Training Language Models to Generate Quality Code with Program Analysis Feedback

TL;DR

This paper introduces REAL, a reinforcement learning framework that uses program analysis and unit tests to guide large language models in generating secure, maintainable, and functionally correct code without manual annotations.

Contribution

We propose a prompt-agnostic, reference-free reinforcement learning approach that improves code quality and security in LLM-generated code through automated feedback signals.

Findings

REAL outperforms existing methods in code quality and functionality.

The framework is scalable and does not require manual annotations.

Experiments show improved security and maintainability in generated code.

Abstract

Code generation with large language models (LLMs), often termed vibe coding, is increasingly adopted in production but fails to ensure code quality, particularly in security (e.g., SQL injection vulnerabilities) and maintainability (e.g., missing type annotations). Existing methods, such as supervised fine-tuning and rule-based post-processing, rely on labor-intensive annotations or brittle heuristics, limiting their scalability and effectiveness. We propose REAL, a reinforcement learning framework that incentivizes LLMs to generate production-quality code using program analysis-guided feedback. Specifically, REAL integrates two automated signals: (1) program analysis detecting security or maintainability defects and (2) unit tests ensuring functional correctness. Unlike prior work, our framework is prompt-agnostic and reference-free, enabling scalable supervision without manual intervention. Experiments across multiple datasets and model scales demonstrate that REAL outperforms state-of-the-art methods in simultaneous assessments of functionality and code quality. Our work bridges the gap between rapid prototyping and production-ready code, enabling LLMs to deliver both speed and quality.