Core Concepts
AGENTFL is a multi-agent system based on ChatGPT for automated fault localization, addressing the limitations of existing LLM-based techniques in handling large codebases.
Abstract
Fault Localization (FL) is crucial in software debugging, with developers spending significant time on this phase.
Large Language Models (LLMs) show promise in bug diagnosis but struggle with long contexts.
AGENTFL decomposes fault localization into comprehension, navigation, and confirmation stages using multiple agents.
Evaluation on Defects4J-V1.2.0 shows AGENTFL outperforms other LLM-based approaches in localizing bugs within Top-1.
Components like Test Behavior Tracking and Document-Guided Search enhance AGENTFL's performance.
Ablation study highlights the importance of tasks like Test Failure Analysis and Method Doc Enhancement.
Stats
AGENTFLは395のバグのうち、Top-1で157をローカライズできます。
AGENTFLはDefects4J-V1.2.0で他のLLMベースのアプローチを上回ります。
AGENTFLはClosureプロジェクトで他のアプローチよりも効果が低いことがわかりました。