Which Models Are Our Models Built On? Auditing invisible dependencies in modern LLMs

arXiv·medium signal

This paper (arXiv 2606.12385, Adhikesaven, Sun, Min) argues that modern LLM training pipelines increasingly depend on other models to generate synthetic data, filter corpora, judge outputs, and guide development — creating invisible model-on-model dependencies that are rarely disclosed. The authors propose auditing methods to trace these hidden lineage relationships. For agent builders, it reframes 'AI supply chain' to include the upstream models baked into a model's own training, a transparency and provenance risk distinct from package or MCP supply chains.