New Apple studies show AI can detect defects and repair code With 98% accuracy

Apple researchers have released three new studies outlining advanced AI models designed to predict software defects, automate testing, and even repair faulty code — highlighting how artificial intelligence could reshape software engineering workflows.

The research, published in October 2025 on Apple’s Machine Learning Research blog, explores different stages of the software lifecycle — from identifying potential defects before they occur, to generating and validating tests, and ultimately training AI agents capable of fixing bugs autonomously.

AI Model Predicts Software Defects with 98% Accuracy

In one study, titled Software Defect Prediction using Autoencoder Transformer Model, Apple researchers introduced a new hybrid AI model known as ADE-QVAET (Adaptive Differential Evolution–Quantum Variational Autoencoder Transformer).

The model combines multiple AI techniques — including differential evolution, quantum-inspired learning, and transformer architecture — to analyze large-scale software metrics and predict where bugs are most likely to appear.

Unlike traditional large language models that process code directly, ADE-QVAET focuses on structural and complexity-based metadata to identify risk areas. When tested on a Kaggle dataset for bug prediction, the system achieved 98.08% accuracy, 92.45% precision, and 94.67% recall, significantly outperforming existing machine learning approaches.

The researchers say this technique could help developers catch software defects early, reducing debugging workloads and improving release reliability.

Multi-Agent AI System Automates Software Testing

A second study, Agentic RAG for Software Testing with Hybrid Vector-Graph and Multi-Agent Orchestration, presents an AI-driven testing system that uses large language models (LLMs) and autonomous agents to generate and manage test plans, cases, and validation reports automatically.

By combining retrieval-augmented generation (RAG) with a hybrid vector-graph data structure, the system maintains traceability between business requirements and test results — a key challenge in enterprise-scale software projects.

In trials involving corporate systems and SAP migration projects, Apple’s framework delivered a 94.8% accuracy rate, reduced testing timelines by 85%, and cut costs by 35%. Researchers estimate the approach could accelerate project completion by up to two months.

However, Apple notes that the model’s testing was limited to specific enterprise environments, suggesting further work is needed before broader deployment.

Training AI Agents to Read and Repair Code

The third study, Training Software Engineering Agents and Verifiers with SWE-Gym, shifts focus from detection to correction. It introduces SWE-Gym, a large-scale training environment containing over 2,400 real-world Python programming tasks drawn from open-source repositories.

SWE-Gym enables AI models to practice debugging in a controlled environment that includes runnable codebases and test suites. Apple’s AI agents trained with this setup solved 72.5% of the given tasks — outperforming prior benchmarks by over 20 percentage points.

A smaller variant, SWE-Gym Lite, achieved comparable results while halving training time, suggesting potential for more efficient model training without sacrificing performance.

The framework also introduces “verifiers” — secondary AI models that evaluate the reasoning steps of code-editing agents, further improving solution accuracy.

Together, these studies mark a significant step in AI-assisted development. Apple’s research points toward a future where agentic systems can predict code-level failures, generate comprehensive test documentation, and autonomously correct errors — all with minimal human intervention.

While practical integration remains limited to experimental and enterprise settings, the findings demonstrate how machine learning continues to move beyond code completion toward end-to-end automation in software engineering.