Recent advancements in large language models (LLMs) have exhibited promising capabilities in addressing various tasks such as defect detection and program repair. Despite their prevalence, LLMs still face limitations in effectively handling these tasks. Common strategies to adapt them and improve their performance for specific tasks involve fine-tuning models based on user data or employing in-context learning with examples of desired inputs and outputs. However, they pose challenges for practical adoption due to the need for high-quality data, extensive computational resources, and continuous maintenance. Furthermore, neither strategy can explain or reason about the ineffectiveness of LLMs in the given tasks.

We propose Calico to address the high cost of fine-tuning, eliminate the necessity for task-specific examples, and provide explanations of LLM deficiency. At the heart of Calico is an evolutionary approach that interleaves knowledge calibration and AI deficiency diagnosis. The key essence of Calico is as follows. First, it focuses on identifying knowledge gaps in LLMs’ program comprehension. Second, it conducts automated code refactoring to integrate the overlooked knowledge into the source code for mitigating those gaps. Third, it employs what-if analysis and counterfactual reasoning to determine a minimum set of overlooked knowledge necessary to improve the performance of LLMs in code tasks.

We extensively evaluated Calico over 8,938 programs on three most commonly seen code tasks. Our experimental results show that vanilla ChatGPT cannot fully understand code structures. With knowledge calibration, Calico improves the vanilla LLM by 20% and exhibits comparable proficiency compared to fine-tuned LLMs. Deficiency diagnosis contributes to 8% reduction in program sizes while ensuring performance. These impressive results demonstrate the feasibility of utilizing a vanilla LLM for automated SE tasks, thereby avoiding the high computational costs associated with a fine-tuned model.