자동 프로그램 수리의 프로세스효율 최적화를 위한 우선순위 기반 LLM-자율 에이전트 = Priority-Based LLM Autonomous Agent for Process Efficiency Optimization in Automated Program Repair

다국어 초록 (Multilingual Abstract)

Recent advances in Large Language Models (LLMs) have significantly improved automated program repair (APR) in terms of code understanding and reasoning.However, most existing methods still rely on FSM-based static workflows, which often cause ineffic ncies in tool invocation and validation phases. To
overcome these limitations, this study proposes the Priority-Based RepairAgent (PB-RepairAgent),an LLM-driven autonomous repair framework that integrates a priority-based control structure.The proposed system dynamically adjusts the order and intensity of patch generation, validation, and testing according to bug type, complexity, and test failure patterns, thereby minimizing redundant resource usage and enhancing overall repair efficiency.Experiments conducted on the Defects4J v1.2 and v2 datasets (835 bugs in total) demonstrate that PB-RepairAgent improves Fix Accuracy by 0.8–1.0%p,reduces average execution time by 8–10%, and decreases token consumption by approximately 30K compared to the baseline FSM-based RepairAgent.Additionally, the LightGBM-based State Controller achieved the most stable and efficient exploration strategy among evaluated models.In conclusion, PB-RepairAgent effectively combines LLM autonomy with priority-based decision control, achieving notable improvements in accuracy, efficiency, and generalization.These results empirically validate the feasibility
of dynamic autonomous control and resource optimization in LLM-based automated program repair

목차 (Table of Contents)

• 제 1장 서론 1
• 제 2장 배경지식 및 관련 연구 5
• 제 3장 연구 방법 15
• 제 4장 연구 설정 24
• 제 5장 연구 결과 30
• 제 6장 위협 요소 40
• 제 7장 토의 : PB-RepairAgent의실용적적용가이드라인 42
• 제 8장 결론 및 향후과제 46