Supervisor
Dr. Mehmet Kerem Turkcan
Description
There are abundant researchers currently attempting to utilize GPT for natural language programming (i.e. Parsel, LEVER). Because GPT (i.e. ChatGPT) can produce wrong code when programmers use it for code generation, an overall idea was to have it create each function individually with multiple passes, starting from the leaves in the function-tree hierarchy until the root is hit.
I am to study the effectiveness of different GPT debugging frameworks, such as ideas like the one aforementioned, with heuristics such as token cost and accuracy. As debugging is an arduous task (i.e. understanding the bug type, understanding the schemantics of what the code is supposed to do and what it is doing), a considerable amount of design decisions must be understood for this to be accomplished.
More to come soon as this project has not yet been submitted to a journal/conference!