Scholarlink

<p>Large Language Models (LLMs) increasingly support automated code generation, yet ensuring code quality–including correctness, maintainability, and reliability–remains challenging. This study investigates whether ensemble-based approaches with learned model selection can improve code quality compared to individual commercial LLMs. We evaluate three configurations across 133 programming problems in C++ and Java: three commercial LLMs, an open-source ensemble synthesizing outputs from six smaller models, and a reinforcement learning–based selector that adaptively chooses optimal model subsets. Quality assessment includes functional correctness, cyclomatic complexity, maintainability index, and execution efficiency. Results show that the RL-based selector achieves the lowest error rates (6–20%) while maintaining competitive code quality metrics and achieving 100% correctness on multi-threading tasks (3 tasks) where all commercial models failed. The RL-based approach further reduces inference latency by 40–45% and token usage by 50% compared to full ensemble baselines while maintaining or improving code quality. These findings suggest that learned model selection can enable quality-focused code generation that balances correctness, maintainability, and efficiency on algorithmic programming benchmarks, motivating further investigation into integration with quality-assured development workflows.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Scalable code generation with large language models: an open-source ensemble and reinforcement learning–based selector

  • 期刊论文

  • 发表日期：

    2026年6月20日

查看全文
  • Mohammed A. Shehab,
  • Safwan Omari,
  • Mohammad Wardat,
  • Yaser Jararweh

摘要

Large Language Models (LLMs) increasingly support automated code generation, yet ensuring code quality–including correctness, maintainability, and reliability–remains challenging. This study investigates whether ensemble-based approaches with learned model selection can improve code quality compared to individual commercial LLMs. We evaluate three configurations across 133 programming problems in C++ and Java: three commercial LLMs, an open-source ensemble synthesizing outputs from six smaller models, and a reinforcement learning–based selector that adaptively chooses optimal model subsets. Quality assessment includes functional correctness, cyclomatic complexity, maintainability index, and execution efficiency. Results show that the RL-based selector achieves the lowest error rates (6–20%) while maintaining competitive code quality metrics and achieving 100% correctness on multi-threading tasks (3 tasks) where all commercial models failed. The RL-based approach further reduces inference latency by 40–45% and token usage by 50% compared to full ensemble baselines while maintaining or improving code quality. These findings suggest that learned model selection can enable quality-focused code generation that balances correctness, maintainability, and efficiency on algorithmic programming benchmarks, motivating further investigation into integration with quality-assured development workflows.