The paper discusses AutoCodeRover, a novel method for autonomously solving Github issues that encompass fixes and feature additions, enhancing software maintenance and evolution. By combining LLMs with program representations and sophisticated code searching based on program structure, the approach enhances LLM’s understanding, facilitating effective context retrieval and spectrum-based fault localization.

Summary:

• Proposes AutoCodeRover, leveraging LLMs for autonomous program improvement.
• Utilizes program structure for enhanced code searching and root cause analysis.
• Employs spectrum-based fault localization for pinpointing issues.
• Experiments on SWE-bench-lite demonstrate increased efficacy in issue resolution.
• Aims to enable future autonomous software engineering with LLMs auto-improving code.

The integration of LLMs into software development processes, as demonstrated by AutoCodeRover, highlights significant potential for future autonomous software engineering. By enhancing the AI’s ability to tackle maintainability and evolution of code, researchers lay the groundwork for more sophisticated and reliable AI-driven coding assistants. This has the potential to revolutionize the way developers interact with code, transitioning from manual intervention to strategic oversight of automated processes. Read more on arXiv.