FGDM: Reasoning Aware Multi-Agentic Framework for Software Bug Detection using Chain of Thought and Tree of Thought Prompting

A new multi-agent AI framework called FGDM (Flow-Graph-Driven Multi-Agent) tackles automated software bug detection by combining four specialized LLM agents that work sequentially. Each agent uses Chain-of-Thought (CoT) and Tree-of-Thought (ToT) prompting to reason through code step by step. The framework first converts source code into a flow graph to capture structural dependencies, then identifies erroneous segments, and finally generates repaired code. It also integrates with FAISS, a vector database, to retrieve similar past bugs and their fixes for context-aware debugging.

The researchers tested FGDM on 100 programs from 10 real-world projects including Ansible, Black, FastAPI, Keras, Matplotlib, Pandas, Scrapy, SpaCy, and Tornado, covering both Python and C codebases. Results showed FGDM outperformed existing deep learning and LLM-based approaches, achieving mean cosine similarity scores of 0.951 for Python and 0.974 for C, while reducing Levenshtein edit distance by 24.33 and 8.37 respectively. This demonstrates the framework's ability to handle large, interconnected codebases where traditional methods struggle with global context.