key tasks: elicitation, modeling, verification, and specification, each assigned to specific agents capable of multiple actions
- MARE consists of five agents and nine distinct actions, providing a dedicated workspace for seamless collaboration
- Experiments show that MARE outperformed existing methods by 15.4% in generating correct requirements models
- Human evaluations confirmed the quality of requirements specifications generated by MARE across three aspects
- MARE represents a significant advancement in enhancing collaboration among agents involved in Requirements Engineering processes

## Summary

, , , , 

In the field of software development, Requirements Engineering (RE) plays a crucial role in translating stakeholders' needs into detailed requirements specifications. Recent advancements in deep learning techniques have shown promise in enhancing various RE tasks. However, effective collaboration across different tasks and roles involved in the process is necessary to achieve high-quality requirements specifications. To address this challenge, a novel framework called MARE (Multi-Agents Collaboration Framework for Requirements Engineering) has been proposed. MARE leverages the power of large language models (LLMs) to facilitate collaboration throughout the entire RE process. The framework divides the RE process into four key tasks: elicitation, modeling, verification, and specification. Each task is assigned to one or two specific agents, with each agent capable of performing multiple actions. In total, MARE consists of five agents and nine distinct actions. One of the key features of MARE is the provision of a dedicated workspace where agents can upload their generated intermediate requirements artifacts and access necessary information for seamless collaboration. To evaluate the effectiveness of MARE, experiments were conducted on a variety of public cases as well as new cases developed specifically for this study. A comparison with three baseline approaches using commonly used metrics revealed that MARE outperformed existing methods by 15.4% in generating correct requirements models. Furthermore, human evaluations were conducted to assess the quality of the requirements specifications generated by MARE across three aspects. The insights gained from these evaluations provide valuable feedback on the performance and capabilities of the framework in producing high-quality requirements specifications. Overall, MARE represents a significant advancement in enhancing collaboration among agents involved in Requirements Engineering processes. It demonstrates superior performance compared to state-of-the-art approaches and offers valuable insights into improving the quality of generated requirements models and specifications.