.png){kind=icon}
Return
LLM-Based Explainable Detection of LLM-Generated Code in Python Programming Courses
.png){kind=icon}
PRE
AI
DOI：10.1145/3770762.3772605.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In introductory programming courses, students increasingly submit code generated by large language models (LLMs) instead of solving problems independently. This growing reliance raises concerns about students' development of programming skills. To reduce the overreliance and promote independent problem solving, we propose an explainable detection framework that predicts whether a code submission was generated by an LLM and provides an explanation for its prediction. Such explanations are crucial in educational settings, where transparency and feedback are essential. To support this, we construct a dataset of student-written and LLM-generated code, paired with explanations automatically produced using GPT-4o. We finetune several code-specialized LLMs using both binary labels (student-written or LLM-generated) and explanations. The resulting models achieve over 99% accuracy and generate informative explanations aligned with their predictions, both validated by human instructors. We also apply our detector to past programming course data, revealing a sharp increase in LLM-generated submissions-from nearly 0% in 2022A to over 40% in 2024A. We conclude with a discussion of false positives and suggest how explainable detection can be responsibly deployed in programming courses. Code and prompts are publicly available at https://github.com/ku21fan/LLM-code-detector.
Keywords:
Large language model
LLM-generated code
Explainable detection
LLM-based detector
Explanation generation
Python programming
Journal
P
IF:
0
Papers:
170
Citations:
0
Organization
Cited Papers
.png){kind=icon}
ShareSave
ShareSaveShareSaveShareSave no more
no more