Evolutionary model merging searches two composition spaces: parameter recipes and data-flow graphs. Across the paper’s experiments, these searches produced a Japanese math model and a Japanese vision-language model by rearranging existing open models rather than training a new monolith. The distinction sharpens the Compositional Intelligence connection: selection can range over reusable parts and, where applicable, their execution order.
Model merging usually depends on someone with good intuition picking what to combine. This runs an evolutionary search over the recipe instead, in two spaces at once.
Parameter space asks which weights to mix and how. Data-flow space asks which layers to keep, from which model, in what order. Searching both together finds combinations that neither space reaches alone.
The method produced a Japanese-language math model and a Japanese vision-language model, in both cases by rearranging existing open models rather than by training a new one.