Scholarlink

Manually specifying and prioritizing user stories in software projects is time-consuming and prone to inconsistency. This paper investigates whether Large Language Models (LLMs) can support these activities through a role-based multi-agent system. The proposed system uses four models: GPT-3.5 Turbo, GPT-4o, LLaMA 3.3, and Mistral-Nemo, to generate user stories from a project description and prioritize them using prompts simulating stakeholder roles. Relevance is evaluated using semantic similarity to the project description, and prioritization consistency is assessed using Kendall’s Tau distance against expert rankings. Results indicate that all models generate functionally relevant requirements with high semantic similarity, although clarity and conciseness vary. In prioritization, the models show moderate alignment with expert rankings, particularly for mid- and low-priority items, while variability across runs remains a challenge.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Multi-agent LLM System for Automated Requirements Analysis: A Study on User Story Generation and Prioritization

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Malik Abdul Sami,
  • Zheying Zhang,
  • Muhammad Waseem,
  • Kai-Kristian Kemell,
  • Zeeshan Rasheed,
  • Tomas Herda,
  • Md. Toufique Hasan,
  • Jussi Rasku,
  • Pekka Abrahamsson

摘要

Manually specifying and prioritizing user stories in software projects is time-consuming and prone to inconsistency. This paper investigates whether Large Language Models (LLMs) can support these activities through a role-based multi-agent system. The proposed system uses four models: GPT-3.5 Turbo, GPT-4o, LLaMA 3.3, and Mistral-Nemo, to generate user stories from a project description and prioritize them using prompts simulating stakeholder roles. Relevance is evaluated using semantic similarity to the project description, and prioritization consistency is assessed using Kendall’s Tau distance against expert rankings. Results indicate that all models generate functionally relevant requirements with high semantic similarity, although clarity and conciseness vary. In prioritization, the models show moderate alignment with expert rankings, particularly for mid- and low-priority items, while variability across runs remains a challenge.