Process evidence triples code authorship detection in education

A new study from Marek Horváth (arXiv:2607.07400) evaluates how well source-code authorship attribution works in programming education compared to competitive programming contests. The paper highlights that existing methods, often trained on contest or open-source data, perform dramatically worse in educational settings where students solve shared assignments with evolving coding habits. While contest data achieves a mean top-1 accuracy of 0.938, educational datasets using only final static code drop to a mere 0.094 — nearly random.

By incorporating process evidence — behavioral features like editing sequences, pause patterns, and compile events captured during development — the educational mean top-1 jumps to 0.233, and the pairwise verification ROC-AUC improves from 0.556 to 0.752. This shows that process data can complement weak final-code signals, but the accuracy remains far below contest levels. The author emphasizes such models are appropriate only as instructor-mediated decision support, not as independent proof of authorship. Educators should treat these tools as alerts for further review, not definitive evidence.