cfgLLM: Automated assistance for software requirements using multi-LLM Fusion Agentic Framework
Open Access
- 19 February 2026
- conference paper
- Published by Association for Computing Machinery (ACM)
- p. 1-10
Abstract
A software requirements analysis document comprises various functional and non-functional requirements of stakeholders and is often expressed in structured natural language. If these requirements are ambiguous or deviate from the end goals, it leads to significant delays and inefficiencies during the software development process. Traditionally, control flow graphs (CFGs) are used to represent the logical flow and complexity of a software system. However, they are generated from the source code and require human effort and experience. In this paper, we propose a novel multi-LLM fusion agentic framework for automated CFG generation directly from natural language software requirements. This serves as a supporting tool, enabling early identification of ambiguities, inconsistencies, and structural issues before implementation. The framework leverages multiple large language models (ChatGPT, LLaMA 3, and Gemini) to extract and interpret control logic. It employs an additional ChatGPT agent to fuse outputs along with contextual information from a vector database via a retrieved augmented generation (RAG) mechanism. To ensure semantic validation, we proposed evaluation metrics: (i) function accuracy, (ii) event accuracy, and (iii) event consistency. Cyclomatic complexity is also used to assess a program’s structural soundness. Experimental results demonstrate that cfgLLM generated valid CFGs, offering assistance for early-stage software design and requirement validation.Keywords
This publication has 12 references indexed in Scilit:
- An automated approach for requirement classification using machine learningInnovations in Systems and Software Engineering, 2025
- Advancing Retrieval-Augmented Generation with Inverted Question Matching for Enhanced QA PerformanceIEEE Access, 2024
- Model Generation with LLMs: From Requirements to UML Sequence Diagrams2024
- Advancing Requirements Engineering Through Generative AI: Assessing the Role of LLMs2024
- Advances in automated support for requirements engineering: a systematic literature reviewRequirements Engineering, 2024
- Large Language Models are Few-shot Testers: Exploring LLM-based General Bug Reproduction2023
- Improving BERT model for requirements classification by bidirectional LSTM-CNN deep modelComputers and Electrical Engineering, 2023
- Decision tree-based Design Defects DetectionIEEE Access, 2021
- Software engineering whispers: The effect of textual vs. graphical software design descriptions on software design communicationEmpirical Software Engineering, 2020
- Automated Extraction of Requirement Entities by Leveraging LSTM-CRF and Transfer Learning2020