This paper introduces the ‘Toggle’ framework, a new approach to Automated Program Repair (APR) that leverages Large Language Models to predict and fix bugs at a granular token level. The framework separates bug localization and fixing into two distinct processes, improving accuracy and efficiency. Here are the key highlights:

• Token-Granulated Bug Localization and Repair: By segmenting the process, the model handles localization and repair separately, improving performance over integrated approaches.
• Inductive Bias Utilization: The study examines different styles of prompting the bug fixing model, identifying those that signi… 1.- New State-of-the-Art Performance: Achieves top results on the CodeXGLUE code refinement benchmark. 2.- Broad Application: Shows potential for adoption in other APR datasets like Defects4J.

The distinction and methodological innovation introduced by the ‘Toggle’ strategy mark a significant advance in the field of Automated Program Repair. It showcases how LLMs can be utilized more effectively within software engineering, potentially leading to broader implications for automated maintenance and quality assurance in software systems.