Iterative program synthesis with code knowledge

Abstract

Component-based synthesis aims to automatically synthesize programs from a collection of components, such as Java programs composed of application programming interfaces (APIs). However, synthesizing complex tasks remains inefficient due to the intractability of program space, which grows exponentially with the length of code. When faced with unfamiliar and complex tasks, programmers often turn to the code available online, reusing existing code knowledge to solve them. Inspired by this, we propose Itas, an explainable iterative program synthesis framework that incorporates code knowledge to reduce the complexity of program space and leverages modular encapsulation to synthesize more functions. Specifically, Itas narrows the program space externally with API knowledge, expands the program space internally with an iterative strategy to improve synthesis efficiency, and encapsulates non-API elements as general APIs to enhance modularity. To evaluate the effectiveness of Itas, we collect a benchmark of real programming tasks and conduct experiments on the conventional program synthesizers SyPet and FrAngel. The experimental results demonstrate that integrating with Itas substantially enhances synthesis efficiency, reducing the synthesis time by 97.02% compared to SyPet and 48.43% compared to FrAngel. Furthermore, comparisons with large language models show that Itas outperforms them in pass rate, successfully solving all tasks.