COAST introduces a novel approach to improving how Large Language Models debug code by simulating the complete human debugging process, not just code repair.

• Addresses four critical debugging stages: Bug Localization, Bug Identification, Code Repair, and Code Recognition
• Uses a communicative multi-agent system to synthesize high-quality training data
• Significantly improves debugging abilities of both open-source and proprietary LLMs
• Creates a new comprehensive evaluation benchmark that tests the full debugging workflow

This research matters for Engineering by providing a more holistic approach to code debugging in AI systems, potentially reducing development time and improving software quality in production environments.

COAST: Enhancing the Code Debugging Ability of LLMs through Communicative Agent Based Data Synthesis