Hebrew University Researchers Map AI Model Lineage With 'Model Atlas'

Researchers from the Hebrew University of Jerusalem have introduced the "Model Atlas," a framework that visualizes populations of machine learning models as an interconnected 3D graph. The project addresses the problem of undocumented AI models by treating the models themselves as data to infer lineage and properties directly from their weights.

By representing models as nodes and weight transformations—such as fine-tuning, quantization, and merging—as directed edges, the Model Atlas provides a structural map of AI evolution. To demonstrate the concept, the team released a public interactive visualization of a subset of Hugging Face models via Cosmograph. In their analysis of over 400,000 documented model relationships, the researchers discovered that 99.41% of quantized models act as leaf nodes, meaning they are the end-point of a transformation chain rather than a base for further edits.

Mapping Model Evolution

The project, proposed as a position paper for NeurIPS 2025, identifies specific structural patterns in how AI evolves. The researchers highlight "Snake" patterns, which represent sequential training checkpoints, and "Fan" patterns, which occur during hyperparameter sweeps. By analyzing these shapes, the framework moves beyond manual documentation, which the authors note is often missing. As Horwitz et al. stated, "Public model repositories now contain millions of models, yet most models remain undocumented and effectively lost."

Implications for Model Forensics

This approach enables a new field of "model forensics" and meta-ML research. By leveraging the graph's structure, the Model Atlas can predict missing attributes for undocumented models—including their intended task, accuracy, and licensing terms—based on the properties of related models in the network. This allows researchers to discover high-performing models that might otherwise remain obscure due to poor documentation.

Furthermore, the atlas helps the industry understand the systemic impact of specific techniques. It allows for a comparative analysis of why certain methods, such as model merging, are widely adopted in Large Language Models (LLMs) but remain less common in vision models. This structural insight provides a roadmap for understanding how LoRA adapters and other fine-tuning methods propagate across different domains.

The Path Forward

As the volume of public models continues to grow, the Model Atlas proposes a shift toward automated mapping to prevent the loss of valuable research. While the current visualization focuses on a subset of Hugging Face, the framework establishes a method for scaling this analysis to millions of models. Future developments will likely focus on refining the predictive accuracy of model attributes and expanding the library of identified structural patterns to better categorize the sprawling landscape of open-source AI.