Change-Aware Defect Prediction with Agentic AI
Analysis
This paper challenges the current evaluation practices in software defect prediction (SDP) by highlighting the issue of label-persistence bias. It argues that traditional models are often rewarded for predicting existing defects rather than reasoning about code changes. The authors propose a novel approach using LLMs and a multi-agent debate framework to address this, focusing on change-aware prediction. This is significant because it addresses a fundamental flaw in how SDP models are evaluated and developed, potentially leading to more accurate and reliable defect prediction.
Key Takeaways
- •Traditional SDP evaluation methods are flawed due to label-persistence bias.
- •The paper proposes a change-aware SDP approach using LLMs and a multi-agent debate framework.
- •The proposed approach shows improved performance in detecting defect introductions compared to traditional methods.
- •The source code is publicly available.
“The paper highlights that traditional models achieve inflated F1 scores due to label-persistence bias and fail on critical defect-transition cases. The proposed change-aware reasoning and multi-agent debate framework yields more balanced performance and improves sensitivity to defect introductions.”