Tag
This paper empirically analyzes the cost-effectiveness of code execution in LLM-based program repair agents, finding that execution is used heavily but often indiscriminately, and that restricting execution can save significant cost with minimal impact on repair success.