Evolution through large models

The paper proposes Evolution through Large Models (ELM), using code-trained LLMs to improve mutation in genetic programming.

In experiments, ELM with MAP-Elites produced many working Python walkers in the Sodarace domain.

Those generated programs were then used to train a new model that can choose a suitable walker for a given terrain.

The result suggests code-generating LLMs can boost evolutionary search and bootstrap training data for new tasks with no prior examples.