Research shows that the GPT-4 model has self-correction capabilities and is expected to promote the further commercialization of AI code

News on July 5th, researchers from the Massachusetts Institute of Technology (MIT) and Microsoft found that the GPT-4 model has excellent code self-correction capabilities, while GPT-3.5 does not have this feature. The paper has been published in ArXiv.

▲ Picture source ArXiv

A number of AI models specifically designed for code have emerged on the market, but currently most of them only serve The role of assisting developers in writing code, such as the Microsoft Copilot Assistant that friends at IT House are familiar with, these AI models can only generate code snippets at the moment, so they cannot completely replace human developers.

Through the study of GPT-4, researchers said that it is actually possible to use "model self-correction" to make the model "reflect on its own shortcomings" to increase the length of code snippets and improve The accuracy of the output results.

▲ Image source ArXiv

▲ Image source ArXiv

Researcher from the University of Edinburgh Fu Yao said that only GPT-4 has the ability to self-improve, while the weaker GPT-3.5 does not have this feature. This finding indicates that large models may have a new type of ability, that is, through a series of user feedback Let AI self-correct and ultimately obtain results that satisfy users. This self-correction ability may only exist in sufficiently mature AI models.

After self-correction, 71% of the code output by the GPT-4 model met the requirements set by the researchers, and the code generated by GPT-3.5 was corrected using GPT-4 Finally, the pass rate of this batch of codes also reached 54%.

The above is the detailed content of Research shows that the GPT-4 model has self-correction capabilities and is expected to promote the further commercialization of AI code. For more information, please follow other related articles on the PHP Chinese website!