Effective Code Generation Using Program Synthesis And CodeLLM

Code generation for a given specification is a holy grail problem in computer science, traditional methods like program synthesis searches for the program which satisfies the given specification from the search space, ideally this become a searching problem when considering program synthesis for code generation. The research focuses on enhancing the accuracy, efficiency, and scalability of automated code generation by combining the strengths of program synthesis techniques and Code Large Language Model (CodeLLMs). Program synthesis involves generating executable code from high-level specifications, ensuring that the generated code adheres to formal correctness and functional requirements. On the other hand, CodeLLMs, which are built on the foundation of deep learning, leverage vast amounts of code-related data to understand, generate, and refine human-like code across various programming languages. By integrating these two approaches, this research addresses critical challenges in the traditional code generation process, such as logical errors, poor handling of ambiguous instructions, and limited contextual comprehension. One of the central aims of this reserach is to improve the contextual awareness of CodeLLMs through program synthesis, ensuring that the generated code not only matches the syntax and semantics of a given problem but also adheres to formal correctness guarantees. This dual approach not only ensures high-quality code generation but also significantly reduces the chances of introducing errors or bugs that could otherwise emerge during manual coding. To evaluate the performance of the proposed system, sophisticated evaluation metrics like CodeBLEU, which takes into account syntactic correctness, semantic similarity, logical consistency. The CodeBELU score is compared for the system which was provided context and without context , the system acheives a improvement of 49.3% in code generation.