Genetic analysis of model merges (multi-parent reproduction)

Develop methods to conduct a genetic analysis of model merges in the Hugging Face ecosystem, treating merges as a form of sexual reproduction with two or more parents and analyzing trait inheritance and mutation in non-tree structures where nodes may have multiple parents.

Background

The paper analyzes genetic similarity and trait diffusion across fine-tuning family trees on Hugging Face, deliberately omitting merges to preserve tree structures analogous to asexual reproduction. In this setting, each model has at most one parent, simplifying the measurement of inheritance and mutation over generations.

The authors note that merges introduce multiple-parent relationships, creating non-tree directed acyclic graphs and more complex local family structures. Because merges are rarer and structurally complex, they leave their genetic analysis as an open direction for future work.

References

We leave this as an open direction the genetic analysis of model merges, which can be thought of as a form of sexual reproduction with two or more parents. Graphs depicting model merges are no longer trees, and the set of local family structures is more complex.

— Anatomy of a Machine Learning Ecosystem: 2 Million Models on Hugging Face  (2508.06811 - Laufer et al., 9 Aug 2025) in Section “Measuring genetic similarity,” Analyzing fine-tuning trees (footnote)

A second problem is epistasis: skills that depend on one another make blind merging dangerous, and merging then pays only when done sparingly and with the offspring screened (39--41; Figs. S10--S11, S13), but nobody yet has a way of measuring how entangled two real skills are.

— The evolution of sex for artificial intelligence: a population-genetic framework for multigenerational model populations  (2609.18560 - Gilestro, 16 Sep 2026) in Discussion, paragraph beginning “The open problem is the fitness function.”