Core Concepts
AGENTFL introduces a multi-agent system for automated fault localization, enhancing Large Language Models (LLMs) to diagnose bugs at the project level.
Abstract
Abstract: AGENTFL addresses limitations of LLMs in fault localization by decomposing the process into comprehension, navigation, and confirmation stages.
Introduction: FL is crucial but time-consuming; SBFL and LBFL are common techniques. LLMs show promise but struggle with long contexts.
Approach: AGENTFL models FL as a three-step process using multiple agents with specialized expertise. Strategies like Test Behavior Tracking and Document-Guided Search are employed.
Evaluation: AGENTFL outperforms LLM-based baselines in localizing bugs within Top-1 on Defects4J-V1.2.0 benchmark. Complementarity with existing techniques is observed.
Experiment Design: Research questions focus on performance, design choices' impact, and practical usability through a user study.
Stats
AGENTFL can localize 157 out of 395 bugs within Top-1.
AGENTFL spends an average of only 0.074 dollars and 97 seconds for a single bug.