Building Family Trees for AI: Studying Large Language Models Like Darwin Studied Finches

This post introduces a 2026 arXiv paper ("Analysis and Explainability of LLMs Via Evolutionary Methods") that proposes treating large language models like biological species and using evolutionary methods to reconstruct their family trees.

The Darwinian Analogy

Just as Darwin inferred relationships among Galápagos finches by observing beaks and behavior—without dissecting cells—we can understand the thousands of LLMs (Llama 3, GPT-4, Claude 3, and countless fine-tunes) by studying their ancestry rather than their internals.

The paper's core mapping:

1. Genotype = model weights: Billions of parameters are a model's genetic code. Fine-tuning is like lab genetic modification—small changes, but recorded in the weights. 2. Phenotype = model outputs: For black-box models (e.g., GPT-4), we can only observe the "beak"—how the model speaks and answers.

Three Things Phylogenetics Can Do

Shannon Gallagher's team used phylogenetic methods to accomplish three things:

1. Catch plagiarists and hybrids

By comparing weights, researchers can reconstruct a model's lineage. If someone secretly fine-tunes another's model or merges two models, the evolutionary tree clearly shows the parentage—essentially a paternity test for AI.

2. Locate where evolution happens

Just as Darwin linked beak changes to environmental adaptation, researchers found that when a general model specializes (e.g., into a medical model), only specific layers undergo drastic change—hinting at where domain knowledge is stored.

3. Trace black-box models by behavior

For models without accessible weights, comparing output (phenotype) similarity still produces highly accurate classification trees, revealing which models are near-duplicates and which represent genuinely new lineages.

Why It Matters

Traditional interpretability tries to explain AI by analyzing individual neurons—like explaining why cats catch mice by tracking atomic collisions. The evolutionary perspective offers a macro view instead:

• Which ancestor did a model's bias inherit from?
• Which training data was the key "mutation" behind its intelligence?
• Is a new model truly innovative, or just a hybrid?

Conclusion: Every LLM response carries the imprint of its long evolutionary history. We are shifting from "engineers who build AI" to "naturalists who observe AI"—and Darwin's theory, 150 years later, has become a powerful tool for understanding artificial intelligence.