LLMs can produce programs for ARC‑AGI‑2 tasks, yet the available compute only permits a few attempts at generation, debugging, and validation. Genetic algorithms can explore a far larger program space, but pure random search rarely starts in a useful neighbourhood of solutions. We bridge the two methods with a compact domain‑specific language (DSL). First, a quantized Qwen3.5‑4B LLM generates an initial batch of programs for each task. These programs seed the initial population of a genetic algorithm, which evolves them while ensuring every mutation remains syntactically valid and executable thanks to the DSL design. In the first 60 tasks of the ARC‑2 public benchmark, the LLM alone solved 2 tasks (3.3%). After applying the genetic search, an additional 4 tasks were solved, yielding 6 correct outputs in total (10.0%). Without the LLM seeding, the evolutionary process failed to produce any solution. The findings demonstrate that genetic search can improve LLM‑generated programs and uncover extra correct solutions.
Review