How OpenAI's CriticGPT Enhances Code Quality in ChatGPT

OpenAI has launched CriticGPT, an innovative AI model designed to enhance the accuracy of code generated by its ChatGPT system. Leveraging Reinforcement Learning from Human Feedback (RLHF), CriticGPT aims to significantly improve alignment within AI systems.

Developed on the foundation of their advanced AI model GPT-4, CriticGPT assists human AI reviewers in scrutinizing code generated by ChatGPT. In a recent research paper titled "LLM Critics Help Catch LLM Bugs," the tool demonstrated effective competence in analyzing code and identifying errors, including subtle issues that human reviewers might overlook.

The study highlighted that CriticGPT’s feedback was preferred by annotators over human-generated notes in 63% of cases involving errors within large language models (LLMs). By employing a novel technique known as "Force Sampling Beam Search," the tool not only aids in reducing the occurrence of hallucinations but also facilitates more comprehensive critiques compared to conventional human or AI-only evaluations.

Users can adjust CriticGPT’s sensitivity to errors, thereby gaining control over its propensity to identify nonexistent issues. However, the tool does have limitations, such as its struggle with evaluating longer and more complex coding tasks, as it is primarily trained on shorter responses from ChatGPT.

Furthermore, in complex coding scenarios where errors span multiple code strings, CriticGPT may face challenges in pinpointing the root causes of discrepancies, potentially leading to its own misjudgments.

Source: The Hindu