Efficient and Scalable Provenance Tracking for LLM-Generated Code Snippets

Researchers proposed SOURCETRACKER and HYBRIDSOURCETRACKER to trace where LLM-generated code snippets may have come from in training data.

HYBRIDSOURCETRACKER combines vector search to find likely matches with fingerprinting to verify provenance, improving scalability over fingerprinting alone.

In tests on THESTACKV2, the approach matched or beat pure fingerprinting on longer code fragments, including cases with identifier renaming.

The method could help with plagiarism checks, license compliance, and large-scale provenance auditing for code models.