UnlearnEvo DSL: An LLM-Driven Evolutionary DSL for Machine Unlearning

Machine unlearning requires designing update procedures that remove the influence of selected data while preserving model utility, yet the space of possible procedures is highly combinatorial and difficult to explore manually. Existing LLM-guided evolutionary methods offer a promising path toward automated algorithm discovery, but their reliance on unconstrained program representations often introduces irrelevant variation that hinders semantic exploration. We introduce UnlearnEvo DSL, a domain-specific language that represents unlearning procedures along four semantic axes---signal, update, scope, and protection---and supports compositional program structures. This DSL representation provides interpretable mutation targets while offering sufficient expressive capacity to move beyond existing shallow algorithms. On TOFU-forget10, UnlearnEvo DSL helps discover a depth-5 unlearning procedure achieving 0.916 Fitness and 0.960 FQ, improving over the human-designed SOTA ADU by 2.3\% and 3.0\%, respectively. These results demonstrate that the design of the algorithmic search space---not just the search strategy---is a critical factor in effective LLM-guided evolutionary discovery.