Appendix B: Case Study—LLMs for EC Operators

Evolutionary computation relies on genetic operators such as mutation, crossover, and selection to explore solution spaces and drive populations toward fitter individuals. Traditional operators treat candidate solutions as opaque bit strings or numeric vectors, applying syntactic transformations (bit flips, arithmetic perturbations, segment exchanges) without regard for the semantics of the representation. This agnosticism is both a strength and a weakness. While it provides generality across problem domains, it becomes a liability when solutions encode structured artifacts such as programs, mathematical expressions, or natural language prompts. In these cases, random perturbations often produce invalid or nonsensical offspring, wasting fitness evaluations and slowing convergence.