Launder-Bench: coding agent rewrites are hard to trace to their source
The point
Coding agents can rewrite repository scale codebases while preserving tested behavior, creating a challenge for plagiarism detection, license auditing, and software provenance. Launder-Bench, built from 282 real world repositories, tests whether such rewrites can still be traced to their source.
Why it matters
For builders, similarity alone is not provenance evidence when agents rewrite whole repos.
Key details
- Two sets: 1,128 pairwise instances and 229 source attribution instances, with abstention allowed when evidence is insufficient
- Strongest static method reached only 0.66 AUROC on pairwise detection
Original ~12 min → 25 sec here Source: arXiv cs.AI, Nov 1, 2026