REPOPILOT: Software Agents To Resolve Software Engineering Tasks at Repository-Level Scale
Abstract: Coding assistants based on Large Language Models (LLMs) have recently surged in popularity. A significant challenge for LLMs is accurately responding to user queries at the scale of entire code repositories. We propose RepoPilot, a multi-agent-based system capable of effectively navigating through source code repositories to collect relevant information, editing code and execute programs. We demonstrate the effectiveness of RepoPilot through extensive evaluations on challenging benchmarks, including SWE-bench and an automatically collected code generation dataset. On SWE-bench Lite, RepoPilot achieves a 17\% pass rate, establishing competitive results compared to the baseline while maintains low cost and also excels in other code intelligence tasks.
Paper Type: Long
Research Area: NLP Applications
Research Area Keywords: NLP Applications
Contribution Types: Model analysis & interpretability, NLP engineering experiment
Languages Studied: English
Submission Number: 5830
Loading