ApexCodium: a Multi-Agent System for Code Generation with Enhanced Self-Reflection

LLMs demonstrate a remarkable performance on natural language tasks, but they are relatively less accurate when it comes to code generation. This gap stems from the inherent complexity of coding, which demands precise logical reasoning, meticulous attention to syntactic and semantic details, and systematic problem-solving-challenges that extend beyond conventional natural language understanding. Several studies have been published proposing different approaches to use LLMs for more accurate code generation. In this paper, we present ApexCodium, a multi-agent architecture for enhanced code generation. Our approach proposes dedicated audit agents at successive workflow stages, facilitating persistent evaluation and enhancement of intermediate output. Further, it uses AIgenerated test cases to refine the generated code. We achieved an accuracy of $\mathbf{93.9\%}$ and $\mathbf{90.2\%}$ on HumanEval and MBPP respectively with GPT 40 mini.