On 13 November 2026, researchers introduced Model Maps, a documentation framework for tracing the ancestry of foundation models. Their paper describes a model map as a knowledge graph of the artifacts and relationships that influence a model, extending earlier documentation approaches such as model cards and datasheets.[1]

The authors argue that foundation models can inherit properties from upstream models, datasets and benchmarks. They call the resulting web of relationships AI entanglement and say that limited visibility can make it harder to identify the origins of vulnerabilities, hallucinations or biased predictions. Their proposed framework identifies 13 types of dependency, including fine-tuning and synthetic data generation.[1]

The paper reports that the authors constructed maps for 15 closed and open-weight foundation models. Across those maps, they report 2,778 unique dependencies and artifacts as far as nine hops from a resulting model. The abstract says the researchers released the maps and argues that transparency requirements should cover upstream dependencies as well as individual models and training data.[1]

Sources

  1. Model Maps: A Documentation Framework for Tracing Foundation Model Ancestry and Understanding AI EntanglementarXiv · 13 Nov 2026

Check or correct this page

Verification record

The digest below shows whether the text you are reading matches the published source. It does not show that the reporting is correct. The build computes it from this page's Markdown source and lists it in the site manifest.

47d5f2494f9739d1efbb7dab604b7c7a24179c3bf5a0e1c975525d9a60aca86c