Analysis of Compiler-Level Static and Dynamic Features for Automated Bug Prediction Using Transformer Models
Veena Janardhan Jadhav, Prakash Devale, Nischal Puri, Priya Parkhi · Automation and Remote Control · 2026
Abstract The rapid evolution of software systems has increased the complexity of codebases, making traditional bug detection and debugging techniques inefficient and resource-intensive. Automated bug prediction systems using cutting-edge machine learning (ML) and deep learning (DL) architectures have drawn a lot of attention in an effort to overcome this difficulty. However, the majority of current approaches rely on historical datasets or static code features, which fail to capture runtime behaviors and intermodule interactions that frequently result in failures. Many frameworks show severe computational overheads, lack scalability across big heterogeneous projects, and are restricted to particular programming languages or benchmark datasets. This paper suggests an integrated method that uses dynamic runtime features and static compiler-level features for thorough bug prediction to get over these restrictions. In this work, transformer-based hybrid models for automated bug prediction are used to analyze compiler-level static and dynamic information in depth. To examine their prediction skills across a variety of datasets, three frameworks: Rust-IR-BERT, nature-based ensemble prediction model, and hybrid CNN-LSTM were compared. To improve the accuracy of bug discovery, the suggested DynastBug framework combines static and dynamic analysis with transformer encoders and hybrid ensemble learning. It captures semantic relationships between code components, monitors execution flow, and leverages attention-based mechanisms to enhance feature representation. Experimental results demonstrated that DynastBug outperformed existing models with an accuracy of 99.2%, precision of 97.8%, recall of 97.3%, and F1-score of 99.3%. The results confirm that integrating compiler-level semantics with runtime behavior significantly improves generalization, reliability, and scalability in automated software bug prediction.