A Hybrid Feature Fusion Framework for Just-in-Time Software Defect Prediction using LLM Reasoning and CodeBERT Embeddings

Just-in-Time (JIT) software defect prediction tries to flag risky commits the moment they are committed, before they can cause problems in a codebase. Most existing approaches rely on simple change metrics, such as how many lines were added or how experienced the developer is, but these numbers say very little about what a change actually does. This paper presents a hybrid framework that looks at a commit from three angles at once: traditional change metrics, a semantic understanding of the code from CodeBERT, and structured reasoning about intent and risk generated by a large language model. We built a multi-language dataset of nearly 60,000 commits from 19 open-source Java, Python, and C/C++ projects, and tested several ways of combining these three sources of information using standard machine learning classifiers. Combining traditional metrics with CodeBERT's code understanding consistently gave the strongest results, clearly outperforming any single source used alone. Adding the LLM-based reasoning on top of that combination did not meaningfully improve things further, though it was genuinely useful when paired with either of the other two sources on its own. Overall, our results suggest that combining structural and semantic information is a practical way to catch defects early, while a third, coarser reasoning layer adds little once the first two are already working together.