QualCoder: A Quality-Driven Multi-Agent Framework for Automated Code Generation

Abstract

Generating high-quality code for problem-solving remains a major challenge as it re quires both linguistic understanding and algorithmic reasoning. Recent multi-agent frameworks have shown great potential for code generation through collaborative planning, implementation, and debugging, yet their performance is hindered by re dundant low-quality plans, the inability to decompose high-level plans into simpler subtasks, and the lack of evaluation against given test cases. We introduce Qual Coder, a novel multi-agent framework that enhances code quality and generation efficiency by using (i) a confidence-aware planner that filters low-quality plans, (ii) a sub-task planning agent that decomposes high-level plans into smaller, executable tasks, and (iii) a quality assurance agent that validates plans and code against test cases. Our experiments across problem-solving benchmarks reveal that QualCoder improves performance over current best approaches by roughly 2% on HumanEval and24.10%onCodeContest,whilereducingthenumberofAPIcallsandtokencount. Further analysis reveals the efficacy of QualCoder in reducing error propagation by filtering plans with lower confidence and using test cases as preliminary evaluations