Which Models Are Our Models Built On? Auditing Invisible Dependencies in Modern LLMs

Sanjay Adhikesaven, Haoxiang Sun, Sewon Min|June 10, 2026arXiv

Key Takeaway

Modern LLMs depend on dozens of upstream models and datasets in complex, recursive ways; ModSleuth makes these invisible dependencies visible and traceable, exposing compliance and transparency issues in LLM development.

Summary

This paper introduces ModSleuth, a system that automatically traces the hidden dependencies between models and datasets used to build modern LLMs.

evaluation data

Key Terms

dependency-graph train-evaluation-coupling artifact source-grounded-evidence