CodeProbe
A universal, interpretable detector for LLM-generated code.
CodeProbe is a proposed framework for the universal detection of LLM-generated code.
The Challenge
Current detection tools rely heavily on “black-box” machine learning models. While effective, they lack interpretability—developers are told code is “AI-generated” without knowing why—and they often struggle when faced with out-of-distribution data (code languages or models they weren’t trained on).
Our Approach
We are investigating the use of Knowledge Units (KUs)—atomic representations of coding patterns—to distinguish between human-written and AI-generated code. By analyzing the density and distribution of these units, CodeProbe aims to provide a detection method that is both highly accurate and human-interpretable.
