AI Times, a Korean AI publication, has published Managing Partner Yongduck Kim's column on generative AI patents.
In this column, we introduce the 'Self-Refine' technology, where generative AI evaluates and improves its own answers, focusing on Google's US Patent No. 12682159. Until now, large language models (LLMs) have sometimes missed some conditions or included errors in the answers once generated, but this patent proposes a structure in which multiple candidate answers are generated and then AI directly evaluates the quality and repeatedly corrects them. It is designed to generate answers that more accurately reflect the user's requirements.
In the column, we look at how answer generation, self-evaluation, progressive refinement, and the 'AI-as-a-Judge' concept of evaluating AI are combined from a patent perspective. We also analyze how the structure of individually analyzing the user's multiple conditions to check whether they are met and using the excellent results as training data can increase the accuracy and efficiency of generative AI.
In addition, we note that this patent is not simply a competition to create a larger AI model, but a technology that seeks to implement the ‘self-reflection’ ability to discover and improve AI errors on its own. This approach has the potential to be used as a core technology to implement more reliable generative AI services in various fields such as document writing, translation, coding, medicine, and law, and is expected to have important implications in future AI patent strategies.
The full article can be found at the link below.
This article reflects the information available when it was published. Contact us to discuss your circumstances.