Apple plans to use AI to detect and correct errors in code

Apple is no longer just working on hardware or user interfaces. The company is also playing an increasingly active role in artificial intelligence research. Three recent studies demonstrate how Apple is trying to make software development processes more efficient with the help of AI. The research focuses on three key areas: error prediction, software testing automation, and the autonomous correction of program errors by AI agents.

Many development teams spend a large portion of their time on repetitive tasks such as error analysis, test case creation, or debugging. Apple is addressing the question of whether AI can not only support these processes but also partially take over them. This creates a picture of software development that is more automated – without sacrificing precision or quality. The three studies offer concrete technical approaches and demonstrate what progress has already been made, but also where the limits still lie.

Fault prediction with the ADE-QVAET model

In the first study , Apple is developing an AI model for predicting software errors. The model is called ADE-QVAET and combines four methods: Adaptive Differential Evolution (ADE), Quantum Variational Autoencoder (QVAE), a transformer layer, and Adaptive Noise Reduction and Augmentation (ANRA). The goal is to avoid weaknesses of existing large-scale language models, such as hallucinations, missing context, or the loss of business-relevant context.

The approach is unusual: Instead of analyzing code directly, the model considers metrics such as code complexity, structure, and size. Based on this data, it recognizes patterns that could indicate potential error locations.

Training was carried out using a specific Kaggle dataset for fault prediction. The results demonstrate high accuracy. With a training rate of 90 %, the model achieved an accuracy of 98.08 %, a precision of 92.45 %, a recall of 94.67 %, and an F1 score of 98.12 %. ADE-QVAET thus significantly outperforms conventional machine learning models such as the Differential Evolution ML model – both in terms of reliability and in avoiding false alarms.

Automated test planning with multi-agent system

The second study focuses on writing and managing software tests. According to Apple, quality engineers spend approximately 30 to 40 percent of their time creating basic test artifacts such as test plans, test cases, and automation scripts. Here, Apple relies on an agent-based framework that combines LLMs and autonomous AI agents. The architecture is based on hybrid vector graphics and multi-agent orchestration.

The system automatically creates test plans, tests requirements, and generates validation reports, while fully maintaining traceability between requirements, business logic, and test results.

The improvements are significant. Accuracy increases from 65% to 94.8% compared to the starting point. At the same time, testing time and resources required for the test suite are reduced by 85% and 85% respectively. Projected cost savings are approximately 35%, and project duration can be reduced by up to two months. The results are based on validations in real corporate systems engineering and SAP migration projects.

There are, however, limitations. The study focuses exclusively on specific business areas such as employee systems, financial applications, and SAP environments. Generalizability to other software contexts is currently not possible.

SWE-Gym: AI agents learn to repair code

The third study goes a step further. The goal is no longer just to predict errors or create tests, but to actually fix code. For this purpose, Apple is developing the SWE-Gym training framework, which uses real-world coding tasks from open-source projects to train AI agents. A total of 2,438 Python problems from 11 open-source repositories were used. Each of these tasks is executable and comes with a test suite, allowing the agents to learn under realistic conditions.

There is also SWE-Gym Lite, a slimmed-down version with 230 simpler tasks that allows for faster training and requires less computing power.

The results show that agents trained with SWE-Gym successfully solved 72.5% of the tasks. This is more than 20 percentage points above previous benchmarks. SWE-Gym Lite achieves comparable results but requires almost 50% less training time. The disadvantage: Due to the simpler tasks, the Lite version is only suitable for complex problems to a limited extent.

Apple presents eight AI studies at ICCV 2025

AI becomes an active part of Apple’s development process

The three studies demonstrate how Apple views AI not just as a tool, but as an active component of the software development process. These studies aren't about visions of the future, but rather about concrete systems that have already been tested. ADE-QVAET enables precise error prediction, automated test agents significantly shorten development cycles, and SWE-Gym provides a foundation for AI that can actively improve code. Its applicability is currently limited by the specific focus on certain enterprise systems and the limitations of training data. Nevertheless, the results clearly indicate that AI can not only assist in software development, but also act independently – from error diagnosis to bug fixes. (Image: Shutterstock / bombermoon)