Context: Large Language Models (LLMs) have recently reframed Automated Program Repair (APR) as a conditional code-generation task. However, practical adoption faces challenges such as overfitting to incomplete test suites, high token/computational costs, and inconsistent evaluation methodologies. Objectives: This paper presents a unified framework to analyze how orchestration strategies (single-shot vs. iterative feedback) and parameter-efficient fine-tuning via Low-Rank Adaptation (LoRA) influence repair effectiveness, efficiency, and patch quality in single-function Java defects. Methods: We implement and compare four repair pipelines—linear and iterative variants, with and without LoRA—across both open-source and proprietary LLMs using standardized benchmarks (Defects4J and HumanEval-Java). The design isolates the impact of iterative orchestration from that of LoRA-based specialization under a consistent evaluation protocol. Results: Our findings show that iterative prompting markedly increases the number of plausible and correct patches while controlling token and time costs. When combined with iterative feedback, LoRA further improves semantic correctness with minimal overhead. Conclusion: This study is the first to systematically disentangle the effects of repair orchestration and lightweight specialization using LoRA within a common framework. We provide empirical evidence that open-source models, enhanced by LoRA and feedback, can match the performance of proprietary models while offering shorter, cheaper patches. By experimentally disentangling orchestration from specialization within a unified protocol, the study clarifies their respective contributions to effectiveness and efficiency in LLM-based APR.
Patricio et al. (Mon,) studied this question.