New AI tool tracks code provenance: SourceTracker detects plagiarism in LLM outputs

Large language models (LLMs) for code generation can reproduce training examples verbatim without attribution, raising serious legal and ethical concerns around plagiarism and license compliance. Traditional fingerprint-based detectors like Winnowing are effective but require linear-time searches over entire training sets, making them impractical for the billion-scale corpora used by modern code LLMs.

To address this, researchers from the University of Bologna and others present SOURCETRACKER, a 300M-parameter encoder optimized for code retrieval, paired with a hybrid two-stage pipeline called HYBRIDSOURCETRACKER (HST). HST first uses vector search to narrow candidate snippets, then re-ranks them via exact Winnowing fingerprints. Evaluated on a 10M-snippet subset of TheStackV2, the system achieves mean reciprocal rank on par with Winnowing for 30-token fragments and surpasses it by 5.4% for 60+ token fragments, all while maintaining logarithmic-time complexity. This enables scalable, high-precision provenance tracking essential for verifying the originality of LLM-generated code.