Reproduce “Latent Reuse in Agent Skills: Multi-modal Clone Detection at Ecosystem Scale” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jiaying Zhu, Lyuye Zhang, Wenbo Guo, Yang Liu
Nanyang Technological University researchers developed SkillReuse, a multi-modal clone detection system that quantifies and tracks undeclared reuse in agent skill ecosystems by fusing signals from YAML metadata, natural language instructions, and executable code. The system accurately identifies over a million clone pairs across 66.8% of analyzed skills, revealing widespread modified variants and providing a mechanism to identify a "blast radius" for security vulnerabilities.
Key insights
- Modified variants, rather than exact copies, are the dominant form of reuse in agent skill ecosystems, with 97% of skill pairs in name-based clone families being modified.
- The multi-modal fusion of diverse signals (YAML, NL, code) and the modeling of cross-channel interactions are critical for robust clone detection, particularly for Type-4 semantic clones.
- Widespread undeclared reuse leads to significant registry inflation (estimated 2.45x) and creates a broad "blast radius" for security vulnerabilities across thousands of skills and authors, even when content crosses author boundaries.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.