LLM‑guided evolutionary optimization

Treat candidate programs, prompts, or model inputs as a population and use an LLM to propose targeted mutations; evaluate with an external score, keep the fittest, and repeat — producing cumulative capability gains across generations. Imbue’s Darwinian Evolver applied this pattern to ARC‑AGI‑2 and achieved large, verifiable jumps in benchmark performance for multiple models. — If LLMs can reliably serve as mutation engines that improve other models or artifacts, that creates a low‑friction path to capability improvements and raises practical questions about governance, competitive dynamics, and safety oversight.