Unlocking LLM Code Correction with Iterative Feedback Loops

This study investigates the ability of Large Language Models (LLMs) to self-correct code through an iterative feedback loop, utilizing compiler error messages and test case feedback. The evaluation spans four models and two programming languages, revealing that reasoning models significantly outperform non-reasoning models in code rectification across iterations. These findings highlight the importance of feedback mechanisms for enhancing LLM performance in real-world programming tasks, particularly in addressing syntactic and runtime errors more effectively than logical or algorithmic failures.