EMASFL: Ensemble of Multi-Agent Small Language Models for Local Fault Localization

Large Language Models (LLM) have significantly advanced Fault Localization (FL) in software debugging but face challenges regarding data privacy, high API costs, and latency. While Small Language Models (SLM) offer a fast, secure, local alternative, they typically lack the reasoning capabilities required for autonomous software repository exploration. This paper proposes EMASFL, a novel local FL system integrating multi-agent architecture with ensemble method. EMASFL overcomes SLM limitations by decomposing the debugging process into a four-step structured workflow executed by specialized agents. The system employs a hybrid strategy, assigning distinct SLM to agents based on their specific strengths in following instructions or code understanding, and integrates outputs from diverse agent configurations using an ensemble method. Evaluations on the Defects4J benchmark demonstrate that EMASFL significantly outperforms existing SLM-based and traditional FL techniques. Notably, it correctly localized 169 bugs, surpassing commercial LLM-based approaches utilizing GPT-3.5 and GPT-4o mini. Operating with an average runtime of 93 seconds and zero API costs, EMASFL provides a practical, high-performance, and privacy-preserving solution for automated software debugging.