一句话：提出业务逻辑驱动的迭代需求自动补全方法 ReqCompleter，用 CRUD 缺口触发 LLM 补全并抑制幻觉。
本文针对需求不完整以及大语言模型生成需求时易产生幻觉的问题，提出业务逻辑驱动的迭代式需求自动补全方法 ReqCompleter。该方法将“用例—实体—操作”三元组视为最小可计算闭环，把用例模型、E-R 图和 CRUD 矩阵融合为统一语义框架；以 CRUD 矩阵中的缺口作为触发器迭代检测缺失功能，同时利用 E-R 图限定实体边界，引导 LLM 在受控范围内生成需求。作者在电商、物流、公共安全等七个案例上评估，相比通用 LLM，需求完整率提升 20%–88%，幻觉率降低 2.4%–13.9%。该工作首次将经典需求工程模型与生成式 AI 紧耦合，形成自动闭环。
To tackle the dual challenges of incomplete requirements and hallucinations in large language models (LLMs), this paper proposes a business-logic-driven iterative requirements auto-completion approach named ReqCompleter. By treating the “use case – entity– operation” triplet as the smallest computable closed loop, ReqCompleter adopts a model-driven iterative mechanism. First, a use-case model, an E-R diagram, and a CRUD (Create, Read, Update, Delete) matrix are fused into a unified semantic framework. Next, gaps in the CRUD matrix act as triggers to iteratively detect missing functionalities, while the E-R diagram delimits entity boundaries to steer the LLM toward generating requirements within a controlled scope. We evaluate our approach across seven cases in e-commerce, logistics, public safety and other domains. Compared to general-purpose LLMs, it improves requirements completeness rate by 20%-88% while reducing hallucination rate by 2.4%-13.9%. To the best of our knowledge, this work represents the first tight coupling of classical requirements engineering models with generative AI, establishing an automated closed-loop system that delivers “what’s missing, as needed" under explicit business logic constraints. This opens a new and practical technical pathway for high-quality, explainable, and continuously evolvable requirements engineering.
@inproceedings{paperbot11902,
title = {Unlocking the Silent Needs: Business-Logic-Driven Iterative Requirements Auto-completion},
author = {Zhujun Wu and Xiaohong Chen and Zhi Jin and Ming Hu and Dongming Jin},
booktitle = {ICSE},
year = {2026},
doi = {10.1145/3744916.3773109}
}