Russian Researchers Teach AI Models to Correct Their Own Coding Mistakes
Researchers from T-Technologies and the Higher School of Economics have developed G-Star+, a new method for faster text and code generation. The approach enables AI models to identify and correct errors in draft responses more accurately without requiring full model retraining.
The new algorithm was presented at the ICML 2026 conference in Seoul. It enables an AI model to identify uncertain portions of its output, mask those sections again, and rewrite only the problematic fragments. This approach speeds up generation while reducing computational costs.
“You can think of G-Star+ as a proofreader reviewing a draft and correcting only the parts that are likely to contain mistakes. After reading the entire text, the ‘proofreader’ returns to the questionable passages and revises them. In practice, our method helps produce higher-quality responses more quickly and with lower computational costs,” said Daniil Gavrilov, head of the Fundamental AI Research Laboratory at T-Technologies.
The method can be applied to chatbots, document drafting services, and code completion tools, where fast response times and efficient use of computing resources are particularly important.
