Multi-Criteria Analysis of AI Code Generation Approaches

By Amine essadik

2026

Abstract

Software engineering has shifted from classical Model-Driven Engineering (MDE) toward Artificial Intelligence- driven software generation paradigms. While classical generative approaches relied on formal metamodels and transformation languages (such as QVT and ATL), contemporary software development increasingly relies on Large Language Models (LLMs), Autonomous Multi-Agent Systems, and Neuro-Symbolic Hybrid architectures. However, selecting an appropriate AI generation paradigm poses complex architectural dilemmas concerning hallucination risks, determinism, execution latency, context limitations, and domain compliance. This paper establishes a comprehensive multi-criteria decision framework to evaluate the three primary AI code generation paradigms: Direct Prompt-driven LLMs, Autonomous Multi-Agent Collaborative Systems, and Neuro-Symbolic Hybrid Architectures. Utilizing the Weighted Sum Method (WSM) across ten normalized engineering criteria, we systematically quantify their relative strengths and weaknesses. The analytical results demonstrate that while Multi-Agent frameworks achieve high domain reasoning and reverse-engineering capabilities, Neuro-Symbolic approaches yield superior overall scores (2.65/3.0) due to formal verification, deterministic boundaries, and rigorous security guarantees. This study provides actionable decision guidelines for researchers and software architects transitioning toward autonomous software engineering

Keywords

Ai , machine learning

References

1. [1] B. Selic, "The pragmatics of model-driven development, " IEEE Software, vol. 20, no. 5, pp. 19–25, 2003. 2. [2] J. Bézivin, "On the unification power of models, " Software & Systems Modeling, vol. 4, no. 2, pp. 171–188, 2005. 3. [3] K. Czarnecki and S. Helsen, "Classification of model transformation approaches, " in Proc. 2nd OOPSLA Workshop on Generative Techniques, 2003, pp. 1–17. 4. [4] M. Chen et al., "Evaluating large language models trained on code, " arXiv preprint arXiv:2107.03374, 2021. 5. [5] S. Bairi et al., "CodePlan: Repository-level coding using planned generation, " Proc. ACM Program. Lang. (OOPSLA), vol. 8, pp. 1–28, 2024. 6. [6] S. Hong et al., "MetaGPT: Meta programming for multi-agent collaborative framework, " in Proc. Int. Conf. Learn. Representations (ICLR), 2024. 7. [7] A. S. d'Avila Garcez et al., "Neurosymbolic AI: The 3rd Wave, " Artificial Intelligence Review, vol. 56, pp. 12387–12406, 2023. 8. [8] M. Zeleny, Multiple Criteria Decision Making, New York: McGraw-Hill, 1982. 9. [9] D. L. Olson, "Comparison of three multicriteria methods to predict known outcomes, " European Journal of Operational Research, vol. 130, no. 3, pp. 576–587, 2001. 10. [10] C. E. Jimenez et al., "SWE-bench: Can language models resolve real-world GitHub issues?, " in Proc. Int. Conf. Learn. Representations (ICLR), 2024. 11. [11] R. Dehbi, M. Talea, and A. Tragha, pp. 2326–2330, 2013. 12. [12] N. Tankovic, D. Vukotic, and M. Zagar, Interfaces (ITI), 2012, pp. 43–48. "The generation approach of Multi-target learning management system, " Advanced Science Letters, vol. 19, no. 8, "Rethinking model driven development: analysis and opportunities, " in Proc. 34th Int. Conf. Inf. Technol.

BibTeX EndNote

Share this publication

X Facebook LinkedIn WhatsApp

Full Paper Access

You are viewing a preview. Access the full manuscript via single purchase or subscription.

You need an account to purchase research papers. Create one for free.