ROCODE introduces a breakthrough approach allowing large language models to detect and fix errors during code generation through strategic backtracking.

• Addresses LLMs' auto-regressive limitations by enabling them to revise previous outputs
• Combines program analysis and backtracking mechanisms to identify and correct code errors
• Outperforms conventional code generation approaches with significantly higher success rates
• Provides practical solutions for real-world software development challenges

This research represents a significant advancement for engineering teams relying on AI-assisted programming, reducing error accumulation and improving overall code quality without requiring additional model training.

ROCODE: Integrating Backtracking Mechanism and Program Analysis in Large Language Models for Code Generation