IEEE/ACM 13th International Conference on Mobile Software Engineering and Systems (MOBILESoft)

A Methodology for Generating Mobile Applications Through Large Language Models

William Niemiec, Anderson Rocha Tavares, Érika Fernandes Cota

Published in

IEEE/ACM 13th International Conference on Mobile Software Engineering and Systems (MOBILESoft)

Year

2026

DOI

10.1145/3795077.3795116

Abstract

The creation of applications using Large Language Models (LLMs) is a fast-growing area in software engineering. While the need for mobile applications increases, few studies have been conducted to investigate how to generate them through LLMs due to the complexity of handling the mobile ecosystem. In this work, we present and validate a platform-independent methodology for generating and evaluating complete mobile applications from high-level natural language descriptions. Our approach uses a multi-agent LLM framework to manage a three-stage process. First, a detailed mobile app description is generated from a document written in natural language. Next, our methodology converts the specification into executable source code. After that, we evaluate the implemented features and the source code quality. This methodology is built on important principles, including chain-of-thought and role-playing. Our findings show that the methodology produces compilable, stable, and crash-free applications with high-quality source code. However, the rates of feature implementation varied, with many functions being incomplete or missing altogether, which may point out the current limits of fully automated generation. This work gives a standardized process and a reproducible benchmark for evaluating LLM-generated applications. More broadly, this methodology enables the creation of functional prototypes from a high-level app description, thereby accelerating the development workflow.

How to cite