OpenAI Develops CriticGPT Model Capable of Spotting GPT-4 Code Generation Errors
OpenAI has recently developed a model called CriticGPT, which has not been released to the public yet. This new artificial intelligence (AI) model aims to identify and rectify errors in code generation by GPT-4. CriticGPT was trained using the reinforcement learning from human feedback (RLHF) framework, utilizing a GPT-4 model. This was reported by SSPDaily.
In their study, OpenAI discovered that when users received assistance from CriticGPT to review code generated by ChatGPT, they outperformed those without such assistance in 60% of cases. The company blog post provides more insights into the details of the CriticGPT model and its development process. The research findings have been published in a paper.
RLHF involves training AI systems by combining machine output with human evaluators who provide feedback on the AI's performance. Known as AI trainers, these evaluators help in adjusting and enhancing the behavior of the AI model. In the case of CriticGPT, the model was trained on a vast dataset of code containing errors. The AI trainers were tasked with identifying and critiquing the mistakes in the code, including both naturally occurring errors and those intentionally inserted by them.
During the training process, CriticGPT presented various versions of code critiques to the trainers, who were then asked to determine if the errors they had introduced were effectively caught by the AI, along with the naturally occurring mistakes. OpenAI's research revealed that CriticGPT performed 63% better than ChatGPT in detecting errors.
However, it is important to note that CriticGPT does have certain limitations. It was primarily trained on short code snippets and has yet to be exposed to more complex and lengthy coding tasks. Additionally, the model still exhibits instances of generating incorrect factual responses and has not been tested extensively with multiple dispersed errors within the code.
Given its focus on enhancing training techniques for generating higher quality outputs, it is unlikely that CriticGPT will be made available to the public. If it does become accessible, it is anticipated that it will be incorporated within the ChatGPT system to assist users in achieving more accurate and reliable code generation.